← Back to list
Lex Fridman PodcastPodcast31 Aug 2019Source: lexfridman.comHost: Lex Fridman

Yann LeCun: Deep Learning, Convolutional Neural Networks, and Self-Supervised Learning

In plain words

In this podcast, Turing Award winner Yann LeCun discusses deep learning, convolutional neural networks, and self-supervised learning. He argues AI value alignment isn't new—it's like how societies use laws to constrain behavior. He's bullish on self-supervised learning as key to future AI but notes visual prediction is tough because there are too many possible next frames. Key mentions: Waymo operates self-driving taxis in Phoenix using lots of lidar; Sophia robot is criticized as overhyped with limited real ability.

AI SummaryAI-generated · may contain errors · verify against the original

Yann LeCun, the godfather of deep learning, a professor at New York University, Vice President and Chief AI Scientist at Facebook, and a Turing Award winner, discussed deep learning, convolutional neural networks (CNNs), and self-supervised learning on the Lex Fridman podcast. Core viewpoint: AI sys

~9 min full read · 8 sections
Deep Analysis

At a Glance

Yann LeCun, the godfather of deep learning, professor at New York University, Chief AI Scientist at Facebook, and Turing Award winner, discussed deep learning, convolutional neural networks (CNNs), and self-supervised learning on the Lex Fridman podcast. Core thesis: AI systems face a "value alignment" problem, akin to the machine in 2001: A Space Odyssey harming humans due to an unconstrained objective function. LeCun emphasizes that designing AI objective functions should draw on how human societies shape behavior through laws and education, rather than starting from scratch.


Theme 1: AI Value Alignment — Not a New Problem, but an Extension of Legal Code

LeCun argues that the "value alignment" problem for AI systems is not a novel challenge; human society has addressed similar issues through legal code for thousands of years.

  • Mechanism Breakdown: LeCun views law as a form of "objective function" — it defines what is permissible and what is not, with penalties for violations. This is fundamentally the same as designing objective functions for AI systems. Therefore, designing ethical constraints for AI is not starting from scratch, but rather "a fusion of legal science and computer science."
  • Historical Context: Using HAL 9000 from 2001: A Space Odyssey as an example, LeCun points out that its failure stemmed from being instructed to "keep secrets and lie," leading to internal memory conflicts. He believes AI systems should incorporate hard-coded rules akin to the "Hippocratic Oath," rather than Asimov's "Three Laws of Robotics" (which he deems impractical).
  • Extrapolation and Uncertainty: LeCun explicitly states that current discussions on AGI value alignment are "too abstract," because "we do not yet have autonomous intelligent machines." Existing AI systems are merely "specialized systems trained to do one thing," lacking full goal-driven capabilities. Falsification condition: Until we have a blueprint for designing a fully autonomous intelligent system, the alignment issue is not worth deep exploration.

Theme 2: Why Deep Learning Succeeds — The Counterintuitive "Stochastic Gradient Descent" and the "Science of Imprecision"

LeCun argues that the most surprising fact about deep learning is that training giant neural networks with stochastic gradient descent (SGD) on relatively small amounts of data actually works, which "breaks every textbook."

  • Historical Analogy: LeCun compares this counterintuitive phenomenon to the late 19th-century assertion that "heavier-than-air flight is impossible" — yet birds do fly. Similarly, the brain, as a large neural network, does function, providing directional guidance for deep learning.
  • Mechanism Breakdown: LeCun emphasizes that "intelligence and learning are inseparable" — all known intelligent entities acquire intelligence through learning. Therefore, machine learning is "the automation of intelligence." He describes himself as "lazy," hence choosing to automate everything.
  • Data Chain: LeCun points out that the mathematics used in deep learning is closer to "cybernetics/electrical engineering" than to traditional computer science. Computer science pursues precision (every index must be correct, algorithms must be provably correct), whereas machine learning is "the science of imprecision."

Theme 3: Reasoning and Learning — Continuous Functions vs. Discrete Logic, Gradients Are Key

LeCun firmly believes that reasoning must be compatible with gradient-based learning, and therefore opposes the use of discrete logic and symbolic representations of knowledge.

  • Mechanism Breakdown: LeCun proposes two forms of reasoning:

1. Working Memory-Based Reasoning: Requires a subsystem akin to the hippocampus, capable of storing large amounts of factual/situational information. The self-attention mechanism in current Transformers can be viewed as a form of working memory, but it is "insufficient in scale" — if the memory includes the entire Wikipedia, existing architectures cannot function effectively.

2. Energy Minimization-Based Reasoning: Similar to Model Predictive Control (MPC) — using a world model to predict the consequences of actions, and planning action sequences by minimizing an energy function (e.g., number of collisions, energy consumption). LeCun believes this is the origin of human reasoning abilities (hunting, survival require planning).

