title
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4

description

detail
{'title': 'Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4', 'heatmap': [{'end': 2474.666, 'start': 2430.957, 'weight': 1}], 'summary': 'Yoshua bengio discusses enhancing artificial neural networks through understanding biological neural networks, emphasizing limitations in lstm architectures, disentangled representations for generalization, ai portrayal in science fiction, addressing bias in machine learning, challenges in human-machine interaction, and the impact of small scientific progress on commerce and applications.', 'chapters': [{'end': 82.022, 'segs': [{'end': 54.439, 'src': 'embed', 'start': 24.368, 'weight': 0, 'content': [{'end': 33.248, 'text': "One of the things I studied recently something that we don't know how biological neural networks do,", 'start': 24.368, 'duration': 8.88}, {'end': 43.532, 'text': 'but would be really useful for artificial ones is the ability to do credit assignment through very long time spans.', 'start': 33.248, 'duration': 10.284}, {'end': 51.838, 'text': "There are things that we can in principle do with artificial neural nets, but it's not very convenient and it's not biologically plausible.", 'start': 44.273, 'duration': 7.565}, {'end': 54.439, 'text': 'And this mismatch.', 'start': 52.718, 'duration': 1.721}], 'summary': 'Studied the challenge in biological neural networks for long-term credit assignment.', 'duration': 30.071, 'max_score': 24.368, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ24368.jpg'}], 'start': 0.229, 'title': 'Biological vs. artificial neural networks', 'summary': 'Delves into the enigmatic features of biological neural networks and how understanding them can enhance artificial neural networks, with a focus on long-term credit assignment capability.', 'chapters': [{'end': 82.022, 'start': 0.229, 'title': 'Biological vs. artificial neural networks', 'summary': 'Discusses the mysterious aspects of biological neural networks and the potential for improving artificial neural networks by understanding these mysteries, particularly focusing on the ability to do credit assignment through long time spans.', 'duration': 81.793, 'highlights': ['The ability of biological neural networks to perform credit assignment through long time spans is a mysterious and captivating aspect, which could potentially improve artificial neural networks.', 'The lack of understanding about how biological neural networks perform credit assignment through long time spans creates a mismatch with artificial neural nets, providing an interesting area for study and potential incorporation of new ideas into artificial neural nets.']}], 'duration': 81.793, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ229.jpg', 'highlights': ['The ability of biological neural networks to perform credit assignment through long time spans is a mysterious and captivating aspect, which could potentially improve artificial neural networks.', 'The lack of understanding about how biological neural networks perform credit assignment through long time spans creates a mismatch with artificial neural nets, providing an interesting area for study and potential incorporation of new ideas into artificial neural nets.']}, {'end': 557.827, 'segs': [{'end': 110.352, 'src': 'embed', 'start': 82.386, 'weight': 1, 'content': [{'end': 84.606, 'text': "So let's break credit assignment up a little bit.", 'start': 82.386, 'duration': 2.22}, {'end': 89.307, 'text': "It's a beautifully technical term, but it could incorporate so many things.", 'start': 86.107, 'duration': 3.2}, {'end': 101.53, 'text': 'So is it more on the RNN memory side thinking like that, or is it something about knowledge, building up common sense, knowledge over time?', 'start': 89.928, 'duration': 11.602}, {'end': 110.352, 'text': "or is it more in the reinforcement learning sense that you're picking up rewards over time for a particular to achieve a certain kind of goal?", 'start': 101.53, 'duration': 8.822}], 'summary': 'Exploring credit assignment in ai: rnn memory, common sense, and reinforcement learning.', 'duration': 27.966, 'max_score': 82.386, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ82386.jpg'}, {'end': 154.834, 'src': 'embed', 'start': 119.255, 'weight': 0, 'content': [{'end': 141.768, 'text': 'episodic memories in our brain which we can access later in order to help us both infer causes of things that we are observing now and assign credit to decisions or interpretations we came up with a while ago,', 'start': 119.255, 'duration': 22.513}, {'end': 143.789, 'text': 'when you know those memories were stored.', 'start': 141.768, 'duration': 2.021}, {'end': 151.732, 'text': 'And then we can change the way we would have reacted or interpreted things in the past.', 'start': 144.569, 'duration': 7.163}, {'end': 154.834, 'text': "And now that's credit assignment used for learning.", 'start': 151.893, 'duration': 2.941}], 'summary': 'Episodic memories aid in inferring causes, assigning credit, and adapting reactions for learning.', 'duration': 35.579, 'max_score': 119.255, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ119255.jpg'}, {'end': 292.341, 'src': 'embed', 'start': 259.16, 'weight': 4, 'content': [{'end': 267.882, 'text': "but what, in your view because we're thinking towards the future is the weakest aspect of the way deep neural networks represent the world?", 'start': 259.16, 'duration': 8.722}, {'end': 269.843, 'text': 'What is, in your view, is missing?', 'start': 268.082, 'duration': 1.761}, {'end': 281.202, 'text': 'So current state-of-the-art neural nets, trained on large quantities of images or texts,', 'start': 272.285, 'duration': 8.917}, {'end': 292.341, 'text': "have some level of understanding of what explains those data sets, but it's very basic,", 'start': 283.939, 'duration': 8.402}], 'summary': 'Deep neural networks lack a comprehensive understanding of data, despite being trained on large datasets.', 'duration': 33.181, 'max_score': 259.16, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ259160.jpg'}, {'end': 410.008, 'src': 'embed', 'start': 355.737, 'weight': 2, 'content': [{'end': 365.403, 'text': "We need to have good world models in our neural nets for them to really understand sentences which talk about what's going on in the world.", 'start': 355.737, 'duration': 9.666}, {'end': 377.39, 'text': 'And I think we need language input to help provide clues about what high-level concepts, like semantic concepts,', 'start': 366.023, 'duration': 11.367}, {'end': 380.992, 'text': 'should be represented at the top levels of these neural nets.', 'start': 377.39, 'duration': 3.602}, {'end': 382.278, 'text': 'In fact,', 'start': 381.917, 'duration': 0.361}, {'end': 397.216, 'text': "there is evidence that the purely unsupervised learning of representations doesn't give rise to high-level representations that are as powerful as the ones we're getting from supervised learning.", 'start': 382.278, 'duration': 14.938}, {'end': 404.224, 'text': "And so the clues we're getting just with the labels, not even sentences, is already very powerful.", 'start': 398.581, 'duration': 5.643}, {'end': 410.008, 'text': "Do you think that's an architecture challenge or is it a dataset challenge? Neither.", 'start': 404.865, 'duration': 5.143}], 'summary': 'Good world models in neural nets are crucial for understanding sentences, and language input provides important clues for high-level concepts. supervised learning yields more powerful high-level representations than unsupervised learning.', 'duration': 54.271, 'max_score': 