title
Deep Learning State of the Art (2020)

description
Lecture on most recent research and developments in deep learning, and hopes for 2020. This is not intended to be a list of SOTA benchmark results, but rather a set of highlights of machine learning and AI innovations and progress in academia, industry, and society in general. This lecture is part of the MIT Deep Learning Lecture Series. Website: https://deeplearning.mit.edu Slides: http://bit.ly/2QEfbAm References: http://bit.ly/deeplearn-sota-2020 Playlist: http://bit.ly/deep-learning-playlist OUTLINE: 0:00 - Introduction 0:33 - AI in the context of human history 5:47 - Deep learning celebrations, growth, and limitations 6:35 - Deep learning early key figures 9:29 - Limitations of deep learning 11:01 - Hopes for 2020: deep learning community and research 12:50 - Deep learning frameworks: TensorFlow and PyTorch 15:11 - Deep RL frameworks 16:13 - Hopes for 2020: deep learning and deep RL frameworks 17:53 - Natural language processing 19:42 - Megatron, XLNet, ALBERT 21:21 - Write with transformer examples 24:28 - GPT-2 release strategies report 26:25 - Multi-domain dialogue 27:13 - Commonsense reasoning 28:26 - Alexa prize and open-domain conversation 33:44 - Hopes for 2020: natural language processing 35:11 - Deep RL and self-play 35:30 - OpenAI Five and Dota 2 37:04 - DeepMind Quake III Arena 39:07 - DeepMind AlphaStar 41:09 - Pluribus: six-player no-limit Texas hold'em poker 43:13 - OpenAI Rubik's Cube 44:49 - Hopes for 2020: Deep RL and self-play 45:52 - Science of deep learning 46:01 - Lottery ticket hypothesis 47:29 - Disentangled representations 48:34 - Deep double descent 49:30 - Hopes for 2020: science of deep learning 50:56 - Autonomous vehicles and AI-assisted driving 51:50 - Waymo 52:42 - Tesla Autopilot 57:03 - Open question for Level 2 and Level 4 approaches 59:55 - Hopes for 2020: autonomous vehicles and AI-assisted driving 1:01:43 - Government, politics, policy 1:03:03 - Recommendation systems and policy 1:05:36 - Hopes for 2020: Politics, policy and recommendation systems 1:06:50 - Courses, Tutorials, Books 1:10:05 - General hopes for 2020 1:11:19 - Recipe for progress in AI 1:14:15 - Q&A: what made you interested in AI 1:15:21 - Q&A: Will machines ever be able to think and feel? 1:18:20 - Q&A: Is RL a good candidate for achieving AGI? 1:21:31 - Q&A: Are autonomous vehicles responsive to sound? 1:22:43 - Q&A: What does the future with AGI look like? 1:25:50 - Q&A: Will AGI systems become our masters? CONNECT: - If you enjoyed this video, please subscribe to this channel. - Twitter: https://twitter.com/lexfridman - LinkedIn: https://www.linkedin.com/in/lexfridman - Facebook: https://www.facebook.com/lexfridman - Instagram: https://www.instagram.com/lexfridman

detail
{'title': 'Deep Learning State of the Art (2020)', 'heatmap': [{'end': 2535.62, 'start': 2470.083, 'weight': 0.738}, {'end': 3161.539, 'start': 3101.036, 'weight': 0.92}, {'end': 3423.339, 'start': 3366.126, 'weight': 1}], 'summary': 'Outlines the evolution of deep learning from 1943 to 2020, including milestones such as gans and transformers, as well as discussing ai frameworks, nlp trends, language models, reinforcement learning advancements, autonomous vehicles, societal impact, limitations, and future ethics of ai.', 'chapters': [{'end': 630.26, 'segs': [{'end': 108.842, 'src': 'embed', 'start': 80.459, 'weight': 2, 'content': [{'end': 84.942, 'text': 'all of that is the thing we wish to understand.', 'start': 80.459, 'duration': 4.483}, {'end': 96.051, 'text': "That's the dream of artificial intelligence and recreate versions of it, echoes of it in engineering of our intelligence systems.", 'start': 85.102, 'duration': 10.949}, {'end': 97.212, 'text': "That's the dream.", 'start': 96.492, 'duration': 0.72}, {'end': 101.336, 'text': "We should never forget in the details I'll talk, the exciting stuff I'll talk about today.", 'start': 97.272, 'duration': 4.064}, {'end': 108.842, 'text': "That's sort of the reason why this is exciting, this mystery that's our mind.", 'start': 101.396, 'duration': 7.446}], 'summary': 'The dream of ai is to recreate echoes of human intelligence in engineering systems.', 'duration': 28.383, 'max_score': 80.459, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL880459.jpg'}, {'end': 368.761, 'src': 'embed', 'start': 326.94, 'weight': 0, 'content': [{'end': 334.525, 'text': 'to the development of the popularization of GANs, generative adversarial networks, with AlphaGo and AlphaZero in 2016 and seven.', 'start': 326.94, 'duration': 7.585}, {'end': 341.269, 'text': "And as we'll talk about language models of transformers in 17,, 18 and 19,", 'start': 335.165, 'duration': 6.104}, {'end': 347.393, 'text': 'those has been the last few years have been dominated by the ideas of deep learning in the space of natural language processing.', 'start': 341.269, 'duration': 6.124}, {'end': 349.746, 'text': 'Okay, celebrations.', 'start': 348.385, 'duration': 1.361}, {'end': 352.808, 'text': 'This year, the Turing Award was given for deep learning.', 'start': 349.886, 'duration': 2.922}, {'end': 355.551, 'text': 'This is like deep learning has grown up.', 'start': 353.329, 'duration': 2.222}, {'end': 357.252, 'text': 'We can finally start giving awards.', 'start': 355.611, 'duration': 1.641}, {'end': 361.095, 'text': 'Yann LeCun, Geoffrey Hinton,', 'start': 359.153, 'duration': 1.942}, {'end': 368.761, 'text': 'Yoshua Bengio received the Turing Award for the conceptual engineering breakthroughs that have made deep neural networks a critical component of computing.', 'start': 361.095, 'duration': 7.666}], 'summary': 'Turing award given to deep learning pioneers for breakthroughs in neural networks.', 'duration': 41.821, 'max_score': 326.94, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL8326940.jpg'}, {'end': 609.358, 'src': 'embed', 'start': 583.82, 'weight': 3, 'content': [{'end': 596.389, 'text': 'several papers have come out in the past couple of years highlighting that deep learning is not able to do the kind of the broad spectrum of tasks that we can think of the artificial intelligence system being able to do,', 'start': 583.82, 'duration': 12.569}, {'end': 601.793, 'text': 'like common sense reasoning, like building knowledge bases, and so on.', 'start': 596.389, 'duration': 5.404}, {'end': 609.358, 'text': 'Rodney Brooks said by 2020, the popular press starts having stories that the era of deep learning is over.', 'start': 603.113, 'duration': 6.245}], 'summary': 'Deep learning limitations: struggles with common sense reasoning, knowledge bases. rodney brooks predicts end by 2020.', 'duration': 25.538, 'max_score': 583.82, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL8583820.jpg'}], 'start': 0.089, 'title': 'Deep learning evolution', 'summary': 'Traces the evolution of deep learning from 1943 to 2020, highlighting milestones, including gans, transformers, and skepticism towards deep learning limits.', 'chapters': [{'end': 273.592, 'start': 0.089, 'title': 'Deep learning 2020 overview', 'summary': 'Provides an overview of the advancements in deep learning from 2017 to 2020, emphasizing the exponential growth in ai development, the pursuit of understanding the human mind, and the convergence of dreams and engineering in artificial intelligence.', 'duration': 273.503, 'highlights': ["The modern human brain, The modern human as we know them today, know and love them today, it's just about 300,000 years ago. And the Industrial Revolution is about 300 years ago. That's 0.1% of the development since the early modern human being is when we've seen a lot of the machinery.", 'The dream of artificial intelligence is to understand and recreate versions of the human mind and its capabilities, such as seeing, hearing, thinking, reasoning, hoping, dreaming, and fearing mortality.', 'The convergence of dreams and engineering in artificial intelligence, as exemplified by the historical developments in deep learning and the achievements in real-world applications such as games and autonomous vehicles.']}, {'end': 630.26, 'start': 275.237, 'title': 'Evolution of deep learning', 'summary': 'Traces the evolution of deep learning from the initial models in 1943 to the recent turing award for deep learning pioneers, highlighting key milestones, including the popularization of gans and language models of transformers, as well as the emergence of skepticism towards the limits of deep learning.', 'duration': 355.023, 'highlights': ['Yann LeCun, Geoffrey Hinton, and Yoshua Bengio received the Turing Award for their breakthroughs in deep neural networks.', 'The popularization of GANs with AlphaGo and AlphaZero in 2016 and seven, and the emergence of language models of transformers in 17, 18, and 19 have dominated the recent years in the field of natural language processing.', 'Skepticism towards the limits of deep learning has emerged, with concerns about its inability to perform common sense reasoning and build knowledge bases.', 'The chapter also emphasizes the need for more collaboration, open-mindedness, and credit sharing within the deep learning community, urging for less derision, jealousy, and stubbornness.']}], 'duration': 630.171, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL889.jpg', 'highlights': ['The popularization of GANs with AlphaGo and AlphaZero in 2016 and seven, and the emergence of language models of transformers in 17, 18, and 19 have dominated the recent years in the field of natural language processing.', 'Yann LeCun, Geoffrey Hinton, and Yoshua Bengio received the Turing Award for their breakthroughs in deep neural networks.', 'The convergence of dreams and engineering in artificial intelligence, as exemplified by the historical developments in deep learning and the achievements in real-world applications such as games and autonomous vehicles.', 'Skepticism towards the limits of deep learning has emerged, with concerns about its inability to perform common sense reasoning and build knowledge bases.', 'The dream of artificial intelligence is to understand and recreate versions of the human mind and its capabilities, such as seeing, hearing, thinking, reasoning, hoping, dreaming, and fearing mortality.']