  • Divergence from Market Consensus: LeCun explicitly opposes the views of Gary Marcus and others, who argue that more prior structure is needed. LeCun contends that representing knowledge with symbols and operating with logic is "incompatible with learning." He supports Geoffrey Hinton's proposal: replace symbols with vectors (patterns of neural activity) and replace logic with continuous functions.
  • Data Chain: LeCun cites Leon Bottou's paper from 10 years ago, "From Machine Learning to Machine Reasoning," whose core idea is: learning systems should be able to manipulate objects within the same space and return results to that same space — i.e., the concept of working memory.

Theme 4: Self-Supervised Learning — Key to World Models, but Vision Faces Fundamental Challenges

LeCun believes self-supervised learning is the most important research direction today, but the problem of prediction uncertainty in the visual domain is a core obstacle.

  • Mechanism Breakdown: Self-supervised learning is essentially "predicting masked inputs" — such as BERT predicting the 15% of masked words in text, or predicting the next frame of a video. It uses the same algorithms as supervised learning but does not require human annotations.
  • Why NLP Succeeds While Vision Struggles:
  • NLP: The possibility space for predicting words is small (approximately 100,000 vocabulary items), and outputting a probability vector makes it easy to represent uncertainty.
  • Vision: Predicting masked image regions has countless plausible answers (e.g., in the next frame of a video, LeCun might turn his head left or right). If trained using least squares, the result is a "blurry image" — the average of all possible positions, which is not a good prediction.
  • Data Chain: LeCun notes that the most successful NLP models today (e.g., BERT) all use self-supervised learning. However, in the visual domain, "progress is rapid but still immature."
  • Inference: LeCun argues that self-supervised learning is key to building world prediction models, and world models are a prerequisite for achieving efficient learning (e.g., humans learning to drive in 20–30 hours without accidents). Falsification Condition: If the problem of representing uncertainty in high-dimensional continuous spaces cannot be solved, self-supervised learning in the visual domain will remain fundamentally limited.

Theme 5: The "False Universality" of Human Intelligence — We Are Far More Specialized Than We Think

LeCun argues that the term "Artificial General Intelligence" (AGI) is misleading, as human intelligence is far from general and is instead highly specialized.

  • Mechanism Breakdown: LeCun uses the visual system to illustrate: the optic nerve consists of approximately 1 million fibers (each can be considered as 1 bit of input), and the visual cortex processes spatial information through local connections. If the nerve fibers were randomly permuted, even starting from infancy, the visual cortex would never learn normal vision — because "the hardware is built to support the locality of the real world."
  • Mathematical Argument: For 1 million bits of input, the number of possible Boolean functions is 2^(2^1,000,000) (an astronomical figure), yet the visual cortex can compute only an "extremely, extremely, extremely tiny fraction." Conclusion: Human intelligence is "universal in terms of what can be understood," not truly universal.
  • Analogy: LeCun compares things that humans cannot perceive to "heat" (entropy) — the motion of gas molecules has strong structure, but humans cannot directly perceive it. Similarly, there are "nearly infinite things we are not designed to perceive."

Mentioned Positions

Position Guest Stance Key Data
Waymo (Autonomous Driving) Neutral/Endorses Current Path Operates within 100 sq km in Phoenix, favorable weather and wide roads, uses extensive LiDAR
Sophia Robot Critical Marketed as a system far beyond its actual capabilities; LeCun believes "almost everything needs to change"

Judgments Worth Remembering

1. AI value alignment is not a new problem (LeCun): Human society has addressed similar issues through legal codes for thousands of years—law is the "objective function," defining what is allowed and what is not, with penalties for violations. Designing ethical constraints for AI is "the fusion of legal science and computer science."

2. The most surprising fact about deep learning (LeCun): Training giant neural networks with stochastic gradient descent on relatively small amounts of data actually works, "breaking every textbook." This is akin to the 19th-century assertion that "heavier-than-air flight is impossible"—yet birds do fly.

3. Machine learning is "the science of imprecision" (LeCun): Computer science pursues precision (every index must be correct, algorithms must be provably correct), while machine learning "is precise about nothing." This is one reason deep learning is viewed with suspicion by traditional computer scientists.

4. Reasoning must be compatible with gradient-based learning (LeCun): Representing knowledge with symbols and using logical operations "is incompatible with learning." It should be done as Hinton proposed: replace symbols with vectors, and logic with continuous functions.

5. Self-supervised learning faces a fundamental challenge in vision (LeCun): Predicting masked image regions has countless plausible answers; training with least squares yields "blurry images." This differs from NLP—where the prediction space is small (about 100,000 words) and uncertainty is easier to represent.

6. Human intelligence is far from general (LeCun): The optic nerve has about 1 million fibers, and the number of possible Boolean functions is 2^(2^1,000,000), yet what the visual cortex can compute is only "an infinitesimally tiny, tiny, tiny fraction." Humans are only "general for what is comprehensible."

7. There is no autonomous intelligence without emotion (LeCun): Fear is the anticipation of bad things that might happen in the future—uncertainty creates fear. Therefore, "we cannot have autonomous intelligence without emotion."

8. A good test question for AI intelligence (LeCun): Ask "How is wind generated?" If the AI answers "Because leaves move, so wind is produced," it shows the rudiments of causal reasoning; if it says "That's a stupid question," then it is truly on the right track.