355.737, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ355737.jpg'}], 'start': 82.386, 'title': 'Neural network learning', 'summary': 'Discusses credit assignment and limitations in lstm architectures, emphasizing the need for robust understanding and joint learning of language and world models for improved comprehension.', 'chapters': [{'end': 185.963, 'start': 82.386, 'title': 'Credit assignment in artificial neural networks', 'summary': 'Discusses credit assignment in artificial neural networks, exploring its implications in memory storage, decision-making, and learning, and identifies current limitations in capturing long-term dependencies in current lstm architectures.', 'duration': 103.577, 'highlights': ['Artificial neural networks store episodic memories to infer causes of current observations and assign credit to past decisions, facilitating learning.', 'Artificial neural networks, particularly current LSTM architectures, struggle to capture long-term dependencies beyond sequences with dozens or hundreds of time steps.', 'Credit assignment in artificial neural networks involves incorporating RNN memory, building common sense knowledge over time, and reinforcement learning for picking up rewards to achieve goals.']}, {'end': 557.827, 'start': 186.063, 'title': 'Neural networks and deep learning', 'summary': 'Discusses the capabilities and limitations of deep neural networks in representing the world, emphasizing the need for more robust and abstract understanding, and the importance of joint learning of language and world models for improved comprehension.', 'duration': 371.764, 'highlights': ['The current state-of-the-art neural nets have a very basic and low-level understanding of large quantities of images or texts. The limitations of deep neural networks in representing the world are highlighted, indicating a need for more robust and abstract understanding.', 'The need for joint learning about language and the world to improve the comprehension capabilities of neural networks is emphasized. The importance of joint learning of language and world models for improved comprehension is stressed, suggesting that language input can provide clues about high-level concepts for representation in neural networks.', "The purely unsupervised learning of representations doesn't give rise to high-level representations as powerful as those obtained from supervised learning. The difference in power between purely unsupervised learning and supervised learning in generating high-level representations is highlighted, indicating the potency of supervised learning in providing powerful clues for representation."]}], 'duration': 475.441, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ82386.jpg', 'highlights': ['Artificial neural networks store episodic memories to infer causes of current observations and assign credit to past decisions, facilitating learning.', 'Credit assignment in artificial neural networks involves incorporating RNN memory, building common sense knowledge over time, and reinforcement learning for picking up rewards to achieve goals.', 'The need for joint learning about language and the world to improve the comprehension capabilities of neural networks is emphasized. The importance of joint learning of language and world models for improved comprehension is stressed, suggesting that language input can provide clues about high-level concepts for representation in neural networks.', "The purely unsupervised learning of representations doesn't give rise to high-level representations as powerful as those obtained from supervised learning. The difference in power between purely unsupervised learning and supervised learning in generating high-level representations is highlighted, indicating the potency of supervised learning in providing powerful clues for representation.", 'The current state-of-the-art neural nets have a very basic and low-level understanding of large quantities of images or texts. The limitations of deep neural networks in representing the world are highlighted, indicating a need for more robust and abstract understanding.']}, {'end': 1239.101, 'segs': [{'end': 634.828, 'src': 'embed', 'start': 614.755, 'weight': 0, 'content': [{'end': 626.002, 'text': 'and labs and grad students will continue to tune architectures and explore all kinds of tweaks to make the current state of the art ever slightly better.', 'start': 614.755, 'duration': 11.247}, {'end': 629.605, 'text': "But I don't think that's going to be nearly enough.", 'start': 627.163, 'duration': 2.442}, {'end': 634.828, 'text': 'I think we need some fairly drastic changes in the way that we are considering learning.', 'start': 629.705, 'duration': 5.123}], 'summary': 'Research on ai needs drastic changes for better results.', 'duration': 20.073, 'max_score': 614.755, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ614755.jpg'}, {'end': 872.13, 'src': 'embed', 'start': 842.376, 'weight': 3, 'content': [{'end': 847.719, 'text': "There are lots of decisions we're taking that we can't really explain, even if sometimes we make up a story.", 'start': 842.376, 'duration': 5.343}, {'end': 856.759, 'text': 'And that knowledge is also necessary for machines to take good decisions.', 'start': 851.115, 'duration': 5.644}, {'end': 864.004, 'text': 'And that knowledge is hard to codify in expert systems, rule-based systems, and, you know, classical AI formalism.', 'start': 856.959, 'duration': 7.045}, {'end': 872.13, 'text': 'And there are other issues, of course, with the old AI, like not really good ways of handling uncertainty.', 'start': 864.784, 'duration': 7.346}], 'summary': 'Challenges in explaining decisions for machines, difficulty in codifying knowledge in classical ai, and handling uncertainty.', 'duration': 29.754, 'max_score': 842.376, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ842376.jpg'}, {'end': 986.306, 'src': 'embed', 'start': 964.898, 'weight': 2, 'content': [{'end': 974.401, 'text': 'that build representations in which the important factors hopefully causal factors are nicely separated and easy to pick up from the representation.', 'start': 964.898, 'duration': 9.503}, {'end': 977.062, 'text': "So that's the idea of disentangled representations.", 'start': 975.142, 'duration': 1.92}, {'end': 986.306, 'text': 'It says transform the data into a space where everything becomes easy, we can maybe just learn with linear models about the things we care about.', 'start': 977.883, 'duration': 8.423}], 'summary': 'Disentangled representations aim to separate causal factors for easier learning.', 'duration': 21.408, 'max_score': 964.898, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ964898.jpg'}], 'start': 560.148, 'title': 'Challenges in neural networks', 'summary': 'Discusses the limitations of increasing network size, emphasizes the need for drastic changes in achieving deep understanding in machine learning, and highlights the potential of disentangled representations to improve generalization power.', 'chapters': [{'end': 662.429, 'start': 560.148, 'title': 'Challenges in achieving deep understanding', 'summary': 'Discusses the limitations of simply increasing the size of neural networks and emphasizes the need for more drastic changes in the approach to achieving deep understanding in machine learning, rather than just focusing on tweaking architectures and increasing network depth.', 'duration': 102.281, 'highlights': ['The need for drastic changes in the way learning is approached to achieve deep understanding in machine learning, rather than simply tweaking architectures and increasing network depth.', 'Emphasizing that increasing the depth of networks, such as going from 100 to 10,000 layers, will not solve the problem.', 'The acknowledgment that engineers, companies, labs, and grad students will continue to tune architectures and explore tweaks, but it may not be nearly enough to achieve the necessary progress in machine learning.', 'Questioning the effectiveness of simply increasing the size of neural networks and the number of parameters as a solution to achieving deep understanding in machine learning.']