Yann LeCun on Neural Networks, Reasoning, and the Future of AI Architectures

Yann LeCun asserts that neural networks can indeed be made to reason, emphasizing that the core challenge lies in determining the appropriate prior structure to embed within them for human-like reasoning to emerge. He highlights a fundamental incompatibility between traditional, discrete, logic-based models of reasoning and gradient-based learning, which he strongly advocates. LeCun characterizes deep learning's mathematical underpinnings as closer to cybernetics and electrical engineering, contrasting it with the exactness of conventional computer science, famously calling machine learning "the science of sloppiness."

To achieve reasoning in AI, LeCun identifies the critical need for a "working memory" system, akin to the human hippocampus, capable of storing a large volume of factual and episodic information for a significant duration. He distinguishes this from the transient memory of the cortex and the long-term synaptic memory. While current architectures like transformers offer some memory through self-attention, they lack the necessary recurrence and scalability to represent vast knowledge bases like Wikipedia, necessitating novel architectural ideas for efficient associative memory access and writing. He also discusses energy minimization as a classical form of reasoning, particularly in planning and model predictive control, suggesting an evolutionary link between planning for survival and the development of reasoning abilities.

LeCun critiques symbolic AI approaches, such as expert systems and knowledge graphs, for their inherent brittleness and rigidity, and their reliance on human experts for knowledge acquisition, which he deems impractical. He points to the Mycin system and subsequent Bayesian networks as examples of this paradigm's limitations. Instead, he champions the idea, advocated by Geoff Hinton and elaborated by Leon Bottou, of replacing discrete symbols with continuous vectors (patterns of neural activity) and logic with continuous functions. This shift, he argues, makes reasoning compatible with gradient-based learning.

The discussion underscores a significant ongoing debate in AI research, particularly with figures like Gary Marcus, regarding the extent of prior structure required for reasoning to spontaneously emerge in learning systems. LeCun's perspective leans towards a continuous, learning-centric approach, drawing inspiration from biological memory systems and the evolutionary pressures that shaped human intelligence, while acknowledging the current architectural gaps that prevent large-scale, flexible reasoning in contemporary neural networks.

do you think neural networks can be made to reason yes there's no question about that

all of our kind of model of what reasoning is that are based on logic are discrete and and are therefore incompatible with gradient based learning

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

the longer term memory is in the synapse the synapses right so what you need if you want for a system that's capable of reasoning is that you want the hippocampus like thing

a chain of reasoning is a process by which you you you can you update your knowledge about the state of the world

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

knowledge graphs are your brittle or logic representation so basically you know variables that have values and constraint between them that are represented by rules is real too rigid and too brittle

the main issue with this is is knowledge acquisition how do you reduce a bunch of data to a graph of this type near relies on the expert on a human being to encode at add knowledge and that's especially in practical question

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