A quick take on Yann LeCun’s Path towards Autonomous Machine Intelligence.

This is just a quick take (high-level summary) on Yann LeCun’s paper “A Path Towards Autonomous Machine Intelligence” which is available here https://openreview.net/pdf?id=BZ5a1r-kVsf.

Thoroughly impressed with the newly published paper by Yann LeCun. He breaks down the overall architecture as follows:

1. The configurator module performs executive control:
2. The perception module receives signals from sensors and estimates the current state of the world.
3. The world model module constitutes the most complex piece of architecture. Its role is twofold: (1) to estimate missing information about the state of the world not provided by perception, and (2) to predict plausible future states of the world.
4. The cost module computes a single scalar output that predicts the level of discomfort of the agent.
5. The actor module computes proposals for action sequences.
6. The short-term memory module keeps track of the current and predicted world state, as well as associated costs.

The paper is clever to push the aleatoric (type 1 to 3) and epistemic (type 1 to 4) uncertainty of the world away from a modeling distribution like in RL (Which models the manifolds as action, critic, and policies on a Bellman curve with generalized reward) to a latent variable that can be optimized or sampled.

This “reparameterization trick” is proposed to be achieved through a Hierarchical Joint Embedding Predictive Architecture (H-JEPA). The JEPA captures the dependencies between two inputs, x and y, and these inputs x and y are fed to trainable encoders that extract abstract representations of them. A predictor module is then trained to predict the abstract representation using a latent variable, z.

To date, only contraction methods were used to train such a JEPA which does not work for high-dimensional models. But LeCun’s paper on VICReg makes it possible to train high-dimensional models using regularized methods (Strongly suggest reading this)

The open question remains on how to precisely regularize the latent variable so as to minimize entropy. I find it to be clever that LeCun has considered Reasoning as an Energy-Based Model and has considered using LVEBMs (Latent Variable Energy Based Models).

Specific Thoughts.

1) A lot of thought has gone into building world models which is the most complex piece of architecture. It is not clear if such a world model representation is possible through H-JEPA and similar methods. Category Theory in Machine Learning holds some key prospects in “Invariant and Equivariant Learning” through Functorial Unsupervised Learning that can be explored.

2) The Role of Multi-Agent Reinforcement Learning is not clear in the proposal. The model heavily relies on optimal control as against Self Supervised Learning, especially in Large Language Models.

I will blog more detailed thoughts on this as I study more deeply.