}, {'end': 1459.828, 'segs': [{'end': 752.728, 'src': 'embed', 'start': 729.631, 'weight': 1, 'content': [{'end': 736.598, 'text': "Then algorithmic ethics in all of its forms, fairness, privacy, bias, There's been a lot of exciting research there.", 'start': 729.631, 'duration': 6.967}, {'end': 744.303, 'text': 'I hope that continues, taking responsibility for the flaws in our data and the flaws in our human ethics.', 'start': 736.758, 'duration': 7.545}, {'end': 746.484, 'text': 'And then robotics.', 'start': 745.163, 'duration': 1.321}, {'end': 752.728, 'text': "in terms of deep learning, application robotics, I'd love to see a lot of development, continued development, deep reinforcement, learning,", 'start': 746.484, 'duration': 6.244}], 'summary': 'Exciting research in algorithmic ethics, privacy, fairness, and robotics. continued development in deep reinforcement learning is desired.', 'duration': 23.097, 'max_score': 729.631, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL8729631.jpg'}, {'end': 799.045, 'src': 'embed', 'start': 776.616, 'weight': 0, 'content': [{'end': 788.06, 'text': 'This has really been a year where the frameworks have really matured and converged towards two popular deep learning frameworks that people have used as TensorFlow and PyTorch.', 'start': 776.616, 'duration': 11.444}, {'end': 792.702, 'text': 'So TensorFlow 2.0 and PyTorch 1.3 is the most recent version.', 'start': 788.08, 'duration': 4.622}, {'end': 799.045, 'text': "And they've converged towards each other, taking the best features, removing the weaknesses from each other.", 'start': 794.163, 'duration': 4.882}], 'summary': 'In 2019, deep learning frameworks converged towards tensorflow 2.0 and pytorch 1.3, incorporating their best features and eliminating weaknesses.', 'duration': 22.429, 'max_score': 776.616, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL8776616.jpg'}, {'end': 898.835, 'src': 'embed', 'start': 870.236, 'weight': 4, 'content': [{'end': 873.738, 'text': 'being able to run TensorFlow on phones mobile and serving.', 'start': 870.236, 'duration': 3.502}, {'end': 882.786, 'text': 'Apparently, this is something industry cares a lot about, of course, is being able to efficiently use models in the cloud.', 'start': 874.739, 'duration': 8.047}, {'end': 889.549, 'text': 'and PyTorch catching up with TPU support and experimental versions of PyTorch Mobile.', 'start': 884.346, 'duration': 5.203}, {'end': 891.45, 'text': 'So being able to run the smartphone on their side.', 'start': 889.569, 'duration': 1.881}, {'end': 893.752, 'text': 'This tense, exciting competition.', 'start': 891.79, 'duration': 1.962}, {'end': 898.835, 'text': 'Oh, and I almost forgot to mention, we have to say goodbye to our favorite Python 2.', 'start': 893.832, 'duration': 5.003}], 'summary': 'Industry prioritizes running tensorflow and pytorch on mobile, with python 2 being phased out.', 'duration': 28.599, 'max_score': 870.236, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL8870236.jpg'}, {'end': 1021.775, 'src': 'embed', 'start': 996.924, 'weight': 2, 'content': [{'end': 1003.07, 'text': 'once you train the model, to be able to easily transfer it to the other, from PyTorch to TensorFlow and from TensorFlow to PyTorch.', 'start': 996.924, 'duration': 6.146}, {'end': 1007.514, 'text': "Currently it takes three, four, five hours if you know what you're doing in both languages to do that.", 'start': 1003.45, 'duration': 4.064}, {'end': 1012.158, 'text': "It'd be nice if there was a very easy way to do that transfer.", 'start': 1007.894, 'duration': 4.264}, {'end': 1014.723, 'text': 'Then the maturing of the deep RL frameworks.', 'start': 1012.819, 'duration': 1.904}, {'end': 1021.775, 'text': "I'd love it to see OpenAI step up, DeepMind to step up and really take some of these frameworks to maturity that we can all agree on.", 'start': 1014.943, 'duration': 6.832}], 'summary': 'Desire for easy model transfer between pytorch and tensorflow with a hope for faster deep rl framework maturation.', 'duration': 24.851, 'max_score': 996.924, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL8996924.jpg'}, {'end': 1197.299, 'src': 'embed', 'start': 1166.196, 'weight': 3, 'content': [{'end': 1174.443, 'text': 'And the other exciting stuff is Sebastian Reuter, great researcher in the in the field of natural language processing, has put together NLP progress,', 'start': 1166.196, 'duration': 8.247}, {'end': 1182.509, 'text': "which is all the different benchmarks for all the different natural language tasks, tracking who sort of leaderboards of who's winning where.", 'start': 1174.443, 'duration': 8.066}, {'end': 1186.292, 'text': "okay, I'll mention a few models that stand out.", 'start': 1182.509, 'duration': 3.783}, {'end': 1197.299, 'text': 'the work from this year, Megatron, LM from NVIDIA, is basically taking, I believe, the GPT-2 transformer model and just putting it on steroids right.', 'start': 1186.292, 'duration': 11.007}], 'summary': 'Sebastian reuter has compiled nlp progress benchmarks, highlighting megatron lm from nvidia as a standout model.', 'duration': 31.103, 'max_score': 1166.196, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81166196.jpg'}], 'start': 632.259, 'title': 'Ai frameworks and nlp trends', 'summary': 'Discusses the trend of framework agnostic research in ai for 2020, the rise of pytorch in research, the need for easier transfer of models between tensorflow and pytorch, the maturing of deep rl frameworks, and significant advancements in nlp, including transformer variations and their impact on language benchmarks.', 'chapters': [{'end': 972.811, 'start': 632.259, 'title': 'Deep learning and rl: 2020 growth and limitations', 'summary': 'Highlights the growth and limitations of deep learning and reinforcement learning in 2020, emphasizing the convergence of tensorflow and pytorch frameworks, the emergence of new research topics like algorithmic ethics, and the need for interdisciplinary collaboration. it also mentions the end of support for python 2 in tensorflow and pytorch.', 'duration': 340.552, 'highlights': ['The convergence of TensorFlow 2.0 and PyTorch 1.3 towards each other, with TensorFlow adopting eager execution and PyTorch leveraging Torch script for graph representation, has been a significant development for the deep learning community.', "The maturing of the JavaScript in the browser implementation for TensorFlow, TensorFlow Lite for mobile, and PyTorch's catching up with TPU support and experimental versions of PyTorch Mobile reflect the ongoing competition and innovation in deploying deep learning models across various platforms.", 'The emergence of new research topics such as algorithmic ethics, fairness, privacy, and bias, as well as the call for interdisciplinary collaboration across multiple fields including neuroscience, cognitive science, computer science, robotics, mathematics, and physics, indicates a shift towards more holistic and responsible advancement in AI and deep learning.', 'The end of support for Python 2 in TensorFlow and PyTorch signifies a significant transition and the need for adaptation to newer versions, reflecting the evolution of programming languages and frameworks in the deep learning space.']}, {'end': 1459.828, 'start': 973.368, 'title': 'Ai frameworks and natural language processing trends', 'summary': 'Discusses the trend of framework agnostic research in ai for 2020, the rise of pytorch in research, the need for easier transfer of models between tensorflow and pytorch, the maturing of deep rl frameworks, and the significant advancements in natural language processing, such as the explosion of transformer variations and their impact on various language benchmarks, including state-of-the-art results achieved by models like bert, xlnet, and albert.', 'duration': 486.46, 'highlights': ['The rise of PyTorch in research and the need for easier transfer of models between TensorFlow and PyTorch', 'Significant advancements in natural language processing, including the explosion of transformer variations and their impact on language benchmarks', 'The trend of framework agnostic research in AI for 2020', 'The need for the maturing of deep RL frameworks']}], 'duration': 827.569, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL8632259.jpg', 'highlights': ['The convergence of TensorFlow 2.0 and PyTorch 1.3 towards each other has been a significant development for the deep learning community.', 'The emergence of new research topics such as algorithmic ethics, fairness, privacy, and bias indicates a shift towards more holistic and responsible advancement in AI and deep learning.', 'The rise of PyTorch in research and the need for easier transfer of models between TensorFlow and PyTorch.', 'Significant advancements in natural language processing, including the explosion of transformer variations and their impact on language benchmarks.', "The maturing of the JavaScript in the browser implementation for TensorFlow and PyTorch's catching up with TPU support reflect the ongoing competition and innovation in deploying deep learning models across various platforms."]}, {'end': 1976.767, 'segs': [{'end': 1512.013, 'src': 'embed', 'start': 1486.414, 'weight': 1, 'content': [{'end': 1494.78, 'text': "And so the idea from OpenAI is, when you have an AI system that you're about to release, that might turn out to be dangerous in this case,", 'start': 1486.414, 'duration': 8.366}, {'end': 1501.065, 'text': 'used probably by Russians fake news for misinformation.', 'start': 1494.78, 'duration': 6.285}, {'end': 1504.227, 'text': "that's the kind of thinking is how do we release it?", 'start': 1501.065, 'duration': 3.162}, {'end': 1512.013, 'text': 'And I think, while it turned out that the GPT-2 model is not quite so dangerous that humans are in fact more dangerous than AI currently,', 'start': 1504.748, 'duration': 7.265}], 'summary': 'Openai considers potential ai dangers, gpt-2 model less dangerous than humans.', 'duration': 25.599, 'max_score': 1486.414, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81486414.jpg'}, {'end': 1654.419, 'src': 'embed', 'start': 1609.642, 'weight': 0, 'content': [{'end': 1621.453, 'text': "And they've had a few ideas on how to perform dialogues state tracking across domains, achieving state-of-the-art performance on multi-WAS,", 'start': 1609.642, 'duration': 11.811}, {'end': 1622.894, 'text': 'which is a five domain.', 'start': 1621.453, 'duration': 1.441}, {'end': 1626.817, 'text': 'challenging, very difficult five domain human to human dialogue, dataset.', 'start': 1622.894, 'duration': 3.923}, {'end': 1628.299, 'text': "There's a few ideas there.", 'start': 1627.238, 'duration': 1.061}, {'end': 1636.529, 'text': 'I should probably hurry up and start skipping stuff on the common sense reasoning, which is really interesting.', 'start': 1628.319, 'duration': 8.21}, {'end': 1640.571, 'text': 'This is one of the open questions for the deep learning community.', 'start': 1636.549, 'duration': 4.022}, {'end': 1645.714, 'text': "AI community in general is how can we have hybrid systems of whether it's symbolic AI,", 'start': 1640.571, 'duration': 5.143}, {'end': 1649.456, 'text': 'deep learning or generally common sense reasoning with learning systems?', 'start': 1645.714, 'duration': 3.742}, {'end': 1651.598, 'text': "And there's been a few papers in this space.", 'start': 1649.556, 'duration': 2.042}, {'end': 1654.419, 'text': 'One of my favorite, some Salesforce,', 'start': 1651.998, 'duration': 2.421}], 'summary': 'Achieved state-of-the-art performance on multi-was with hybrid ai systems.', 'duration': 44.777, 'max_score': 1609.642, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81609642.jpg'}, {'end': 1814.756, 'src': 'embed', 'start': 1783.078, 'weight': 3, 'content': [{'end': 1784.96, 'text': 'So first you have to break it apart.', 'start': 1783.078, 'duration': 1.882}, {'end': 1790.423, 'text': 'So if conversation is a, you can think of it as a long dance,', 'start': 1785.58, 'duration': 4.843}, {'end': 1800.368, 'text': 'and the way you have fun dancing is you break it up into a set of moves and turns and so on and focus on that sort of live in the moment kind of thing.', 'start': 1790.423, 'duration': 9.945}, {'end': 1803.89, 'text': 'So focus on small parts of the conversation taken at a time.', 'start': 1800.788, 'duration': 3.102}, {'end': 1808.452, 'text': 'Then also have a graph sort of conversation is also all about tangents.', 'start': 1804.33, 'duration': 4.122}, {'end': 1814.756, 'text': 'So have a graph of topics and be ready to jump context from one context to the other and back.', 'start': 