}, {'end': 1239.101, 'start': 663.01, 'title': 'Limitations of current neural networks', 'summary': 'Discusses the limitations of current neural networks in understanding simple environments, the need for vast computing power, the importance of common sense knowledge, and the potential of disentangled representations to improve generalization power.', 'duration': 576.091, 'highlights': ['Current deep learning methods require millions of examples to learn models for very simple tasks, highlighting the inefficiency compared to human learning which needs only dozens of examples. Current deep learning methods struggle to learn models for simple environments, requiring millions of examples, showcasing the inefficiency compared to human learning, which needs only dozens of examples.', 'The need for vast computing power to build neural nets with broad knowledge of the world that typical adult humans have, and the potential for academics to advance the state of the art in training frameworks and learning models. There is a need for vast computing power to build neural nets with broad knowledge, and academics have the potential to advance the state of the art in training frameworks and learning models.', 'The importance of common sense knowledge and intuition, which is not consciously accessible but crucial for machines to make good decisions, and the limitations of classical expert systems in codifying such knowledge. Common sense knowledge and intuition are crucial for machines to make good decisions, but classical expert systems have limitations in codifying such knowledge.', 'The potential of disentangled representations to improve generalization power by separating important factors and relationships between variables, thereby avoiding catastrophic forgetting and enhancing generalization to new distributions. Disentangled representations have the potential to improve generalization power by separating important factors and relationships between variables, avoiding catastrophic forgetting and enhancing generalization to new distributions.']}], 'duration': 678.953, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ560148.jpg', 'highlights': ['The acknowledgment that engineers, companies, labs, and grad students will continue to tune architectures and explore tweaks, but it may not be nearly enough to achieve the necessary progress in machine learning.', 'The need for drastic changes in the way learning is approached to achieve deep understanding in machine learning, rather than simply tweaking architectures and increasing network depth.', 'The potential of disentangled representations to improve generalization power by separating important factors and relationships between variables, avoiding catastrophic forgetting and enhancing generalization to new distributions.', 'The importance of common sense knowledge and intuition, which is not consciously accessible but crucial for machines to make good decisions, and the limitations of classical expert systems in codifying such knowledge.']}, {'end': 1472.117, 'segs': [{'end': 1269.667, 'src': 'embed', 'start': 1239.101, 'weight': 2, 'content': [{'end': 1243.962, 'text': 'you can now make sense of what is going on on this planet where, like visually, for example, things are totally different.', 'start': 1239.101, 'duration': 4.861}, {'end': 1254.515, 'text': "Taking that analogy further and distorting it, let's enter a science fiction world of, say, Space Odyssey 2001 with Hal.", 'start': 1246.107, 'duration': 8.408}, {'end': 1260.44, 'text': 'Yeah Or maybe, which is probably one of my favorite AI movies.', 'start': 1254.775, 'duration': 5.665}, {'end': 1261.561, 'text': 'Me too.', 'start': 1260.96, 'duration': 0.601}, {'end': 1269.667, 'text': "And then there's another one that a lot of people love that maybe a little bit outside of the AI community is Ex Machina.", 'start': 1262.902, 'duration': 6.765}], 'summary': 'Discussion about ai movies like space odyssey 2001 and ex machina.', 'duration': 30.566, 'max_score': 1239.101, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1239101.jpg'}, {'end': 1328.571, 'src': 'embed', 'start': 1300.263, 'weight': 1, 'content': [{'end': 1309.865, 'text': 'so, uh, what do you think is the best way to talk about ai safety, to think about it, to have discourse about it within ai community and outside,', 'start': 1300.263, 'duration': 9.602}, {'end': 1316.706, 'text': 'and grounded in the fact that ex machina is one of the main sources of information for the general public about ai?', 'start': 1309.865, 'duration': 6.841}, {'end': 1328.571, 'text': "so i think i think you're putting it right there's a big difference between the sort of discussion we ought to have within the AI community and the sort of discussion that really matter in the general public.", 'start': 1316.706, 'duration': 11.865}], 'summary': 'Discuss ai safety within the community and address public perception influenced by ex machina.', 'duration': 28.308, 'max_score': 1300.263, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1300263.jpg'}, {'end': 1374.859, 'src': 'embed', 'start': 1346.179, 'weight': 0, 'content': [{'end': 1354.104, 'text': 'because For the public discussion the things I believe really matter are the short-term and medium-term,', 'start': 1346.179, 'duration': 7.925}, {'end': 1365.092, 'text': "very likely negative impacts of AI on society, whether it's from security, like big brother scenarios with face recognition or killer robots,", 'start': 1354.104, 'duration': 10.988}, {'end': 1374.859, 'text': 'or the impact on the job market or concentration of power and discrimination, all kinds of social issues which could actually,', 'start': 1365.092, 'duration': 9.767}], 'summary': 'Ai may have negative impacts on society, including job market, security, power concentration, and discrimination.', 'duration': 28.68, 'max_score': 1346.179, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1346179.jpg'}, {'end': 1427.933, 'src': 'embed', 'start': 1404.639, 'weight': 3, 'content': [{'end': 1412.305, 'text': 'in the same way that you could say should we study what could happen if a meteorite you know came to Earth and destroyed it?', 'start': 1404.639, 'duration': 7.666}, {'end': 1417.769, 'text': "So I think it's very unlikely that this is going to happen or happen in a reasonable future.", 'start': 1412.825, 'duration': 4.944}, {'end': 1418.97, 'text': "It's very..", 'start': 1417.849, 'duration': 1.121}, {'end': 1427.933, 'text': 'The sort of scenario of an AI getting loose goes against my understanding of at least current machine learning and current neural nets and so on.', 'start': 1420.727, 'duration': 7.206}], 'summary': 'Unlikely for ai to go rogue; similar to rare meteorite event.', 'duration': 23.294, 'max_score': 1404.639, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1404639.jpg'}], 'start': 1239.101, 'title': 'Ai in science fiction and ai safety discourse', 'summary': 'Explores the portrayal of ai in popular science fiction movies like space odyssey 2001, hal, and ex machina, and discusses the importance of distinguishing between short-term and medium-term negative impacts of ai on society for public discussion, while highlighting the unlikely but still worth investigating existential risks of ai, and the need for ai safety discourse within and outside the ai community.', 'chapters': [{'end': 1279.634, 'start': 1239.101, 'title': 'Ai in science fiction', 'summary': 'Delves into the portrayal of ai in popular science fiction movies like space odyssey 2001, hal, and ex machina, exploring personal preferences and reactions to these portrayals.', 'duration': 40.533, 'highlights': ['The chapter explores the portrayal of AI in popular science fiction movies such as Space Odyssey 2001, Hal, and Ex Machina, reflecting personal preferences and reactions.', "The conversation touches on the audience's ability to enjoy and appreciate the movie Ex Machina.", 'The speaker expresses personal preferences for AI movies, citing Space Odyssey 2001 and Hal as favorites.']