Yann LeCun on Deep Learning's Foundations, AI Reasoning, and Ethical Alignment

This podcast features Yann LeCun, a pioneer in deep learning, discussing the foundational principles, current challenges, and future directions of artificial intelligence. The conversation begins with an exploration of AI ethics, using the fictional Hal 9000 as a thought experiment to illustrate "value misalignment" – the critical issue of ensuring AI objectives align with human well-being. LeCun argues that designing objective functions for AI is analogous to human lawmaking, a millennia-old practice of shaping behavior through rules and consequences, suggesting that the science of lawmaking and computer science will converge to address these complex ethical dilemmas. He emphasizes that while current AI systems are specialized, the abstract discussion of AGI ethics serves as a useful tool for understanding human ethical codes.

LeCun expresses his profound surprise and conviction regarding the empirical success of deep learning, particularly the ability of gigantic neural networks to learn effectively from relatively small datasets despite violating traditional textbook rules against excessive parameters and non-convex objective functions. He posits that intelligence is inseparable from learning, advocating for gradient-based learning as the most promising path, contrasting it with discrete, logic-based approaches common in traditional computer science. He describes machine learning as "the science of sloppiness," highlighting its continuous, approximate nature over the exactness of discrete mathematics.

The discussion delves into the future of AI reasoning, where LeCun outlines the need for systems with "working memory" akin to the human hippocampus, capable of storing and iteratively processing episodic information. He mentions memory networks, neural Turing machines, and transformers as steps in this direction but notes the challenge of scaling associative memory for vast knowledge bases like Wikipedia. LeCun also connects reasoning to planning, suggesting that energy minimization and model predictive control offer a framework for AI to plan actions, drawing parallels to the survival strategies of early species. He critiques symbolic AI and knowledge graphs as too rigid and brittle, preferring the manipulation of "vectors" and "continuous functions" as advocated by Geoff Hinton.

Finally, LeCun addresses the contentious topic of causal inference, acknowledging Judea Pearl's concerns but pointing to ongoing research, including recent work by Leon Bottou, on enabling neural networks to discern causal relationships. He touches on the historical "AI winter" of the 1990s, attributing it to the difficulty of implementing and debugging early neural networks without modern software tools and platforms. LeCun shares a personal anecdote about the patenting of convolutional neural networks at Bell Labs, expressing his skepticism about software patents and advocating for open-source collaboration to accelerate progress in the field.

There's no notion of evil in that in that context other than the fact that people die but it was an example of what people call value misalignment right you give an objective to a machine and the Machine strives to achieve this objective and if you don't put any constraints on this objective like don't kill people and don't do things like this the Machine given the power will do stupid things just to achieve this dis objective or damaging things to achieve its objective.

designing objective functions for people is something that we know how to do and we don't do it by you know programming things although the legal code is called code so that tells you something and it's actually the design of an object you function that's really what legal code is right it tells you you can do it what you can't do if you do it you pay that much that's that's an objective function.

the fact that you gigantic neural nets trying to train them on you know relatively small amounts of data relatively with the caste grid in the center that it actually works breaks everything you read in every textbook right every pre deep learning textbook that told you you need to have fewer parameters and you have data samples you know if you have non-convex objective function you have no guarantee of convergence you know all the things that you read in textbook and they tell you stay away from this and they were all wrong.

intelligence is inseparable from running so you the idea somehow that you can create an intelligent machine by basically programming for me was a non-starter you know from the start every intelligent entity that we know about arrives at this intelligence to learning so learning you know machine learning was completely obvious path.

machine learning is the science of sloppiness really that's beautiful.

what you need to have is a working memory so you need to have some device if you want some subsystem they can store a relatively large number of factual episodic information for you know a reasonable amount of time.

planning is a form of reasoning and perhaps what led to the ability of humans to reason is the fact that or you know species you know that appear before us had to do some sort of planning to be able to hunt and survive and survive the winter in particular.

one suggestion with geoff hinton has been advocating for many decades is replace symbols by vectors think of it as pattern of activities in a bunch of neurons or units or whatever you wanna call them and replace logic by continuous functions.

I don't really believe in patterns for this kind of stuff yes so that's that's a great thing so I tell you a war story yeah you so what happens was the the first the first pattern of a condition that was about kind of the early version Congress on that that didn't have separate pudding layers it had the conditional layers which tried more than one if you want right and then there was a second one on commercial nets with separate pudding layers train with back probably in 89 and 1992 something like this at the time the life life of a pattern was 17 years.