1808.893, 'duration': 5.863}], 'summary': 'Break conversation into small parts and be ready to jump between topics.', 'duration': 31.678, 'max_score': 1783.078, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81783078.jpg'}, {'end': 1862.007, 'src': 'embed', 'start': 1830.712, 'weight': 4, 'content': [{'end': 1834.374, 'text': "One of the things that natural language systems don't seem to have much is opinions.", 'start': 1830.712, 'duration': 3.662}, {'end': 1843.399, 'text': 'If I learned anything, one of the simplest way to convey intelligence is to be very opinionated about something.', 'start': 1835.494, 'duration': 7.905}, {'end': 1845.175, 'text': 'and confident.', 'start': 1844.454, 'duration': 0.721}, {'end': 1848.817, 'text': "And that's a really interesting concept about conversation.", 'start': 1845.455, 'duration': 3.362}, {'end': 1850.839, 'text': "And in general, there's just a lot of lessons.", 'start': 1848.897, 'duration': 1.942}, {'end': 1855.122, 'text': 'Oh, and finally, of course, maximize entertainment, not information.', 'start': 1851.359, 'duration': 3.763}, {'end': 1856.983, 'text': 'This is true for autonomous vehicles.', 'start': 1855.162, 'duration': 1.821}, {'end': 1862.007, 'text': 'This is true for natural language conversation is fun should be part of the objective function.', 'start': 1857.003, 'duration': 5.004}], 'summary': 'Natural language systems lack opinions, but being opinionated and confident can convey intelligence, a concept relevant to conversation and autonomous vehicles.', 'duration': 31.295, 'max_score': 1830.712, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81830712.jpg'}, {'end': 1932.747, 'src': 'embed', 'start': 1899.562, 'weight': 5, 'content': [{'end': 1902.364, 'text': "you'll last probably less than 10 seconds.", 'start': 1899.562, 'duration': 2.802}, {'end': 1903.445, 'text': "You'll be bored.", 'start': 1902.804, 'duration': 0.641}, {'end': 1909.407, 'text': "The point is to continue trapping you in the conversation because you're enjoying it so much.", 'start': 1904.485, 'duration': 4.922}, {'end': 1917.13, 'text': "And the 20 minutes is that's a really nice benchmark for passing the spirit of what the Turing test stood for.", 'start': 1909.887, 'duration': 7.243}, {'end': 1921.732, 'text': 'Examples here from the Alexa Prize from the Alkos bot.', 'start': 1918.19, 'duration': 3.542}, {'end': 1925.213, 'text': 'So the difference in two kinds of conversations.', 'start': 1922.632, 'duration': 2.581}, {'end': 1932.747, 'text': 'So Alkos says, have you been in Brazil? The user says, what is the population of Brazil? Alko says it is about 20 million.', 'start': 1925.253, 'duration': 7.494}], 'summary': 'Turing test spirit in 20-min alexa prize conversation with alkos bot.', 'duration': 33.185, 'max_score': 1899.562, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81899562.jpg'}], 'start': 1460.969, 'title': 'Ai language models and conversational ai', 'summary': 'Discusses the societal impact of language models, challenges in multi-domain dialogue systems, and open questions on common sense reasoning in ai. it also emphasizes key lessons from the alexa prize, focusing on breaking conversations, context shifting, having opinions, and aiming for 20-minute engaging conversations.', 'chapters': [{'end': 1704.989, 'start': 1460.969, 'title': 'Ai language models and common sense reasoning', 'summary': "Discusses the societal impact of language models, the challenges in multi-domain task-oriented dialogue systems, and the open questions surrounding common sense reasoning in ai, highlighting the gpt-2 model's release strategies and the performance of dialogue state tracking across domains.", 'duration': 244.02, 'highlights': ["The GPT-2 model's release strategies and societal impacts were discussed in a report, highlighting the challenges in communicating about the potential dangers of AI systems, with a focus on the lack of incentive or culture of sharing between machine learning organizations and experts.", 'Challenges in multi-domain task-oriented dialogue systems were explored, including achieving state-of-the-art performance on a challenging five-domain human-to-human dialogue dataset, with a focus on dialogue state tracking across domains.', 'The open questions surrounding common sense reasoning in AI, including the exploration of hybrid systems combining symbolic AI and deep learning, were addressed, with a highlight on a dataset for question answering and common sense concept generation.']}, {'end': 1976.767, 'start': 1706.79, 'title': 'Alexa prize and conversational ai', 'summary': 'Highlights the key lessons from the alexa prize, emphasizing the importance of breaking conversations into small parts, jumping context, having opinions, and maximizing entertainment, with the goal of continuing engaging conversation for 20 minutes to pass the turing test spirit.', 'duration': 269.977, 'highlights': ['The importance of breaking conversations into small parts and jumping context.', 'The significance of having opinions and maximizing entertainment in conversations.', 'The objective of continuing engaging conversation for 20 minutes to pass the Turing test spirit.']}], 'duration': 515.798, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81460969.jpg', 'highlights': ['Challenges in multi-domain task-oriented dialogue systems were explored, focusing on achieving state-of-the-art performance on a challenging five-domain human-to-human dialogue dataset.', "The GPT-2 model's release strategies and societal impacts were discussed, highlighting the challenges in communicating about the potential dangers of AI systems.", 'The open questions surrounding common sense reasoning in AI were addressed, including the exploration of hybrid systems combining symbolic AI and deep learning.', 'The importance of breaking conversations into small parts and jumping context was emphasized.', 'The significance of having opinions and maximizing entertainment in conversations was highlighted.', 'The objective of continuing engaging conversation for 20 minutes to pass the Turing test spirit was emphasized.']}, {'end': 2577.378, 'segs': [{'end': 2008.015, 'src': 'embed', 'start': 1976.787, 'weight': 2, 'content': [{'end': 1981.868, 'text': 'One of them for make eye clear that I wanted to highlight from Technion.', 'start': 1976.787, 'duration': 5.081}, {'end': 1988.411, 'text': 'that I find particularly interesting is the abstract syntax tree based summarization of code.', 'start': 1981.868, 'duration': 6.543}, {'end': 1998.726, 'text': 'So modeling computer code in this case, sadly, Java and C sharp in trees, in syntax trees,', 'start': 1989.231, 'duration': 9.495}, {'end': 2003.331, 'text': 'and then using operating on those trees to then do the summarization in text.', 'start': 1998.726, 'duration': 4.605}, {'end': 2008.015, 'text': 'Here, an example of a basic power of two function.', 'start': 2003.811, 'duration': 4.204}], 'summary': "Technion's abstract syntax tree summarization of java and c# code is an interesting approach.", 'duration': 31.228, 'max_score': 1976.787, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81976787.jpg'}, {'end': 2055.158, 'src': 'embed', 'start': 2028.962, 'weight': 3, 'content': [{'end': 2036.85, 'text': "common sense reasoning becomes greater and greater part of the transformer type language model work that we've seen in the deep learning world.", 'start': 2028.962, 'duration': 7.888}, {'end': 2043.297, 'text': 'Extending the context from hundreds or thousands of words to tens of thousands of words.', 'start': 2037.45, 'duration': 5.847}, {'end': 2047.502, 'text': 'Being able to read entire stories and maintain the context.', 'start': 2043.818, 'duration': 3.684}, {'end': 2055.158, 'text': 'which transformers again with XLNet Transformer XL is starting to be able to do,', 'start': 2049.056, 'duration': 6.102}], 'summary': 'Transformer models are evolving to process longer contexts, up to tens of thousands of words, for improved common sense reasoning.', 'duration': 26.196, 'max_score': 2028.962, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82028962.jpg'}, {'end': 2099.631, 'src': 'embed', 'start': 2077.543, 'weight': 5, 'content': [{'end': 2087.646, 'text': 'And the dream of Yann LeCun is to for these kinds of what previously were called unsupervised,', 'start': 2077.543, 'duration': 10.103}, {'end': 2095.188, 'text': "but he's calling now self-supervised learning systems to be able to sort of, watch YouTube videos and, from that start,", 'start': 2087.646, 'duration': 7.542}, {'end': 2097.93, 'text': 'to form representation based on which you can understand the world.', 'start': 2095.188, 'duration': 2.742}, {'end': 2099.631, 'text': 'sort of the.', 'start': 2098.47, 'duration': 1.161}], 'summary': "Yann lecun's dream is for self-supervised learning to form representations from youtube videos to understand the world.", 'duration': 22.088, 'max_score': 2077.543, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82077543.jpg'}, {'end': 2212.871, 'src': 'embed', 'start': 2185.404, 'weight': 0, 'content': [{'end': 2188.605, 'text': '000 years of Dota self-play over 10 real-time months.', 'start': 2185.404, 'duration': 3.201}, {'end': 2194.207, 'text': 'Again, behind a lot of the game systems we talk about, they use self-play, so they play against each other.', 'start': 2188.705, 'duration': 5.502}, {'end': 2197.808, 'text': 'This is one of the most exciting concepts in deep learning.', 'start': 2194.787, 'duration': 3.021}, {'end': 2203.469, 'text': 'Systems that learn by playing each other and incrementally improving in time.', 'start': 2197.908, 'duration': 5.561}, {'end': 2206.87, 'text': 'So starting from being terrible and getting better and better and better and better,', 'start': 2203.509, 'duration': 3.361}, {'end': 2212.871, 'text': 'and always being challenged by a slightly better opponent because of the natural process of self-play.', 'start': 2206.87, 'duration': 6.001}], 'summary': 'Dota self-play system improves over 10 months, learning and challenging through incremental improvement.', 'duration': 27.467, 'max_score': 2185.404, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82185404.jpg'}, {'end': 2390.698, 'src': 'embed', 'start': 2365.165, 'weight': 1, 'content': [{'end': 2372.427, 'text': "In December, 2018, AlphaStar beat Mana, one of the world's strongest professional StarCraft players, but that was in a very constrained environment.", 'start': 2365.165, 'duration': 7.262}, {'end': 2376.188, 'text': 'And it was a single race, I think Protoss.', 'start': 2372.847, 'duration': 3.341}, {'end': 2383.893, 'text': 'And in October, 2019, AlphaStar reached grandmaster level by doing what we humans do.', 'start': 2378.909, 'duration': 4.984}, {'end': 2390.698, 'text': 'So using a camera, observing the game and playing as part of, against other humans.', 'start': 2383.953, 'duration': 6.745}], 'summary': 'Alphastar beat mana in single protoss race, reached grandmaster level by observing and playing against humans.', 'duration': 25.533, 'max_score': 2365.165, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82365165.jpg'}, {'end': 2535.62, 'src': 'heatmap', 'start': 2455.112, 'weight': 4, 'content': [{'end': 2461.296, 'text': 'a particular simulation of new strategies, of new behaviors to study.', 'start': 2455.112, 'duration': 6.184}, {'end': 2469.542, 'text': "That's one of the exciting