Greedy dynamical meta-learning

The Problem with Gradients

Modern deep learning relies heavily on gradient descent, which works well for static models but struggles with long-term dependencies. As learning occurs over time, the gradient of a dynamical system often becomes unstable due to the butterfly effect, leading to exploding or vanishing gradients. The authors argue that because intelligence is inherently a process of learning over long timescales, gradient-based methods are fundamentally limited in their ability to optimize the learning process itself.

The DSML Framework

To address these limitations, the authors introduce Dynamical Systems Meta-Learning (DSML). Instead of relying on a trainable policy (as in meta-RL) or static weight updates, DSML treats the agent as a stochastic dynamical system where weights and activations evolve according to fixed laws of motion. The core idea is to generate a population of 'mutants'—variations of the agent's state—and evaluate them over specific time horizons to select the most promising trajectories. By selecting for high-performing agents over long periods, the system implicitly learns to generate beneficial mutations and preserve learned structures.

Balancing Timescales

Central to DSML is the distinction between the mutation timescale (the time at which an agent reaches its peak intelligence) and the evaluation timescale (the time at which performance is most distinguishable). The authors demonstrate that these two are often misaligned: by the time an agent's performance is clearly superior, it may have already passed its peak intelligence. The DSML algorithm requires careful calibration of five parameters—mutation time, evaluation time, and various sampling counts—to ensure that the system successfully ascends the fitness landscape without collapsing back into low-intelligence attractors.

This paper challenges the current paradigm of using gradient-based optimization for all aspects of machine learning. By shifting the focus toward evolutionary strategies applied to dynamical systems, the authors suggest a path toward agents that can autonomously discover how to learn, potentially overcoming the 'predictability horizon' that currently limits the depth and duration of neural network training.