}, {'end': 1472.117, 'start': 1282.435, 'title': 'Ai safety discourse', 'summary': 'Discusses the importance of distinguishing between short-term and medium-term negative impacts of ai on society for public discussion, while highlighting the unlikely but still worth investigating existential risks of ai, and the need for ai safety discourse within and outside the ai community.', 'duration': 189.682, 'highlights': ['The public discussion should focus on short-term and medium-term negative impacts of AI on society, such as security concerns, job market impact, concentration of power, and discrimination, which could threaten democracy. Emphasizes the importance of addressing short-term and medium-term negative impacts of AI on society, including security, job market, power concentration, and discrimination, with potential democratic threats.', 'Existential risk of AI is considered unlikely but worth academic investigation, similar to studying the possibility of a meteorite destroying Earth, although it goes against the understanding of current machine learning and neural nets. Acknowledges the unlikely nature of existential risk of AI but asserts its worth for academic investigation, drawing a parallel with the study of meteorite impact, while expressing skepticism based on current machine learning and neural nets understanding.', 'AI safety discourse should distinguish between discussions within the AI community and those that matter in the general public, advocating for a more pragmatic approach to public discussions and highlighting the limitations of sensationalized portrayals in media. Advocates for a pragmatic approach to AI safety discourse, emphasizing the need to differentiate discussions within the AI community from those relevant to the general public, while critiquing sensationalized media portrayals of AI.']}], 'duration': 233.016, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1239101.jpg', 'highlights': ['The public discussion should focus on short-term and medium-term negative impacts of AI on society, such as security concerns, job market impact, concentration of power, and discrimination, which could threaten democracy.', 'AI safety discourse should distinguish between discussions within the AI community and those that matter in the general public, advocating for a more pragmatic approach to public discussions and highlighting the limitations of sensationalized portrayals in media.', 'The chapter explores the portrayal of AI in popular science fiction movies such as Space Odyssey 2001, Hal, and Ex Machina, reflecting personal preferences and reactions.', 'Existential risk of AI is considered unlikely but worth academic investigation, similar to studying the possibility of a meteorite destroying Earth, although it goes against the understanding of current machine learning and neural nets.']}, {'end': 2005.661, 'segs': [{'end': 1552.378, 'src': 'embed', 'start': 1499.609, 'weight': 0, 'content': [{'end': 1509.377, 'text': 'Science moves by small steps thanks to the collaboration and community of a large number of people interacting.', 'start': 1499.609, 'duration': 9.768}, {'end': 1518.425, 'text': 'And all the scientists who are expert in their field kind of know what is going on even in the industrial labs.', 'start': 1510.238, 'duration': 8.187}, {'end': 1522.628, 'text': 'Information flows and leaks and so on.', 'start': 1519.806, 'duration': 2.822}, {'end': 1528.373, 'text': 'And the spirit of it is very different from the way science is painted in this movie.', 'start': 1522.888, 'duration': 5.485}, {'end': 1530.687, 'text': 'Yeah, let me ask on that point.', 'start': 1528.766, 'duration': 1.921}, {'end': 1539.912, 'text': "It's been the case to this point that kind of, even if the research happens inside Google or Facebook, inside companies, it still kind of comes out.", 'start': 1531.547, 'duration': 8.365}, {'end': 1540.572, 'text': 'ideas come out.', 'start': 1539.912, 'duration': 0.66}, {'end': 1543.173, 'text': 'Do you think that will always be the case with AI?', 'start': 1540.712, 'duration': 2.461}, {'end': 1552.378, 'text': "Is it possible to bottle ideas to the point where there's a set of breakthroughs that go completely undiscovered by the general research community??", 'start': 1543.253, 'duration': 9.125}], 'summary': 'Collaboration and information flow drive scientific progress, even from industrial labs. will ai breakthroughs always be shared?', 'duration': 52.769, 'max_score': 1499.609, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1499609.jpg'}, {'end': 1636.666, 'src': 'embed', 'start': 1581.35, 'weight': 2, 'content': [{'end': 1586.015, 'text': 'So the problem, again, one thing is the movie and you could imagine all kinds of science fiction.', 'start': 1581.35, 'duration': 4.665}, {'end': 1598.365, 'text': 'The problem for me, maybe similar to the question about existential risk is that this kind of movie paints such a wrong picture of what is actual.', 'start': 1586.076, 'duration': 12.289}, {'end': 1606.07, 'text': "you know the actual science and how it's going on that that it can have unfortunate effects on people's understanding of current science.", 'start': 1598.365, 'duration': 7.705}, {'end': 1609.032, 'text': "And so that's kind of sad.", 'start': 1607.311, 'duration': 1.721}, {'end': 1615.939, 'text': "There's an important principle in research, which is diversity.", 'start': 1610.737, 'duration': 5.202}, {'end': 1619.16, 'text': 'So in other words, research is exploration.', 'start': 1616.199, 'duration': 2.961}, {'end': 1621.301, 'text': 'Research is exploration in the space of ideas.', 'start': 1619.36, 'duration': 1.941}, {'end': 1624.902, 'text': 'And different people will focus on different directions.', 'start': 1622.041, 'duration': 2.861}, {'end': 1627.743, 'text': "And this is not just good, it's essential.", 'start': 1625.202, 'duration': 2.541}, {'end': 1636.666, 'text': "So I'm totally fine with people exploring directions that are contrary to mine or look orthogonal to mine.", 'start': 1628.823, 'duration': 7.843}], 'summary': 'Misrepresentation in movies affects public understanding of science; diversity in research is essential.', 'duration': 55.316, 'max_score': 1581.35, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1581350.jpg'}, {'end': 1765.548, 'src': 'embed', 'start': 1738.776, 'weight': 3, 'content': [{'end': 1746.339, 'text': 'And we can use machine learning techniques to try to build predictors, classifiers that are going to be less biased.', 'start': 1738.776, 'duration': 7.563}, {'end': 1757.743, 'text': 'We can do it, for example, using adversarial methods to make our systems less sensitive to these variables we should not be