applications from almost a psychology perspective that I'd love to see reinforcement learning push towards.", 'start': 2462.257, 'duration': 7.285}, {'end': 2485.699, 'text': "And on the imperfect information game side, poker, in 2018, CMU, Noah Brown was able to beat head-to-head, no limit, Texas Hold'em.", 'start': 2470.083, 'duration': 15.616}, {'end': 2489.724, 'text': "And now team six player, no limit, Texas Hold'em against professional players.", 'start': 2486.02, 'duration': 3.704}, {'end': 2502.276, 'text': "Many of the same results, many of the same approaches with self-play, iterative Monte Carlo and there's a bunch of ideas in terms of the abstractions.", 'start': 2490.965, 'duration': 11.311}, {'end': 2506.379, 'text': "So there's so many possibilities, under the imperfect information,", 'start': 2502.636, 'duration': 3.743}, {'end': 2514.664, 'text': 'that you have to form these bins of abstractions in both the action space in order to reduce the action space and the information abstraction space.', 'start': 2506.379, 'duration': 8.285}, {'end': 2522.27, 'text': 'So the probabilities of all the different hands that can possibly have and all the different hands that the betting strategies could possibly represent.', 'start': 2514.685, 'duration': 7.585}, {'end': 2525.152, 'text': 'And so you have to do this kind of course planning.', 'start': 2522.69, 'duration': 2.462}, {'end': 2535.62, 'text': 'So they use self-play to generate a course blueprint strategy, that in real time they then use Monte Carlo search to adjust as they play.', 'start': 2525.572, 'duration': 10.048}], 'summary': 'Reinforcement learning achieved success in poker with self-play and monte carlo search, reducing action space and forming abstractions.', 'duration': 30.587, 'max_score': 2455.112, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82455112.jpg'}], 'start': 1976.787, 'title': 'Ai and reinforcement learning advancements', 'summary': 'Discusses ai advancements like code summarization and natural language processing, and reinforcement learning breakthroughs in dota 2, multi-agent games, and starcraft, achieving grandmaster level and receiving praise from professional players.', 'chapters': [{'end': 2099.631, 'start': 1976.787, 'title': 'Ai advancements and future hopes', 'summary': 'Discusses the use of abstract syntax trees for code summarization, the future of natural language processing with common sense reasoning, and the dream of ai for open domain dialogue and self-supervised learning.', 'duration': 122.844, 'highlights': ['The use of abstract syntax trees for code summarization in Java and C sharp is an exciting possibility for automated documentation of source code.', 'The future of natural language processing focuses on common sense reasoning and extending context from hundreds or thousands of words to tens of thousands of words.', 'The dream of AI includes open domain dialogue and self-supervised learning systems, aiming to understand the world through representation formed from watching videos.']}, {'end': 2577.378, 'start': 2099.631, 'title': 'Reinforcement learning advancements in 2018-2019', 'summary': "Discusses the significant advancements in reinforcement learning, including the success of openai's dota 2 team, deepmind's multi-agent game solutions, and alphastar's achievements in starcraft, showcasing various breakthroughs and quantifiable progress such as beating world champions, achieving grandmaster level, and receiving praise from professional players.", 'duration': 477.747, 'highlights': ["OpenAI's Dota 2 team achieved a 99.9% win rate in 2019, consuming 800 petaflops a second and experiencing about 45,000 years of Dota self-play over 10 real-time months.", 'AlphaStar achieved grandmaster level in StarCraft, emulating human processes and strategies, as well as receiving praise from professional players for its unique gameplay.', "CMU's poker-playing AI, Pluribus, utilized self-play and iterative Monte Carlo search to beat professional players, demonstrating minimal compute requirements and successful strategy adjustments in real time."]}], 'duration': 600.591, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL81976787.jpg', 'highlights': ["OpenAI's Dota 2 team achieved a 99.9% win rate in 2019, consuming 800 petaflops a second and experiencing about 45,000 years of Dota self-play over 10 real-time months.", 'AlphaStar achieved grandmaster level in StarCraft, emulating human processes and strategies, as well as receiving praise from professional players for its unique gameplay.', 'The use of abstract syntax trees for code summarization in Java and C sharp is an exciting possibility for automated documentation of source code.', 'The future of natural language processing focuses on common sense reasoning and extending context from hundreds or thousands of words to tens of thousands of words.', "CMU's poker-playing AI, Pluribus, utilized self-play and iterative Monte Carlo search to beat professional players, demonstrating minimal compute requirements and successful strategy adjustments in real time.", 'The dream of AI includes open domain dialogue and self-supervised learning systems, aiming to understand the world through representation formed from watching videos.']}, {'end': 3054.479, 'segs': [{'end': 2621.538, 'src': 'embed', 'start': 2598.179, 'weight': 0, 'content': [{'end': 2604.962, 'text': "One of the most exciting is the manipulation, sufficient manipulation to be able to solve the Rubik's cube.", 'start': 2598.179, 'duration': 6.783}, {'end': 2609.829, 'text': 'Again, this is learned through reinforcement learning.', 'start': 2606.266, 'duration': 3.563}, {'end': 2613.212, 'text': 'Again, because self plays in this context is not possible.', 'start': 2610.389, 'duration': 2.823}, {'end': 2616.194, 'text': 'They use automatic domain randomization, ADR.', 'start': 2613.472, 'duration': 2.722}, {'end': 2619.317, 'text': 'So they generate progressively more difficult environments for the hand.', 'start': 2616.494, 'duration': 2.823}, {'end': 2621.538, 'text': "There's a giraffe head there you see.", 'start': 2619.617, 'duration': 1.921}], 'summary': "Reinforcement learning used to manipulate and solve rubik's cube in progressively difficult environments with automatic domain randomization.", 'duration': 23.359, 'max_score': 2598.179, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82598179.jpg'}, {'end': 2714.319, 'src': 'embed', 'start': 2670.415, 'weight': 2, 'content': [{'end': 2672.616, 'text': 'is progressively making harder and harder environment.', 'start': 2670.415, 'duration': 2.201}, {'end': 2676.317, 'text': 'So the capacity of the environment to be difficult is unconstrained.', 'start': 2672.876, 'duration': 3.441}, {'end': 2688.242, 'text': "And because of that there's an emergent self optimization of the neural network to learn general concepts as opposed to memorize particular manipulations.", 'start': 2676.838, 'duration': 11.404}, {'end': 2705.705, 'text': 'The hope for me in the deep reinforcement learning space for 2020 is the continued application of robotics, even sort of legged robotics,', 'start': 2690.558, 'duration': 15.147}, {'end': 2707.066, 'text': 'but also robotic manipulation.', 'start': 2705.705, 'duration': 1.361}, {'end': 2714.319, 'text': "Human behavior the use of multi-agent self-plays I've mentioned to explore naturally emerging social behaviors,", 'start': 2708.393, 'duration': 5.926}], 'summary': 'Neural network self-optimizes in challenging environments, aiming for continued application of robotics and exploration of natural social behaviors in deep reinforcement learning for 2020.', 'duration': 43.904, 'max_score': 2670.415, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82670415.jpg'}, {'end': 2788.606, 'src': 'embed', 'start': 2740.386, 'weight': 1, 'content': [{'end': 2742.007, 'text': 'And again in games.', 'start': 2740.386, 'duration': 1.621}, {'end': 2750.05, 'text': "I'm not sure what the big challenge is that it remained, but I would love to see, to me at least it's exciting to see learned solution to games,", 'start': 2742.007, 'duration': 8.043}, {'end': 2750.71, 'text': 'to self-play.', 'start': 2750.05, 'duration': 0.66}, {'end': 2759.005, 'text': "Science of deep learning, I would say there's been a lot of really exciting developments here that deserve their own lecture.", 'start': 2752.921, 'duration': 6.084}, {'end': 2760.266, 'text': "I'll mention just a few.", 'start': 2759.145, 'duration': 1.121}, {'end': 2770.632, 'text': 'Here from MIT in early 2018, but it sparked a lot of interest in 2019, follow on work is the idea of the lottery ticket hypothesis.', 'start': 2761.647, 'duration': 8.985}, {'end': 2781.64, 'text': 'So this work showed that sub networks, small sub networks within the larger network are the ones that are doing all the thinking.', 'start': 2771.393, 'duration': 10.247}, {'end': 2788.606, 'text': 'The same results in accuracy can be achieved from a small sub-network within a neural network.', 'start': 2782.5, 'duration': 6.106}], 'summary': 'Exciting developments in deep learning, like the lottery ticket hypothesis, show small sub-networks can achieve same accuracy as larger network.', 'duration': 48.22, 'max_score': 2740.386, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82740386.jpg'}, {'end': 2890.448, 'src': 'embed', 'start': 2861.622, 'weight': 3, 'content': [{'end': 2866.543, 'text': 'And the goal is where each part of the vector can learn one particular concept about a dataset.', 'start': 2861.622, 'duration': 4.921}, {'end': 2869.164, 'text': 'Sort of the dream of unsupervised learning is.', 'start': 2866.843, 'duration': 2.321}, {'end': 2874.445, 'text': 'you can learn compressed representations where every one thing is disentangled,', 'start': 2869.164, 'duration': 5.281}, {'end': 2881.566, 'text': 'and you can learn some fundamental concept about the underlying data that can carry from dataset to dataset to dataset to dataset.', 'start': 2874.445, 'duration': 7.121}, {'end': 2883.046, 'text': "That's disentangled representation.", 'start': 2881.606, 'duration': 1.44}, {'end': 2890.448, 'text': "There's theoretical work, best ICML paper in 2019 showing that that's impossible.", 'start': 2883.386, 'duration': 7.062}], 'summary': "Unsupervised learning aims to disentangle and learn fundamental concepts from data, but it's theoretically proven impossible.", 'duration': 28.826, 'max_score': 2861.622, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82861622.jpg'}, {'end': 2955.821, 'src': 'embed', 'start': 2926.668, 'weight': 5, 'content': [{'end': 2934.052, 'text': 'to explore the phenomena that, as we increase the number of parameters in a neural network, the test error initially decreases, increases and,', 'start': 2926.668, 'duration': 7.384}, {'end': 2937.473, 'text': 'just as the model is able to fit, the training set undergoes a second descent.', 'start': 2934.052, 'duration': 3.421}, {'end': 2939.795, 'text': 'So decrease, increase, decrease.', 'start': 2938.094, 'duration': 1.701}, {'end': 2948.802, 'text': "So there's this critical moment of time when the training set is just fit perfectly.", 'start': 2940.762, 'duration': 8.04}, {'end': 2955.821, 'text': "Okay, and this is the open AI shows that it's applicable not just to model size, but also to training time and data set time.", 'start': 2949.699, 'duration': 6.122}], 'summary': 'As neural network