Error driven learning

Error-driven learning is a powerful form of learning that drives changes in synaptic weights to reduce errors. In Axon, these errors are generally prediction errors: the differences between a prediction and what actually happens (see predictive learning). Prediction-error-driven learning is what drives learning in the large language models (LLMs) that power ChatGPT and related models.

Error-driven learning requires a mechanism for these error signals to drive synaptic changes. In LLMs and other current abstract neural network models, error backpropagation is used to drive synaptic learning. However, this mechanism is not directly compatible with known neurobiological mechanisms. Thus, Axon uses temporal derivatives, which can be computed locally using biologically based mechanisms in the context of bidirectional connectivity, to drive error-driven learning using the kinase algorithm.

Because error-driven learning is mathematically derived to minimize performance errors on any given task, it provides a very general and flexible form of learning. By contrast, Hebbian learning, which is more obviously supported by the known synaptic plasticity mechanisms, is mathematically related to principal components analysis, and generally learns to encode the strong correlational structure present in its inputs. Empirically, it does not learn to solve arbitrary tasks, and is thus not a plausible basis for a general-purpose learning mechanism in the neocortex.

See combinatorial vs conjunctive for an analysis of the tradeoffs at play in error-driven learning, that help clarify the relationship to Hebbian learning.