sensitive to.', 'start': 1748.16, 'duration': 9.583}, {'end': 1762.146, 'text': 'So these are clear, well-defined ways of trying to address the problem.', 'start': 1758.303, 'duration': 3.843}, {'end': 1765.548, 'text': 'Maybe they have weaknesses and more research is needed and so on.', 'start': 1762.246, 'duration': 3.302}], 'summary': 'Use machine learning for less biased predictors, classifiers. address weaknesses with more research.', 'duration': 26.772, 'max_score': 1738.776, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1738776.jpg'}, {'end': 1827.703, 'src': 'embed', 'start': 1789.026, 'weight': 5, 'content': [{'end': 1790.187, 'text': 'All right, so this is short term.', 'start': 1789.026, 'duration': 1.161}, {'end': 1799.055, 'text': "Long term, I'm really interested in thinking of how we can instill moral values into computers.", 'start': 1790.828, 'duration': 8.227}, {'end': 1802.758, 'text': "Obviously, this is not something we'll achieve in the next five or 10 years.", 'start': 1799.215, 'duration': 3.543}, {'end': 1805.45, 'text': 'how can we?', 'start': 1804.15, 'duration': 1.3}, {'end': 1813.293, 'text': "there's already work in detecting emotions, for example in images, in sounds, in texts,", 'start': 1805.45, 'duration': 7.843}, {'end': 1827.703, 'text': 'and also studying how different agents interacting different ways may correspond to patterns of, say, injustice which could trigger anger.', 'start': 1813.293, 'duration': 14.41}], 'summary': 'Long-term goal: instill moral values in computers; detect emotions and study interactions for patterns of injustice.', 'duration': 38.677, 'max_score': 1789.026, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1789026.jpg'}, {'end': 1971.825, 'src': 'embed', 'start': 1946.764, 'weight': 6, 'content': [{'end': 1952.03, 'text': 'And can we design, train a system that is gonna be a good teacher?', 'start': 1946.764, 'duration': 5.266}, {'end': 1966.063, 'text': "So in my group we have a project called baby I or baby I game where there is a game or scenario where there's a learning agent and a teaching agent.", 'start': 1953.132, 'duration': 12.931}, {'end': 1969.424, 'text': 'Presumably, the teaching agent would eventually be a human.', 'start': 1967.203, 'duration': 2.221}, {'end': 1971.825, 'text': "But we're not there yet.", 'start': 1970.564, 'duration': 1.261}], 'summary': 'Researching a system for teaching, with a project called baby i game, aiming for human teaching agent.', 'duration': 25.061, 'max_score': 1946.764, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1946764.jpg'}], 'start': 1473.863, 'title': 'Misrepresentation of science in movies and addressing bias in machine learning', 'summary': 'Addresses the misrepresentation of science in movies, emphasizes collaborative nature of scientific progress, and improbability of bottling up breakthrough ideas in ai. it also discusses the importance of diversity in research, the need for addressing bias in machine learning, and the long-term goal of instilling moral values into computers, highlighting short-term techniques for reducing bias and long-term efforts towards teaching machines and detecting emotions.', 'chapters': [{'end': 1606.07, 'start': 1473.863, 'title': 'Misrepresentation of science in movies', 'summary': 'Discusses the misrepresentation of science in movies, emphasizing the collaborative nature of scientific progress and the improbability of bottling up breakthrough ideas in ai.', 'duration': 132.207, 'highlights': ['The collaborative nature of scientific progress is emphasized, highlighting that science moves forward through the collaboration and community of a large number of people interacting.', 'The improbability of bottling up breakthrough ideas in AI is discussed, with the assertion that it is unlikely and not how scientific progress is currently achieved.', 'The detrimental effects of movies misrepresenting science are mentioned, suggesting that it can lead to misunderstandings about current scientific processes.']}, {'end': 2005.661, 'start': 1607.311, 'title': 'Diversity in research and addressing bias in machine learning', 'summary': 'Discusses the importance of diversity in research, the need for addressing bias in machine learning, and the long-term goal of instilling moral values into computers, highlighting the short-term techniques for reducing bias and the long-term efforts towards teaching machines and detecting emotions.', 'duration': 398.35, 'highlights': ['Addressing bias in machine learning through short-term techniques such as building less biased predictors and classifiers using adversarial methods and the need for government regulation to enforce their use.', 'The importance of diversity in research and the exploration of different directions, highlighting that disagreement is a sign of good research and science.', 'The long-term goal of instilling moral values into computers, including detecting emotions and modeling human emotional reactions, particularly focusing on detecting situations triggering anger.', 'The concept of machine teaching and the need for designing strategies to train systems to be good teachers, with a project called baby I game aimed at exploring this concept.']}], 'duration': 531.798, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ1473863.jpg', 'highlights': ['The collaborative nature of scientific progress is emphasized, highlighting that science moves forward through the collaboration and community of a large number of people interacting.', 'The improbability of bottling up breakthrough ideas in AI is discussed, with the assertion that it is unlikely and not how scientific progress is currently achieved.', 'The detrimental effects of movies misrepresenting science are mentioned, suggesting that it can lead to misunderstandings about current scientific processes.', 'Addressing bias in machine learning through short-term techniques such as building less biased predictors and classifiers using adversarial methods and the need for government regulation to enforce their use.', 'The importance of diversity in research and the exploration of different directions, highlighting that disagreement is a sign of good research and science.', 'The long-term goal of instilling moral values into computers, including detecting emotions and modeling human emotional reactions, particularly focusing on detecting situations triggering anger.', 'The concept of machine teaching and the need for designing strategies to train systems to be good teachers, with a project called baby I game aimed at exploring this concept.']