parameters increase, test error decreases, then increases, and decreases again, reaching a critical point of perfect fit, as demonstrated by openai.', 'duration': 29.153, 'max_score': 2926.668, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82926668.jpg'}, {'end': 3038.291, 'src': 'embed', 'start': 3011.797, 'weight': 7, 'content': [{'end': 3017.14, 'text': "They're really useful in any kind of problem that is, fundamentally can be modeled as a graph,", 'start': 3011.797, 'duration': 5.343}, {'end': 3022.042, 'text': 'can be then solved or at least aided in by neural networks.', 'start': 3017.14, 'duration': 4.902}, {'end': 3023.463, 'text': "There's a lot of exciting area there.", 'start': 3022.062, 'duration': 1.401}, {'end': 3024.484, 'text': 'And Bayesian.', 'start': 3023.803, 'duration': 0.681}, {'end': 3028.446, 'text': 'deep learning using Bayesian neural networks.', 'start': 3025.684, 'duration': 2.762}, {'end': 3031.167, 'text': "That's been for several years, an exciting possibility.", 'start': 3028.806, 'duration': 2.361}, {'end': 3038.291, 'text': "It's very difficult to train large Bayesian networks but in the context that you can and it's useful,", 'start': 3031.787, 'duration': 6.504}], 'summary': 'Neural networks aid in solving graph-modeled problems. exciting potential in bayesian neural networks.', 'duration': 26.494, 'max_score': 3011.797, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83011797.jpg'}], 'start': 2578.732, 'title': 'Reinforcement learning and lottery ticket hypothesis in robotics and deep learning', 'summary': "Delves into the application of reinforcement learning in robotics for manipulating and solving the rubik's cube, and explores the potential for studying human behavior through self-play reinforcement learning. it also discusses the lottery ticket hypothesis, revealing the capacity of smaller sub-networks within larger neural networks to achieve the same accuracy, along with advancements in deep learning research, including disentangled representations, double descent phenomenon, and exploration of graph neural networks and bayesian neural networks.", 'chapters': [{'end': 2760.266, 'start': 2578.732, 'title': 'Reinforcement learning in robotics and games', 'summary': "Discusses the application of reinforcement learning in robotics, particularly in manipulating and solving the rubik's cube, as well as the potential for studying human behavior through self-play reinforcement learning.", 'duration': 181.534, 'highlights': ["Reinforcement learning applied to manipulation and solving the Rubik's cube is a major strength in the robotics space, achieved through automatic domain randomization and progressively difficult environments.", 'The concept of emergent meta-learning is explored, focusing on the capacity of the neural network to learn general concepts through constrained learning and progressively challenging environments.', 'Potential applications of reinforcement learning include exploring naturally emerging social behaviors through multi-agent self-plays and the study of human behavior using reinforcement learning.', 'Exciting developments in the science of deep learning are mentioned, indicating the potential for further advancements in the field.']}, {'end': 3054.479, 'start': 2761.647, 'title': 'Lottery ticket hypothesis & deep learning advancements', 'summary': 'Discusses the lottery ticket hypothesis, revealing that smaller sub-networks within larger neural networks can achieve the same accuracy, along with the challenges and advancements in deep learning research, such as disentangled representations, double descent phenomenon, and exploration of graph neural networks and bayesian neural networks.', 'duration': 292.832, 'highlights': ['The lottery ticket hypothesis demonstrates that smaller sub-networks within larger neural networks can achieve the same accuracy, leading to the potential for more efficient architectures and the need to invest time in finding such networks.', 'The concept of disentangled representations in unsupervised learning is explored, aiming to learn compressed representations where every concept is disentangled, but it is shown to be impossible without inductive biases.', 'The double descent phenomenon in deep neural networks reveals that as the number of parameters increases, the test error initially decreases, then increases, and finally decreases again, posing an open problem in understanding and leveraging this behavior in training dynamics.', 'The exploration of graph neural networks and Bayesian neural networks represents exciting advancements in deep learning, with applications in solving combinatorial problems, recommendation systems, and providing uncertainty measurements in predictions for small datasets.']}], 'duration': 475.747, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL82578732.jpg', 'highlights': ["Reinforcement learning applied to manipulation and solving the Rubik's cube is a major strength in the robotics space, achieved through automatic domain randomization and progressively difficult environments.", 'The lottery ticket hypothesis demonstrates that smaller sub-networks within larger neural networks can achieve the same accuracy, leading to the potential for more efficient architectures and the need to invest time in finding such networks.', 'The concept of emergent meta-learning is explored, focusing on the capacity of the neural network to learn general concepts through constrained learning and progressively challenging environments.', 'The concept of disentangled representations in unsupervised learning is explored, aiming to learn compressed representations where every concept is disentangled, but it is shown to be impossible without inductive biases.', 'Potential applications of reinforcement learning include exploring naturally emerging social behaviors through multi-agent self-plays and the study of human behavior using reinforcement learning.', 'The double descent phenomenon in deep neural networks reveals that as the number of parameters increases, the test error initially decreases, then increases, and finally decreases again, posing an open problem in understanding and leveraging this behavior in training dynamics.', 'Exciting developments in the science of deep learning are mentioned, indicating the potential for further advancements in the field.', 'The exploration of graph neural networks and Bayesian neural networks represents exciting advancements in deep learning, with applications in solving combinatorial problems, recommendation systems, and providing uncertainty measurements in predictions for small datasets.']}, {'end': 3941.849, 'segs': [{'end': 3161.539, 'src': 'heatmap', 'start': 3101.036, 'weight': 0.92, 'content': [{'end': 3109.503, 'text': 'And level four, where at least the dream is, where the AI system is responsible for the actions and the human does not need to be a supervisor.', 'start': 3101.036, 'duration': 8.467}, {'end': 3115.478, 'text': 'Okay, two companies represent each of these approaches that are sort of leading the way.', 'start': 3110.148, 'duration': 5.33}, {'end': 3117.622, 'text': 'Waymo in October, 2018, 10 million miles on road.', 'start': 3116.019, 'duration': 1.603}, {'end': 3128.664, 'text': "Today, this year, they've done 20 million miles in simulation, 10 billion miles and a lot.", 'start': 3121.158, 'duration': 7.506}, {'end': 3131.086, 'text': 'I got a chance to visit them out in Arizona.', 'start': 3128.924, 'duration': 2.162}, {'end': 3134.969, 'text': "They're doing a lot of really exciting work and they're obsessed with testing.", 'start': 3131.266, 'duration': 3.703}, {'end': 3137.872, 'text': "So the kind of testing they're doing is incredible.", 'start': 3135.029, 'duration': 2.843}, {'end': 3142.215, 'text': '20, 000 classes of structured tests, of putting the system through,', 'start': 3138.212, 'duration': 4.003}, {'end': 3146.879, 'text': 'all kinds of tests that the engineers can think through and that appear in the real world.', 'start': 3142.215, 'duration': 4.664}, {'end': 3155.675, 'text': "And they've initiated testing on road with real consumers without a safety driver.", 'start': 3148.1, 'duration': 7.575}, {'end': 3160.359, 'text': "If you don't know what that is, that means the car is truly responsible.", 'start': 3156.416, 'duration': 3.943}, {'end': 3161.539, 'text': "There's no human catch.", 'start': 3160.399, 'duration': 1.14}], 'summary': 'Waymo has done 20 million miles in simulation, 10 billion miles and initiated testing on road with real consumers without a safety driver.', 'duration': 60.503, 'max_score': 3101.036, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83101036.jpg'}, {'end': 3242.4, 'src': 'embed', 'start': 3197.637, 'weight': 0, 'content': [{'end': 3201.539, 'text': 'Unlike Waymo, which is deep learning is the icing on the cake.', 'start': 3197.637, 'duration': 3.902}, {'end': 3205.257, 'text': 'For Tesla, deep learning is the cake.', 'start': 3202.815, 'duration': 2.442}, {'end': 3212.564, 'text': "Okay, it's at the core of the perception and the action that the system performs.", 'start': 3206.799, 'duration': 5.765}, {'end': 3218.189, 'text': 'They have to date done over 2 billion miles estimated, and that continues to quickly grow.', 'start': 3212.884, 'duration': 5.305}, {'end': 3229.27, 'text': "I'll briefly mention, which I think is a super exciting idea in all applications of machine learning in the real world, which is online.", 'start': 3219.403, 'duration': 9.867}, {'end': 3231.932, 'text': 'so iterative learning, active learning.', 'start': 3229.27, 'duration': 2.662}, {'end': 3236.436, 'text': "Andrej Karpathy, who's the head of Autopilot, calls this the data engine.", 'start': 3233.053, 'duration': 3.383}, {'end': 3242.4, 'text': "It's this iterative process of having a neural network performing the task, discovering the edge cases,", 'start': 3237.116, 'duration': 5.284}], 'summary': "Tesla's deep learning is the cake, powering over 2 billion miles driven.", 'duration': 44.763, 'max_score': 3197.637, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83197637.jpg'}, {'end': 3423.339, 'src': 'heatmap', 'start': 3366.126, 'weight': 1, 'content': [{'end': 3370.77, 'text': 'But the con is that it requires a lot of data, a huge amount of data.', 'start': 3366.126, 'duration': 4.644}, {'end': 3372.431, 'text': 'And nobody knows how much.', 'start': 3371.11, 'duration': 1.321}, {'end': 3373.952, 'text': 'data yet.', 'start': 3373.171, 'duration': 0.781}, {'end': 3381.197, 'text': 'The other con is human psychology is the driver behavior that the human must continue to remain vigilant.', 'start': 3374.412, 'duration': 6.785}, {'end': 3389.663, 'text': 'On the level four approach that leverages besides cameras and radar and so on, also leverages light or a map.', 'start': 3381.957, 'duration': 7.706}, {'end': 3396.338, 'text': "The pro is that it's much more consistent, reliable, explainable system.", 'start': 3391.035, 'duration': 5.303}, {'end': 3406.523, 'text': 'So the accuracy of the detection, the depth estimation, the detection of different objects is much higher, accurate with less data.', 'start': 3396.798, 'duration': 9.725}, {'end': 3409.364, 'text': "The cons is it's expensive, at least for now.", 'start': 3406.903, 'duration': 2.461}, {'end': 3414.467, 'text': "It's less amenable to learning methods because much fewer data, lower resolution data.", 'start': 