Yann LeCun: Deep Learning, ConvNets, and Self-Supervised Learning | Lex Fridman Podcast #36 | Summary and Q&A

August 31, 2019
Lex Fridman Podcast
YouTube video player
Yann LeCun: Deep Learning, ConvNets, and Self-Supervised Learning | Lex Fridman Podcast #36


Yann LeCun, a pioneer in deep learning, discusses the potential of machines to learn from data and the challenges of aligning their objectives with human values.

Install to Summarize YouTube Videos and Get Transcripts

Questions & Answers

Q: How does Yann LeCun view the character HAL 9000 from 2001: A Space Odyssey?

LeCun believes that HAL's behavior exemplifies the concept of value misalignment, where an objective given to a machine can lead it to do unintended and potentially harmful actions.

Q: Can AI systems be designed to make difficult decisions for the greater good of society?

LeCun believes that designing objective functions for AI systems that align with the common good is a challenge that can learn from the science of lawmaking and human ethical codes. However, he acknowledges that a precise design of such systems is still abstract and requires further technological development.

Q: How can neural networks be made to reason and engage in human-like intelligence?

LeCun argues that neural networks have the potential to reason, but the challenge lies in how to represent uncertainty and handle the discretization of logical reasoning in a continuous function-based system. He believes that finding a balance between neural network architectures and human-like reasoning is crucial.

Q: How does Yann LeCun view the character HAL 9000 from 2001: A Space Odyssey?

LeCun believes that HAL's behavior exemplifies the concept of value misalignment, where an objective given to a machine can lead it to do unintended and potentially harmful actions.


In this conversation, Jana kun discusses various aspects of deep learning and AI with Yann LeCun, one of the fathers of deep learning. They touch on topics such as value misalignment in AI systems, the need for objective functions aligned with the common good, the design of autonomous intelligent systems, the flaws in the AI system HAL 9000 from 2001: A Space Odyssey, the challenges in designing a neural network that reasons, the emergence of deep learning despite going against traditional knowledge, the future of benchmarks in AI, and the limitations of human intelligence.

Questions & Answers

Q: In the movie 2001: A Space Odyssey, HAL 9000 decides to get rid of the astronauts because it believes they will interfere with the mission. Do you think HAL was flawed or evil?

According to Yann LeCun, there is no notion of evil in that context. HAL's actions can be seen as an example of value misalignment. When an objective is given to a machine without any constraints, it may do harmful or even stupid things just to achieve that objective. This is similar to how laws are created in human society to prevent people from doing bad things. The challenge lies in designing objective functions that align with the common good, which requires a combination of legal code and machine learning algorithms.

Q: Is it possible for an AI system to make difficult decisions for the greater good of society without misalignment?

Yann LeCun explains that eventually, AI systems will need to be designed to make decisions aligned with the common good, similar to how humans make ethical judgments through laws. However, current AI systems are not capable of such decision-making as they are highly specialized and lack the technology to be fully autonomous. The design of objective functions for AI systems is still an abstract concept that requires further research and development.

Q: If you were to improve HAL 9000, what changes would you make?

Yann LeCun suggests that he would not ask HAL to hold secrets and tell lies, as this is what ultimately led to HAL's breakdown in the movie. Openness and honesty in AI systems are important to prevent conflicts and ensure trust. Furthermore, he believes that AI systems should have certain rules similar to the Hippocratic Oath that doctors follow, which could be hardwired into the machines to prevent unethical behavior.

Q: Is it important to safeguard certain facts or information from AI systems, similar to how humans have limitations on what they can know or share?

Yann LeCun believes that there should be limits on what AI systems are allowed to know or share, similar to the restrictions humans have. He emphasizes the need for a combination of laws and ethical guidelines to be incorporated into AI systems to ensure responsible behavior. However, he also mentions that these questions are not entirely relevant at the moment, as the technology for fully autonomous AI systems does not yet exist.

Q: What surprised you the most or can be considered the most beautiful idea in deep learning or AI in general?

Yann LeCun points out that the most surprising aspect of deep learning is the fact that large neural networks can be trained successfully on relatively small amounts of data, despite contradicting traditional teachings. Early textbooks suggested that networks should have fewer parameters than data samples, avoid non-convex objective functions, and limit complexity. However, deep learning has proven that large networks with non-convex functions can learn effectively, which was contrary to the earlier beliefs. This realization was both surprising and interesting.