}, {'end': 2240.14, 'segs': [{'end': 2074.947, 'src': 'embed', 'start': 2005.661, 'weight': 0, 'content': [{'end': 2012.045, 'text': "maybe teach it the things that the learner has most trouble with, or just add the boundary between what it knows and doesn't know, and so on.", 'start': 2005.661, 'duration': 6.384}, {'end': 2021.473, 'text': "So there's a tradition of these kind of ideas from other fields and like tutorial systems, for example, and AI.", 'start': 2012.705, 'duration': 8.768}, {'end': 2026.658, 'text': 'And of course, people in the humanities have been thinking about these questions.', 'start': 2023.175, 'duration': 3.483}, {'end': 2033.944, 'text': "But I think it's time that machine learning people look at this, because in the future we'll have more and more uh,", 'start': 2026.678, 'duration': 7.266}, {'end': 2040.527, 'text': 'human machine interaction with a human in the loop and, i think, understanding how to make this work better.', 'start': 2033.944, 'duration': 6.583}, {'end': 2044.188, 'text': 'all the problems around that are very interesting and not sufficiently addressed.', 'start': 2040.527, 'duration': 3.661}, {'end': 2047.63, 'text': "you've done a lot of work with language too.", 'start': 2044.188, 'duration': 3.442}, {'end': 2056.112, 'text': 'what aspect of um, the traditionally formulated turing test, a test of natural language understanding and generation,', 'start': 2047.63, 'duration': 8.482}, {'end': 2059.415, 'text': 'in your eyes is the most difficult of of conversation?', 'start': 2056.112, 'duration': 3.303}, {'end': 2063.498, 'text': 'what, in your eyes, is the hardest part of conversation to solve for machines??', 'start': 2059.415, 'duration': 4.083}, {'end': 2074.947, 'text': "So I would say it's everything having to do with the non-linguistic knowledge which implicitly you need in order to make sense of sentences,", 'start': 2064.619, 'duration': 10.328}], 'summary': 'Machine learning must address human-machine interaction for better understanding and problem-solving in conversation.', 'duration': 69.286, 'max_score': 2005.661, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2005661.jpg'}, {'end': 2145.889, 'src': 'embed', 'start': 2087.83, 'weight': 1, 'content': [{'end': 2101.895, 'text': 'I think these are interesting challenges for machine learning because they point in the direction of building systems that both understand how the world works and its causal relationships in the world,', 'start': 2087.83, 'duration': 14.065}, {'end': 2109.778, 'text': 'and associate that knowledge with how to express it in language, either for reading or writing.', 'start': 2101.895, 'duration': 7.883}, {'end': 2114.3, 'text': "You speak French? Yes, it's my mother tongue.", 'start': 2112.019, 'duration': 2.281}, {'end': 2116.383, 'text': "It's one of the romance languages.", 'start': 2114.701, 'duration': 1.682}, {'end': 2123.089, 'text': 'Do you think passing the Turing test and all the underlying challenges we just mentioned depend on language?', 'start': 2117.203, 'duration': 5.886}, {'end': 2125.711, 'text': 'Do you think it might be easier in French than it is in English?', 'start': 2123.189, 'duration': 2.522}, {'end': 2128.013, 'text': 'Or is it independent of language??', 'start': 2126.252, 'duration': 1.761}, {'end': 2130.336, 'text': "I think it's independent of language.", 'start': 2128.994, 'duration': 1.342}, {'end': 2145.889, 'text': 'I would like to build systems that can use the same principles, the same learning mechanisms to learn from human agents, whatever their language.', 'start': 2132.197, 'duration': 13.692}], 'summary': 'Challenges in building systems that understand causal relationships and express knowledge in language, independent of language.', 'duration': 58.059, 'max_score': 2087.83, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2087830.jpg'}, {'end': 2240.14, 'src': 'embed', 'start': 2208.599, 'weight': 5, 'content': [{'end': 2214.463, 'text': 'And what have you learned from the experience? Listen to your inner voice.', 'start': 2208.599, 'duration': 5.864}, {'end': 2222.73, 'text': "Don't, you know, be trying to just please the crowds and the fashion.", 'start': 2215.644, 'duration': 7.086}, {'end': 2232.679, 'text': 'And if you have a strong intuition about something that is not contradicted by actual evidence, go for it.', 'start': 2223.071, 'duration': 9.608}, {'end': 2235.616, 'text': 'I mean, it could be contradicted by people.', 'start': 2233.854, 'duration': 1.762}, {'end': 2240.14, 'text': "But not your own instinct of based on everything you've learned.", 'start': 2236.957, 'duration': 3.183}], 'summary': 'Trust your intuition over crowd-pleasing, if not contradicted by evidence.', 'duration': 31.541, 'max_score': 2208.599, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2208599.jpg'}], 'start': 2005.661, 'title': 'Challenges in machine learning', 'summary': 'Discusses the need for machine learning to address challenges in human-machine interaction, focusing on understanding non-linguistic knowledge in conversation and causal relationships with language, and passing the turing test.', 'chapters': [{'end': 2074.947, 'start': 2005.661, 'title': 'Challenges in machine learning for conversational understanding', 'summary': 'Discusses the need for machine learning to address the challenges of human-machine interaction, particularly in understanding non-linguistic knowledge in conversation.', 'duration': 69.286, 'highlights': ['Understanding non-linguistic knowledge is a significant challenge in conversation for machines, as it is implicitly required to make sense of sentences.', 'The future will involve increased human-machine interaction, necessitating a better understanding of how to make it work, with unresolved and intriguing problems around that.', 'There is a tradition of ideas from other fields, such as tutorial systems and AI, that can be beneficial for addressing the challenges in machine learning for human-machine interaction.']}, {'end': 2240.14, 'start': 2074.947, 'title': 'Challenges for machine learning', 'summary': 'Discusses the challenges for machine learning in understanding the world and its causal relationships, associating that knowledge with language, and the independence of language in passing the turing test.', 'duration': 165.193, 'highlights': ['Building systems that understand the world and its causal relationships and associate that knowledge with language is an interesting challenge for machine learning. None', 'Passing the Turing test and the underlying challenges are independent of language, and the goal is to build systems that can learn from human agents regardless of their language. None', 'The importance of staying warm with friends and listening to your inner voice in continuing research, and the value of intuition not contradicted by actual evidence. None']}], 'duration': 234.479, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2005661.jpg', 'highlights': ['Understanding non-linguistic knowledge is a significant challenge in conversation for machines.', 'Building systems that understand the world and its causal relationships and associate that knowledge with language is an interesting challenge for machine learning.', 'The future will involve increased human-machine interaction, necessitating a better understanding of how to make it work.', 'There is a tradition of ideas from other fields, such as tutorial systems and AI, that can be beneficial for addressing the challenges in machine learning for human-machine interaction.', 'Passing the Turing test and the underlying challenges are independent of language, and the goal is to build systems that can learn from human agents regardless of their language.', 'The importance of staying warm with friends and listening to your inner voice in continuing research, and the value of intuition not contradicted by actual evidence.']