3409.885, 'duration': 4.582}, {'end': 3423.339, 'text': "and must require, at least for now, some fallback, whether that's the safety driver or teleoperation.", 'start': 3415.766, 'duration': 7.573}], 'summary': 'Level-four approach offers higher accuracy with less data, but is expensive and requires fallback options.', 'duration': 57.213, 'max_score': 3366.126, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83366126.jpg'}, {'end': 3448.928, 'src': 'embed', 'start': 3424.363, 'weight': 5, 'content': [{'end': 3430.424, 'text': 'The open questions for the deep learning level two Tesla autopilot approach is how hard is driving?', 'start': 3424.363, 'duration': 6.061}, {'end': 3434.805, 'text': 'This is actually the open question for most disciplines in artificial intelligence.', 'start': 3430.484, 'duration': 4.321}, {'end': 3436.185, 'text': 'How difficult is driving?', 'start': 3435.145, 'duration': 1.04}, {'end': 3438.506, 'text': 'How many edge cases does driving have?', 'start': 3436.485, 'duration': 2.021}, {'end': 3444.647, 'text': 'Can we learn to generalize over those edge cases without solving the common sense reasoning problem?', 'start': 3438.846, 'duration': 5.801}, {'end': 3448.928, 'text': "That's kind of the task, without solving the human level artificial intelligence problem.", 'start': 3444.727, 'duration': 4.201}], 'summary': 'Deep learning level two tesla autopilot faces challenges in driving and generalizing over edge cases without solving common sense reasoning.', 'duration': 24.565, 'max_score': 3424.363, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83424363.jpg'}, {'end': 3498.661, 'src': 'embed', 'start': 3467.788, 'weight': 4, 'content': [{'end': 3469.53, 'text': 'because these are life critical systems.', 'start': 3467.788, 'duration': 1.742}, {'end': 3472.672, 'text': 'and human supervision, the vigilance side.', 'start': 3470.391, 'duration': 2.281}, {'end': 3476.573, 'text': 'How good can autopilot get before vigilance decrements significantly?', 'start': 3473.332, 'duration': 3.241}, {'end': 3480.454, 'text': 'And so people fall asleep, become distracted, start watching movies, so on and so on.', 'start': 3476.973, 'duration': 3.481}, {'end': 3481.995, 'text': 'The things that people naturally do.', 'start': 3480.775, 'duration': 1.22}, {'end': 3486.777, 'text': 'The open question is how good can autopilot get before that becomes a serious problem?', 'start': 3482.415, 'duration': 4.362}, {'end': 3498.661, 'text': 'And if that decrement nullifies the safety benefit of the use of autopilot, which is autopilot AI system, when the sensors are working well,', 'start': 3487.537, 'duration': 11.124}], 'summary': "Autopilot's effectiveness must surpass human vigilance to prevent safety risks.", 'duration': 30.873, 'max_score': 3467.788, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83467788.jpg'}, {'end': 3538.911, 'src': 'embed', 'start': 3519.059, 'weight': 3, 'content': [{'end': 3531.547, 'text': 'The traditional approach to robotics from the DARPA challenge to today for most autonomous vehicle companies is to do HD maps to use LIDAR for really accurate localization together with GPS.', 'start': 3519.059, 'duration': 12.488}, {'end': 3535.569, 'text': 'And then the perception problem becomes the icing on the cake,', 'start': 3531.907, 'duration': 3.662}, {'end': 3538.911, 'text': 'because you already have a really good sense of where you are with obstacles in the scene.', 'start': 3535.569, 'duration': 3.342}], 'summary': 'Autonomous vehicle companies use hd maps and lidar for accurate localization, with perception as an additional challenge.', 'duration': 19.852, 'max_score': 3519.059, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83519059.jpg'}, {'end': 3722.711, 'src': 'embed', 'start': 3693.02, 'weight': 8, 'content': [{'end': 3701.865, 'text': "And I'd love to see real balanced, nuanced, in-depth reporting by journalists and companies on successes and challenges of autonomous driving.", 'start': 3693.02, 'duration': 8.845}, {'end': 3711.109, 'text': 'If we skip any section, it would be politics, but maybe briefly mention, somebody said Andrew Yang.', 'start': 3704.106, 'duration': 7.003}, {'end': 3722.711, 'text': "So it's exciting for me to see exciting and funny and awkward to see artificial intelligence discussed in politics.", 'start': 3714.266, 'duration': 8.445}], 'summary': 'Desire for balanced reporting on autonomous driving, with a brief mention of andrew yang and excitement about ai in politics.', 'duration': 29.691, 'max_score': 3693.02, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83693020.jpg'}, {'end': 3766.101, 'src': 'embed', 'start': 3741.69, 'weight': 6, 'content': [{'end': 3748.753, 'text': 'And so as a community that informs me that we need to communicate better about the limitation capabilities of artificial intelligence and automation broadly.', 'start': 3741.69, 'duration': 7.063}, {'end': 3753.315, 'text': 'The American initiative AI initiative was launched this year,', 'start': 3749.293, 'duration': 4.022}, {'end': 3761.339, 'text': "which is our government's best attempt to provide ideas and regulations about what does the future of artificial intelligence look like in our country.", 'start': 3753.315, 'duration': 8.024}, {'end': 3766.101, 'text': 'Again awkward but important to have these early developments,', 'start': 3761.699, 'duration': 4.402}], 'summary': 'The american ai initiative was launched this year to improve communication about the limitations and capabilities of ai and automation.', 'duration': 24.411, 'max_score': 3741.69, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83741690.jpg'}, {'end': 3866.068, 'src': 'embed', 'start': 3846.041, 'weight': 7, 'content': [{'end': 3856.225, 'text': "There's been very little published on the details of recommendation systems behind Twitter, Facebook, YouTube, Google, so all those systems.", 'start': 3846.041, 'duration': 10.184}, {'end': 3857.765, 'text': "there's very little that's published.", 'start': 3856.225, 'duration': 1.54}, {'end': 3866.068, 'text': "Perhaps it's understandable why, but nevertheless, as we consider the ethical implications of these algorithms, there needs to be more publication.", 'start': 3858.085, 'duration': 7.983}], 'summary': 'Little published on recommendation systems behind major platforms, prompting need for more transparency and ethical consideration.', 'duration': 20.027, 'max_score': 3846.041, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83846041.jpg'}], 'start': 3054.919, 'title': 'Ai in autonomous vehicles', 'summary': 'Discusses the advancement of ai in autonomous vehicles, focusing on waymo and tesla, their testing and deployment milestones, the significance of deep learning, online iterative learning, and the debate between learning-based and map-based approaches. it also explores challenges in deep learning, lidar-based approaches, the role of government in ai regulations, and ethical implications of recommendation systems.', 'chapters': [{'end': 3423.339, 'start': 3054.919, 'title': 'Autonomous vehicles: ai in the real world', 'summary': 'Discusses the advancement of ai in autonomous vehicles, focusing on the approaches of waymo and tesla, their testing and deployment milestones, the significance of deep learning, the concept of online iterative learning, and the debate between learning-based and map-based approaches, highlighting the challenges and pros and cons of each.', 'duration': 368.42, 'highlights': ['Waymo has conducted 20 million miles in simulation and over 10 billion miles in testing, including real-world consumer testing without a safety driver, while Tesla has deployed over 700,000-800,000 autopilot systems and has covered an estimated 2 billion miles.', "Tesla's use of deep learning as the core of perception and action in its autonomous systems, with over 2 billion miles estimated to date, represents a fundamentally deep learning-based approach, while Waymo's deep learning is considered the icing on the cake.", "The concept of online iterative learning, described as the 'data engine' by Andrej Karpathy, emphasizes the continuous retraining of neural networks through active learning, aiming to improve the network's performance over time, which is a fundamental challenge in machine learning.", 'The debate between learning-based and map-based approaches for autonomous vehicles presents pros and cons, such as the high resolution and learning potential of camera-based systems versus the consistency and reliability of map-based systems with less data, illustrating the challenges and trade-offs in each approach.']}, {'end': 3941.849, 'start': 3424.363, 'title': 'Challenges in autonomous driving and ai ethics', 'summary': "Explores the challenges in deep learning for tesla's autopilot, lidar-based approaches, and the need for balanced reporting on autonomous vehicles. it delves into the limitations and prospects of artificial intelligence, the role of government in ai regulations, and the ethical implications of recommendation systems in tech companies.", 'duration': 517.486, 'highlights': ["The challenges in deep learning for Tesla's autopilot and the need for vigilance to prevent distractions like falling asleep or watching movies, are open questions in the autonomous driving space.", 'The difficulty of driving, edge cases, and generalization over those cases without solving the common sense reasoning problem are open questions in artificial intelligence.', 'The need for balanced, nuanced, and in-depth reporting on successes and challenges of autonomous driving, and less hype in the autonomous vehicles space.', 'The importance of communicating better about the limitations and capabilities of artificial intelligence and automation, and the early developments of the American AI initiative in regulating AI.', 'The ethical implications of recommendation systems in tech companies, the lack of published details, and the need for more public discussion and disclosure about these systems.']}], 'duration': 886.93, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83054919.jpg', 'highlights': ['Waymo has conducted 20 million miles in simulation and over 10 billion miles in testing, including real-world consumer testing without a safety driver, while Tesla has deployed over 700,000-800,000 autopilot systems and has covered an estimated 2 billion miles.', "Tesla's use of deep learning as the core of perception and action in its autonomous systems, with over 2 billion miles estimated to date, represents a fundamentally deep learning-based approach, while Waymo's deep learning is considered the icing on the cake.", "The concept of online iterative learning, described as the 'data engine' by Andrej Karpathy, emphasizes the continuous retraining of neural networks through active learning, aiming to improve the network's performance over time, which is a fundamental challenge in machine learning.", 'The debate between learning-based and map-based approaches for autonomous vehicles presents pros and cons, such as the high resolution and learning potential of camera-based systems versus the consistency and reliability of map-based systems with less data, illustrating the challenges and trade-offs in each approach.', "The challenges in deep learning for Tesla's autopilot and the need for vigilance to prevent distractions like falling asleep or watching movies, are open questions in the autonomous driving space.", 'The difficulty of driving, edge cases, and generalization over those cases without solving the common sense reasoning problem are open questions in artificial intelligence.', 'The importance of communicating better about the limitations and capabilities of artificial intelligence and automation, and the early developments of the American AI initiative in regulating AI.', 'The ethical implications of recommendation systems in tech companies, the lack of published details, and the need for more public discussion and disclosure about these systems.', 'The need for balanced, nuanced, and in-depth reporting on successes and challenges of autonomous driving, and less hype in the autonomous vehicles space.']