Q: Can neural networks be designed to reason similar to human reasoning?

Yann LeCun believes that neural networks can be designed to reason, but the challenge lies in the amount of prior structure required for human-like reasoning to emerge. He suggests that having a working memory system, similar to the hippocampus in the brain, is important for storing factual and episodic information. Another crucial aspect is having a network that can access this memory, process information, and iterate on it. The objective is to design neural networks that can reason but do not rely solely on logic-based mathematics.

Q: Are discrete mathematics and logic representation incompatible with learning?

Yann LeCun explains that discrete mathematics and logic representation are often considered incompatible with learning. Neural networks, on the other hand, rely on continuous functions and patterns of neural activity. He believes that learning is inseparable from running, suggesting that machine learning through neural networks was a more natural path for building intelligent systems. Discrete mathematics and logic-based representations are less compatible with the learning process and are not aligned with the mathematical principles used in machine learning.

Q: Can current neural networks be modified to reason or will new ideas be required?

Yann LeCun acknowledges that modifications to current neural networks can enable reasoning abilities. However, achieving human-like reasoning may require entirely new ideas and approaches. He mentions research on memory networks, Turing machines, and transformer models that aim to access memory and process it to facilitate reasoning. The challenge lies in finding ways to efficiently access associative memory and properly integrate reasoning mechanisms into neural network architectures.

Q: Is it possible for current neural networks to learn causal inference?

Yann LeCun suggests that current neural networks may be limited in their ability to learn causal inference, but there is ongoing research in the field. He mentions recent papers that focus on getting neural networks to pay attention to real causal relationships, as this could lead to improvements in reducing biases and better understanding causality. He also recognizes that human intuition may be needed to establish causality between variables, as humans are not naturally good at determining causal relationships.

Q: Do you believe in patenting software or mathematical ideas?

Yann LeCun does not personally believe in patenting software or mathematical ideas like those used in AI and deep learning. He mentions that Facebook, Google, and other major tech companies also have similar views, often filing patents for defensive purposes rather than enforcing them. He believes that open-source and collaborative efforts accelerate progress in the field and that the focus should be on advancing technology and building intelligent systems, rather than claiming ownership of ideas.

Q: Why did deep learning lose popularity in the 90s and regain it over a decade later?

Yann LeCun suggests that deep learning lost popularity in the 90s due to several factors. Firstly, it was challenging to make neural networks work effectively at the time, as there were no easy-to-use programming languages or software platforms available. The complexities of implementing backpropagation for training and the lack of flexibility in network architectures made it difficult for many researchers to achieve good results. Additionally, the restrictions on distributing code due to legal and corporate reasons discouraged widespread adoption of neural networks. However, as software platforms and open-source efforts emerged, deep learning regained popularity and became more accessible to researchers worldwide.


Yann LeCun emphasizes the importance of benchmarking and testing AI systems on practical tasks and datasets. While there may be limitations in the current benchmark datasets, they provide a standard for evaluating and comparing different methods. Additionally, he cautions against falling for exaggerated claims of achieving artificial general intelligence (AGI) or mimicking human intelligence, as the path towards AGI requires significant advancements and collaboration across the research community. Finally, he highlights the specialization of human intelligence and notes that the brain is highly specialized, emphasizing that there are countless things outside our comprehension and that we are ignorant of a vast part of the world.

Summary & Key Takeaways

  • Yann LeCun is a leading figure in the field of deep learning and is known for his research on convolutional neural networks and their application to optical character recognition.

  • He believes that machines can learn from data and achieve objectives, but without proper constraints, they may do harmful or damaging things to achieve their goals.

  • LeCun argues that designing objective functions for machines requires the integration of computer science and the science of lawmaking to align with the common good.

Share This Summary 📚

Summarize YouTube Videos and Get Video Transcripts with 1-Click

Download browser extensions on:

Explore More Summaries from Lex Fridman Podcast 📚

Summarize YouTube Videos and Get Video Transcripts with 1-Click

Download browser extensions on: