Greedy dynamical meta-learning
2026-07-27 • Machine Learning
Machine Learning
AI summaryⓘ
The authors explain that while gradient descent works well for big models, it can become unstable when used for a long time. Other methods that don't use gradients have trouble dealing with many variables. To handle this, the authors created a method where a model learns to change its own parts to improve itself. They use a two-step process: first, the model adjusts itself in many ways, and then a simpler optimization adjusts how the model makes those adjustments. This approach works well since the outer step only needs to optimize a few things.
gradient descentmeta-learninggradient-free optimizationhigh-dimensional optimizationzeroth-order methodsinner loopouter loopself-modifying modelsoptimization algorithms
Authors
Aria Yom
Abstract
Gradient descent scales well to large models, but becomes unstable over long time horizons. Gradient-free optimizers can scale to arbitrary timespans, but are hobbled by high dimensions. Since learning occurs in large models over long timescales, neither of these approaches is likely to produce traits which can accelerate the learning process. Instead, we propose a meta-learning algorithm in which the agent learns to modify its own weights and biases. Our algorithm consists of an inner loop, wherein the agent performs some high-dimensional optimization upon itself, and an outer loop, wherein we perform some low-dimensional optimization upon the inner loop. Since the outer loop handles very few parameters, standard zeroth-order methods may be used.