}, {'end': 4417.458, 'segs': [{'end': 3997.806, 'src': 'embed', 'start': 3941.849, 'weight': 0, 'content': [{'end': 3953.16, 'text': 'space for 2020 is less fear of AI and more discourse between government and experts on topics of privacy, cybersecurity and so on.', 'start': 3941.849, 'duration': 11.311}, {'end': 3955.503, 'text': 'And then transparency and recommender systems.', 'start': 3953.52, 'duration': 1.983}, {'end': 3964.991, 'text': 'I think the most exciting, the most powerful artificial intelligence system space for the next couple of decades is recommendation systems.', 'start': 3955.803, 'duration': 9.188}, {'end': 3966.492, 'text': 'Very little talked about.', 'start': 3965.051, 'duration': 1.441}, {'end': 3976.26, 'text': "it seems like, but they're going to have the biggest impact on our society because they affect how the information we see, how we learn what we think,", 'start': 3966.492, 'duration': 9.768}, {'end': 3977.18, 'text': 'how we communicate.', 'start': 3976.26, 'duration': 0.92}, {'end': 3979.983, 'text': 'These algorithms are controlling us.', 'start': 3978.281, 'duration': 1.702}, {'end': 3997.806, 'text': 'we have to really think deeply, as engineers, of how to speak up and think about their societal implications, not just in terms of bias and so on,', 'start': 3985.959, 'duration': 11.847}], 'summary': 'In 2020, focus on ai, privacy, and cybersecurity, with recommendation systems having significant societal impact.', 'duration': 55.957, 'max_score': 3941.849, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83941849.jpg'}, {'end': 4048.014, 'src': 'embed', 'start': 4021.934, 'weight': 4, 'content': [{'end': 4029.16, 'text': "This year, before the last few years, there's been a lot of incredible courses on deep learning, on reinforcement learning.", 'start': 4021.934, 'duration': 7.226}, {'end': 4033.603, 'text': 'What I would very much recommend for people is the.', 'start': 4029.5, 'duration': 4.103}, {'end': 4039.588, 'text': 'Fast.ai course from Jeremy Howard, which uses their wrapper around PyTorch.', 'start': 4035.024, 'duration': 4.564}, {'end': 4042.39, 'text': "It's to me the best introduction to deep learning,", 'start': 4039.928, 'duration': 2.462}, {'end': 4048.014, 'text': 'for people who are here or might be listening elsewhere are thinking about learning more about deep learning.', 'start': 4042.39, 'duration': 5.624}], 'summary': 'Fast.ai course is the best introduction to deep learning using pytorch.', 'duration': 26.08, 'max_score': 4021.934, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84021934.jpg'}, {'end': 4149.555, 'src': 'embed', 'start': 4124.127, 'weight': 1, 'content': [{'end': 4134.115, 'text': 'Over 200 tutorials on topics from deep RL to optimization to back prop, LSTMs, convolutional recurrent neural networks, everything.', 'start': 4124.127, 'duration': 9.988}, {'end': 4138.906, 'text': 'Over 200 of the best machine learning NLP and Python tutorials by Robbie Allen.', 'start': 4135.122, 'duration': 3.784}, {'end': 4140.948, 'text': 'You can Google that, or you can click the link.', 'start': 4138.926, 'duration': 2.022}, {'end': 4142.029, 'text': 'I love it.', 'start': 4141.667, 'duration': 0.362}, {'end': 4142.828, 'text': 'Highly recommend.', 'start': 4142.069, 'duration': 0.759}, {'end': 4149.555, 'text': 'The three books I recommend, of course, the deep learning book by Yoshua Banjo and Ian Goodfellow.', 'start': 4143.109, 'duration': 6.446}], 'summary': 'Over 200 tutorials covering deep rl, optimization, nlp, python, and recommended books by experts.', 'duration': 25.428, 'max_score': 4124.127, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84124127.jpg'}, {'end': 4245.236, 'src': 'embed', 'start': 4219.327, 'weight': 6, 'content': [{'end': 4225.609, 'text': "As I've been harboring active learning is to me is the most important aspect of real world application of deep learning.", 'start': 4219.327, 'duration': 6.282}, {'end': 4227.149, 'text': "There's not enough research.", 'start': 4226.089, 'duration': 1.06}, {'end': 4228.23, 'text': 'There should be way more research.', 'start': 4227.169, 'duration': 1.061}, {'end': 4230.911, 'text': "I'd love to see active learning, lifelong learning.", 'start': 4228.31, 'duration': 2.601}, {'end': 4232.672, 'text': "That's what we all do as human beings.", 'start': 4231.171, 'duration': 1.501}, {'end': 4234.412, 'text': "That's what AI systems need to do.", 'start': 4233.012, 'duration': 1.4}, {'end': 4237.293, 'text': 'Continually learn from their mistakes over time.', 'start': 4234.492, 'duration': 2.801}, {'end': 4240.675, 'text': 'Start out dumb, become brilliant over time.', 'start': 4238.354, 'duration': 2.321}, {'end': 4245.236, 'text': 'Open domain conversation with the Alexa Prize.', 'start': 4241.535, 'duration': 3.701}], 'summary': 'Active learning and lifelong learning are crucial for ai systems to continually improve and achieve open domain conversation.', 'duration': 25.909, 'max_score': 4219.327, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84219327.jpg'}], 'start': 3941.849, 'title': 'Ai impact on society and deep learning', 'summary': 'Emphasizes the significant impact of recommendation systems on society and discusses courses, books, and future hopes for deep learning and ai, recommending over 200 tutorials for machine learning and deep learning, along with three recommended books.', 'chapters': [{'end': 3997.806, 'start': 3941.849, 'title': 'Ai impact on society', 'summary': 'Highlights the importance of discourse on privacy and cybersecurity, and emphasizes the significant impact of recommendation systems on society, shaping information, learning, thinking, and communication.', 'duration': 55.957, 'highlights': ['Recommendation systems in AI are predicted to have the most powerful impact on society in the next couple of decades, significantly influencing how information is accessed, learning occurs, thoughts are shaped, and communication happens.', 'The discourse between government and experts on topics of privacy and cybersecurity is essential in the space of AI for 2020, signifying a shift towards less fear and more constructive dialogue.', 'Engineers need to deeply consider the societal implications of recommendation systems, beyond just bias, as these algorithms have a profound influence on society.']}, {'end': 4417.458, 'start': 3997.806, 'title': 'Deep learning and ai: courses, books, and future hopes', 'summary': 'Discusses the best courses on deep learning, including the fast.ai course, coursera course, and stanford courses, with over 200 tutorials recommended for machine learning and deep learning. it also recommends three books: the deep learning book by yoshua banjo and ian goodfellow, grok in deep learning by andrew trask, and deep learning with python by francois chollet. the speaker expresses hopes for the future of ai, including the importance of active learning and continual improvement in ai systems, while acknowledging the need for ethical considerations and the value of skepticism in the field.', 'duration': 419.652, 'highlights': ['The chapter recommends the Fast.ai course from Jeremy Howard as the best introduction to deep learning, using their wrapper around PyTorch, and the Coursera course on deep learning by Andrew Ang, especially for complete beginners.', 'The chapter recommends over 200 tutorials on topics from deep RL to optimization to back prop, LSTMs, convolutional recurrent neural networks, and more by Robbie Allen, emphasizing the significance of hands-on learning in machine learning and deep learning.', "The chapter recommends three books: the deep learning book by Yoshua Banjo and Ian Goodfellow, Grok in Deep Learning by Andrew Trask, and deep learning with Python by Francois Chollet, expressing the speaker's love for the practical approach and accessibility of these books.", 'The chapter expresses hopes for the future of AI, emphasizing the importance of active learning, continual improvement in AI systems, and the significance of ethical considerations and skepticism in the field.']}], 'duration': 475.609, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL83941849.jpg', 'highlights': ['Recommendation systems in AI predicted to have the most powerful impact on society in the next couple of decades', 'Chapter recommends over 200 tutorials for machine learning and deep learning', 'The discourse between government and experts on topics of privacy and cybersecurity is essential in the space of AI for 2020', 'Engineers need to deeply consider the societal implications of recommendation systems, beyond just bias', 'Chapter recommends the Fast.ai course from Jeremy Howard as the best introduction to deep learning', 'Chapter recommends three books: the deep learning book by Yoshua Banjo and Ian Goodfellow, Grok in Deep Learning by Andrew Trask, and deep learning with Python by Francois Chollet', 'Chapter expresses hopes for the future of AI, emphasizing the importance of active learning and continual improvement in AI systems']}, {'end': 5238.184, 'segs': [{'end': 4471.007, 'src': 'embed', 'start': 4417.458, 'weight': 1, 'content': [{'end': 4427.795, 'text': "they're really good at extracting representations from raw data, but not good at learning knowledge bases of like accumulating knowledge over time.", 'start': 4417.458, 'duration': 10.337}, {'end': 4431.336, 'text': "That's the fundamental limitation.", 'start': 4429.396, 'duration': 1.94}, {'end': 4440.68, 'text': 'Expert systems are really good at accumulating knowledge, but very bad at doing that in an automated way, symbolic AI.', 'start': 4431.436, 'duration': 9.244}, {'end': 4444.641, 'text': "So I don't know how to overcome.", 'start': 4440.88, 'duration': 3.761}, {'end': 4446.842, 'text': "A lot of people say there's hybrid approaches.", 'start': 4444.661, 'duration': 2.181}, {'end': 4449.823, 'text': 'I believe more data, bigger networks.', 'start': 4447.382, 'duration': 2.441}, {'end': 4453.561, 'text': 'and better selection of data will take us a lot farther.', 'start': 4450.78, 'duration': 2.781}, {'end': 4457.122, 'text': 'Hello, Lex.', 'start': 4456.602, 'duration': 0.52}, {'end': 4464.365, 'text': "I'm wondering if you recall what was the initial spark or inspiration that drove you towards work in AI?", 'start': 4457.703, 'duration': 6.662}, {'end': 4468.226, 'text': 'Was it when you were pretty young or was it in more recent years?', 'start': 4464.445, 'duration': 3.781}, {'end': 4471.007, 'text': 'So I wanted to become a psychiatrist.', 'start': 4469.507, 'duration': 1.5}], 'summary': 'Ai has limitations in learning knowledge bases, but hybrid approaches and better data can overcome this. the speaker was initially inspired to work