}, {'end': 2517.957, 'segs': [{'end': 2332.75, 'src': 'embed', 'start': 2240.34, 'weight': 0, 'content': [{'end': 2246.286, 'text': 'So, of course, you have to adapt your beliefs when your experiments contradict those beliefs.', 'start': 2240.34, 'duration': 5.946}, {'end': 2250.71, 'text': 'But you have to stick to your beliefs otherwise.', 'start': 2247.687, 'duration': 3.023}, {'end': 2254.894, 'text': "It's what allowed me to go through those years.", 'start': 2251.851, 'duration': 3.043}, {'end': 2256.255, 'text': "It's what allowed me to..", 'start': 2254.974, 'duration': 1.281}, {'end': 2266.413, 'text': 'persist in directions that, you know, took time, whatever other people think, took time to mature and bring fruits.', 'start': 2257.276, 'duration': 9.137}, {'end': 2272.486, 'text': 'So history of AI is marked with these.', 'start': 2268.004, 'duration': 4.482}, {'end': 2279.789, 'text': "of course it's marked with technical breakthroughs, but it's also marked with these seminal events that capture the imagination of the community.", 'start': 2272.486, 'duration': 7.303}, {'end': 2287.772, 'text': 'Most recent, I would say, AlphaGo beating the world champion human Go player was one of those moments.', 'start': 2280.829, 'duration': 6.943}, {'end': 2297.919, 'text': 'What do you think the next such moment might be? Okay, so first of all, I think that these so-called seminal events are overrated.', 'start': 2289.013, 'duration': 8.906}, {'end': 2305.587, 'text': 'As I said, science really moves by small steps.', 'start': 2302.584, 'duration': 3.003}, {'end': 2320.199, 'text': "Now what happens is you make one more small step and it's like the drop that fills the bucket, and then you have drastic consequences,", 'start': 2305.967, 'duration': 14.232}, {'end': 2325.284, 'text': "because now you're able to do something you were not able to do before or now, say,", 'start': 2320.199, 'duration': 5.085}, {'end': 2332.75, 'text': 'the cost of building some device or solving a problem becomes cheaper than what existed, and you have a new market that opens up.', 'start': 2325.284, 'duration': 7.466}], 'summary': 'Beliefs guide persistence, history of ai marked by seminal events like alphago beating world champion, but small steps drive science forward.', 'duration': 92.41, 'max_score': 2240.34, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2240340.jpg'}, {'end': 2391.431, 'src': 'embed', 'start': 2363.285, 'weight': 2, 'content': [{'end': 2368.589, 'text': 'what are the two hottest topics? GANs and reinforcement learning.', 'start': 2363.285, 'duration': 5.304}, {'end': 2377.438, 'text': 'Even though in Montreal in particular, reinforcement learning was something pretty much absent just two or three years ago.', 'start': 2369.39, 'duration': 8.048}, {'end': 2384.865, 'text': "So there's really a big interest from students and there's a big interest from people like me.", 'start': 2378.038, 'duration': 6.827}, {'end': 2391.431, 'text': "So I would say this is something where we're going to see more progress,", 'start': 2387.788, 'duration': 3.643}], 'summary': 'In montreal, interest in reinforcement learning has surged in the past two to three years, particularly among students and professionals.', 'duration': 28.146, 'max_score': 2363.285, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2363285.jpg'}, {'end': 2474.666, 'src': 'heatmap', 'start': 2430.957, 'weight': 1, 'content': [{'end': 2441.829, 'text': 'Well, GANs or other generative models, I believe, will be crucial ingredients in building agents that can understand the world.', 'start': 2430.957, 'duration': 10.872}, {'end': 2451.093, 'text': 'A lot of the successes in reinforcement learning in the past has been with policy gradient, where you just learn a policy,', 'start': 2443.07, 'duration': 8.023}, {'end': 2452.574, 'text': "you don't actually learn a model of the world.", 'start': 2451.093, 'duration': 1.481}, {'end': 2454.694, 'text': 'But there are lots of issues with that.', 'start': 2453.394, 'duration': 1.3}, {'end': 2461.157, 'text': "And we don't know how to do model-based RL right now, but I think this is where we have to go.", 'start': 2455.615, 'duration': 5.542}, {'end': 2471.484, 'text': 'in order to build models that can generalize faster and better like to new distributions that capture, to some extent at least,', 'start': 2462.077, 'duration': 9.407}, {'end': 2474.666, 'text': 'the underlying causal mechanisms in the world.', 'start': 2471.484, 'duration': 3.182}], 'summary': 'Gans and generative models are crucial for building agents to understand the world and improve model-based rl.', 'duration': 43.709, 'max_score': 2430.957, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2430957.jpg'}, {'end': 2517.957, 'src': 'embed', 'start': 2491.308, 'weight': 5, 'content': [{'end': 2496.33, 'text': 'You know, when I was an adolescent, I was reading a lot and then I started reading science fiction.', 'start': 2491.308, 'duration': 5.022}, {'end': 2499.311, 'text': 'There you go.', 'start': 2497.83, 'duration': 1.481}, {'end': 2500.131, 'text': "That's it.", 'start': 2499.771, 'duration': 0.36}, {'end': 2501.952, 'text': "That's where I got hooked.", 'start': 2500.711, 'duration': 1.241}, {'end': 2510.234, 'text': 'And then, you know, I had one of the first personal computers and I got hooked in programming.', 'start': 2502.592, 'duration': 7.642}, {'end': 2512.275, 'text': 'And so it just, you know.', 'start': 2511.115, 'duration': 1.16}, {'end': 2514.636, 'text': 'Start with fiction and then make it a reality.', 'start': 2512.515, 'duration': 2.121}, {'end': 2515.156, 'text': "That's right.", 'start': 2514.656, 'duration': 0.5}, {'end': 2517.517, 'text': 'Josje, thank you so much for talking today.', 'start': 2515.176, 'duration': 2.341}, {'end': 2517.957, 'text': 'My pleasure.', 'start': 2517.557, 'duration': 0.4}], 'summary': 'Reading science fiction as an adolescent led to a passion for programming and technology.', 'duration': 26.649, 'max_score': 2491.308, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2491308.jpg'}], 'start': 2240.34, 'title': 'Ai beliefs and small scientific progress', 'summary': "Discusses adapting beliefs in ai, emphasizing seminal events like alphago's victory, and the impact of small scientific progress on commerce and applications, focusing on reinforcement learning and gans.", 'chapters': [{'end': 2297.919, 'start': 2240.34, 'title': 'Beliefs and seminal events in ai history', 'summary': 'Emphasizes the importance of adapting beliefs in the face of contradicting experiments, while also highlighting the significance of seminal events in the history of ai, such as alphago beating the world champion human go player, and questions the overrating of such events.', 'duration': 57.579, 'highlights': ['The importance of adapting beliefs when experiments contradict them, while also emphasizing the necessity of sticking to beliefs that allow for persistence and eventual fruition of ideas.', "The history of AI is marked by technical breakthroughs as well as seminal events that capture the community's imagination, with AlphaGo beating the world champion human Go player being one of these significant moments.", 'Questioning the overrating of so-called seminal events in the field of AI, raising doubts about the significance and impact of such events.']}, {'end': 2517.957, 'start': 2302.584, 'title': 'The impact of small scientific progress', 'summary': 'Discusses how small scientific progress can lead to drastic consequences in commerce and applications, with reinforcement learning and gans being the two hottest topics in the ai community, despite not yet providing much industrial fallout.', 'duration': 215.373, 'highlights': ['Reinforcement learning and GANs are the two hottest topics in the AI community, showing a big interest from students and professionals, despite not yet providing significant industrial benefits.', 'Small scientific progress can lead to drastic consequences in commerce and applications, opening up new markets and reducing costs for building devices or solving problems.', "The speaker's fascination with artificial intelligence began with reading science fiction as an adolescent and later delving into programming with personal computers."]