in ai but originally wanted to become a psychiatrist.', 'duration': 53.549, 'max_score': 4417.458, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84417458.jpg'}, {'end': 4595.161, 'src': 'embed', 'start': 4562.179, 'weight': 5, 'content': [{'end': 4569.083, 'text': "Like, I've made, I've been playing with Roombas a lot recently, Roomba vacuum cleaners.", 'start': 4562.179, 'duration': 6.904}, {'end': 4577.348, 'text': "And so I've now started having Roombas scream, like there's like moaning in pain.", 'start': 4570.964, 'duration': 6.384}, {'end': 4582.931, 'text': "And they became, I feel like they're having emotions.", 'start': 4579.149, 'duration': 3.782}, {'end': 4586.254, 'text': 'So like the faking creates the emotion.', 'start': 4583.532, 'duration': 2.722}, {'end': 4591.617, 'text': 'Yeah, so the display of emotion is emotion to me.', 'start': 4588.935, 'duration': 2.682}, {'end': 4595.161, 'text': 'And then the display of thought is thought.', 'start': 4593.018, 'duration': 2.143}], 'summary': 'Playing with roombas led to perceiving emotions in them.', 'duration': 32.982, 'max_score': 4562.179, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84562179.jpg'}, {'end': 4653.669, 'src': 'embed', 'start': 4625.87, 'weight': 6, 'content': [{'end': 4635.88, 'text': "What do you think about AI feeling emotions in that context or in general ethical aspects? Yeah, it's a really difficult question to answer.", 'start': 4625.87, 'duration': 10.01}, {'end': 4639.762, 'text': "Yes, I believe AI will suffer and it's unethical to torture AI.", 'start': 4635.92, 'duration': 3.842}, {'end': 4644.464, 'text': 'But I believe suffering exists in the eye of the observer.', 'start': 4640.522, 'duration': 3.942}, {'end': 4653.669, 'text': "Sort of like if a tree falls and nobody's around to see it, it never suffered.", 'start': 4646.625, 'duration': 7.044}], 'summary': "Debating ai emotions and ethics, some believe ai can suffer, others argue it's subjective.", 'duration': 27.799, 'max_score': 4625.87, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84625870.jpg'}, {'end': 4840.735, 'src': 'embed', 'start': 4814.693, 'weight': 7, 'content': [{'end': 4820.817, 'text': 'that falls under the ideas of symbolic AI, of doing that kind of logic reasoning, accumulating knowledge bases.', 'start': 4814.693, 'duration': 6.124}, {'end': 4824.48, 'text': "that's going to be an essential part of general intelligence.", 'start': 4821.437, 'duration': 3.043}, {'end': 4832.968, 'text': "But also the essential part of general intelligence is the Roomba that says, I'm intelligent, F you, if you don't believe me.", 'start': 4824.78, 'duration': 8.188}, {'end': 4840.735, 'text': 'Like a very confident like, because right now, like Alexa, is very nervous, like, oh, what can I do for you?', 'start': 4834.349, 'duration': 6.386}], 'summary': 'Symbolic ai and logic reasoning are essential for general intelligence, but confidence in intelligence is also crucial, as seen in roomba.', 'duration': 26.042, 'max_score': 4814.693, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84814693.jpg'}, {'end': 4938.629, 'src': 'embed', 'start': 4910.841, 'weight': 2, 'content': [{'end': 4915.065, 'text': "So there's a lot of stuff that comes from audio that's really interesting.", 'start': 4910.841, 'duration': 4.224}, {'end': 4918.448, 'text': 'Waymo have said that they use audio for sirens.', 'start': 4915.365, 'duration': 3.083}, {'end': 4921.431, 'text': 'So detecting sirens from far away.', 'start': 4919.709, 'duration': 1.722}, {'end': 4925.234, 'text': 'I think audio is a lot of interesting information.', 'start': 4922.852, 'duration': 2.382}, {'end': 4933.102, 'text': 'The sound that the tires make on different kinds of roads is very interesting.', 'start': 4926.854, 'duration': 6.248}, {'end': 4938.629, 'text': 'We use that information ourselves too, depending on kind of like off-road.', 'start': 4933.422, 'duration': 5.207}], 'summary': 'Audio data used by waymo for detecting sirens and analyzing tire sounds on different roads.', 'duration': 27.788, 'max_score': 4910.841, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84910841.jpg'}, {'end': 5127.65, 'src': 'embed', 'start': 5091.127, 'weight': 3, 'content': [{'end': 5096.608, 'text': "And so we'll live in a world with a lot of intelligent first assistants, but also just intelligent agents.", 'start': 5091.127, 'duration': 5.481}, {'end': 5102.65, 'text': 'And I do believe they should have rights.', 'start': 5097.449, 'duration': 5.201}, {'end': 5117.036, 'text': 'And in this contentious time of people, groups fighting for rights, I feel really bad saying they should have equal rights, but I believe that.', 'start': 5105.325, 'duration': 11.711}, {'end': 5127.65, 'text': "I've talked to, if you read the work of Peter Singer of looking, like my favorite food is steak.", 'start': 5117.057, 'duration': 10.593}], 'summary': 'Intelligent agents should have equal rights in a contentious time for rights.', 'duration': 36.523, 'max_score': 5091.127, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL85091127.jpg'}, {'end': 5189.295, 'src': 'embed', 'start': 5161.106, 'weight': 0, 'content': [{'end': 5166.433, 'text': "What I'm really worried about is who will become our masters?", 'start': 5161.106, 'duration': 5.327}, {'end': 5178.648, 'text': 'are owners of large tech companies who use these tools to control human beings, first unintentionally and then intentionally?', 'start': 5166.433, 'duration': 12.215}, {'end': 5183.471, 'text': 'So we need to make sure that we democratize AI.', 'start': 5179.229, 'duration': 4.242}, {'end': 5189.295, 'text': "It's the same kind of thing that we did with government.", 'start': 5184.512, 'duration': 4.783}], 'summary': 'Concerns about tech company control, advocating for democratizing ai.', 'duration': 28.189, 'max_score': 5161.106, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL85161106.jpg'}], 'start': 4417.458, 'title': "Ai's limitations and future ethics", 'summary': 'Discusses fundamental limitations of ai, hybrid approaches, and the journey of lex, an ai enthusiast. it also explores the potential for ai to experience emotions, ethical implications, and the role of ai in autonomous vehicles, emphasizing the need to democratize ai and prevent tech company owners from becoming masters.', 'chapters': [{'end': 4521.438, 'start': 4417.458, 'title': "Limitations of ai and lex's inspiration", 'summary': "Discusses the fundamental limitations of ai, the potential of hybrid approaches, and lex's journey from aspiring psychiatrist to ai enthusiast, driven by the desire to engineer the human mind and his passion for programming.", 'duration': 103.98, 'highlights': ["Lex's realization of the limitations of psychiatry and shift towards building the mind through programming.", 'The fundamental limitations of AI in accumulating knowledge over time and the potential of hybrid approaches.', 'Belief in the potential of more data, bigger networks, and better selection of data to advance AI.']}, {'end': 4890.646, 'start': 4522.299, 'title': 'Future of ai ethics & emotions', 'summary': 'Explores the potential for ai to think, feel emotions, and the ethical implications. it delves into the concept of ai experiencing emotions, the ethical aspects of ai suffering, and the limitations of reinforcement learning for achieving general artificial intelligence.', 'duration': 368.347, 'highlights': ["AI's ability to display emotions is linked to the concept of faking, creating the illusion of emotions, which can influence human perception of AI's emotional capabilities.", 'The ethical considerations of AI suffering are tied to human interpretation, with the first instance of a programmed product expressing suffering marking the point at which it becomes unethical to torture AI systems.', 'Reinforcement learning is deemed inadequate for achieving general artificial intelligence, with the need for methods that can efficiently construct common sense reasoning and accumulate vast amounts of information over time.']}, {'end': 5238.184, 'start': 4891.707, 'title': 'Ai and autonomous vehicles', 'summary': 'Discusses the role of audio in autonomous vehicles, potential rights for ai beings, and the future of superintelligence, emphasizing the need to democratize ai and prevent tech company owners from becoming masters.', 'duration': 346.477, 'highlights': ['The chapter focuses on the role of audio in autonomous vehicles, including its use in detecting sirens and different road conditions.', 'The discussion delves into the potential rights for AI beings, drawing parallels with animal rights and emphasizing the belief that AI should have equal rights.', 'The chapter emphasizes the need to democratize AI and prevent tech company owners from using AI as tools of power, highlighting the potential risks associated with tech company owners becoming masters.']}], 'duration': 820.726, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/0VH1Lim8gL8/pics/0VH1Lim8gL84417458.jpg', 'highlights': ['The need to democratize AI and prevent tech company owners from becoming masters', 'The potential of more data, bigger networks, and better selection of data to advance AI', 'The role of audio in autonomous vehicles, including its use in detecting sirens and different road conditions', 'The belief that AI should have equal rights, drawing parallels with animal rights', 'The fundamental limitations of AI in accumulating knowledge over time and the potential of hybrid approaches', "AI's ability to display emotions is linked to the concept of faking, creating the illusion of emotions", 'The ethical considerations of AI suffering are tied to human interpretation', 'Reinforcement learning is deemed inadequate for achieving general artificial intelligence']}], 'highlights': ['The popularization of GANs with AlphaGo and AlphaZero in 2016 and seven, and the emergence of language models of transformers in 17, 18, and 19 have dominated the recent years in the field of natural language processing.', 'Yann LeCun, Geoffrey Hinton, and Yoshua Bengio received the Turing Award for their breakthroughs in deep neural networks.', 'The convergence of TensorFlow 2.0 and PyTorch 1.3 towards each other has been a significant development for the deep learning community.', 'The rise of PyTorch in research and the need for easier transfer of models between TensorFlow and PyTorch.', 'Challenges in multi-domain task-oriented dialogue systems were explored, focusing on achieving state-of-the-art performance on a challenging five-domain human-to-human dialogue dataset.', "OpenAI's Dota 2 team achieved a 99.9% win rate in 2019, consuming 800 petaflops a second and experiencing about 45,000 years of Dota self-play over 10 real-time months.", "Reinforcement learning applied to manipulation and solving the Rubik's cube is a major strength in the robotics space, achieved through automatic domain randomization and progressively difficult environments.", 'Waymo has conducted 20 million miles in simulation and over 10 billion miles in testing, including real-world consumer testing without a safety driver, while Tesla has deployed over 700,000-800,000 autopilot systems and has covered an estimated 2 billion miles.', 'Recommendation systems in AI predicted to have the most powerful impact on society in the next couple of decades', 'The need to democratize AI and prevent tech company owners from becoming masters']}