}], 'duration': 277.617, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/azOmzumh0vQ/pics/azOmzumh0vQ2240340.jpg', 'highlights': ["The history of AI is marked by technical breakthroughs as well as seminal events that capture the community's imagination, with AlphaGo beating the world champion human Go player being one of these significant moments.", 'Small scientific progress can lead to drastic consequences in commerce and applications, opening up new markets and reducing costs for building devices or solving problems.', 'Reinforcement learning and GANs are the two hottest topics in the AI community, showing a big interest from students and professionals, despite not yet providing significant industrial benefits.', 'The importance of adapting beliefs when experiments contradict them, while also emphasizing the necessity of sticking to beliefs that allow for persistence and eventual fruition of ideas.', 'Questioning the overrating of so-called seminal events in the field of AI, raising doubts about the significance and impact of such events.', "The speaker's fascination with artificial intelligence began with reading science fiction as an adolescent and later delving into programming with personal computers."]}], 'highlights': ['The ability of biological neural networks to perform credit assignment through long time spans is a mysterious and captivating aspect, which could potentially improve artificial neural networks.', 'The lack of understanding about how biological neural networks perform credit assignment through long time spans creates a mismatch with artificial neural nets, providing an interesting area for study and potential incorporation of new ideas into artificial neural nets.', 'Artificial neural networks store episodic memories to infer causes of current observations and assign credit to past decisions, facilitating learning.', 'Credit assignment in artificial neural networks involves incorporating RNN memory, building common sense knowledge over time, and reinforcement learning for picking up rewards to achieve goals.', 'The need for joint learning about language and the world to improve the comprehension capabilities of neural networks is emphasized.', 'The importance of joint learning of language and world models for improved comprehension is stressed, suggesting that language input can provide clues about high-level concepts for representation in neural networks.', "The purely unsupervised learning of representations doesn't give rise to high-level representations as powerful as those obtained from supervised learning.", 'The current state-of-the-art neural nets have a very basic and low-level understanding of large quantities of images or texts.', 'The acknowledgment that engineers, companies, labs, and grad students will continue to tune architectures and explore tweaks, but it may not be nearly enough to achieve the necessary progress in machine learning.', 'The need for drastic changes in the way learning is approached to achieve deep understanding in machine learning, rather than simply tweaking architectures and increasing network depth.', 'The potential of disentangled representations to improve generalization power by separating important factors and relationships between variables, avoiding catastrophic forgetting and enhancing generalization to new distributions.', 'The importance of common sense knowledge and intuition, which is not consciously accessible but crucial for machines to make good decisions, and the limitations of classical expert systems in codifying such knowledge.', 'The public discussion should focus on short-term and medium-term negative impacts of AI on society, such as security concerns, job market impact, concentration of power, and discrimination, which could threaten democracy.', 'AI safety discourse should distinguish between discussions within the AI community and those that matter in the general public, advocating for a more pragmatic approach to public discussions and highlighting the limitations of sensationalized portrayals in media.', 'The chapter explores the portrayal of AI in popular science fiction movies such as Space Odyssey 2001, Hal, and Ex Machina, reflecting personal preferences and reactions.', 'Existential risk of AI is considered unlikely but worth academic investigation, similar to studying the possibility of a meteorite destroying Earth, although it goes against the understanding of current machine learning and neural nets.', 'The collaborative nature of scientific progress is emphasized, highlighting that science moves forward through the collaboration and community of a large number of people interacting.', 'The improbability of bottling up breakthrough ideas in AI is discussed, with the assertion that it is unlikely and not how scientific progress is currently achieved.', 'The detrimental effects of movies misrepresenting science are mentioned, suggesting that it can lead to misunderstandings about current scientific processes.', 'Addressing bias in machine learning through short-term techniques such as building less biased predictors and classifiers using adversarial methods and the need for government regulation to enforce their use.', 'The importance of diversity in research and the exploration of different directions, highlighting that disagreement is a sign of good research and science.', 'The long-term goal of instilling moral values into computers, including detecting emotions and modeling human emotional reactions, particularly focusing on detecting situations triggering anger.', 'The concept of machine teaching and the need for designing strategies to train systems to be good teachers, with a project called baby I game aimed at exploring this concept.', 'Understanding non-linguistic knowledge is a significant challenge in conversation for machines.', 'Building systems that understand the world and its causal relationships and associate that knowledge with language is an interesting challenge for machine learning.', 'The future will involve increased human-machine interaction, necessitating a better understanding of how to make it work.', 'There is a tradition of ideas from other fields, such as tutorial systems and AI, that can be beneficial for addressing the challenges in machine learning for human-machine interaction.', 'Passing the Turing test and the underlying challenges are independent of language, and the goal is to build systems that can learn from human agents regardless of their language.', 'The importance of staying warm with friends and listening to your inner voice in continuing research, and the value of intuition not contradicted by actual evidence.', "The history of AI is marked by technical breakthroughs as well as seminal events that capture the community's imagination, with AlphaGo beating the world champion human Go player being one of these significant moments.", 'Small scientific progress can lead to drastic consequences in commerce and applications, opening up new markets and reducing costs for building devices or solving problems.', 'Reinforcement learning and GANs are the two hottest topics in the AI community, showing a big interest from students and professionals, despite not yet providing significant industrial benefits.', 'The importance of adapting beliefs when experiments contradict them, while also emphasizing the necessity of sticking to beliefs that allow for persistence and eventual fruition of ideas.', 'Questioning the overrating of so-called seminal events in the field of AI, raising doubts about the significance and impact of such events.', "The speaker's fascination with artificial intelligence began with reading science fiction as an adolescent and later delving into programming with personal computers."]}