nazmiefearmutcu/fuze

Fuze is a research project tackling a basic problem in machine learning: when a neural network learns a new skill, it tends to overwrite and forget the old ones. This is called catastrophic forgetting. The project's stated goal is to build a learning system for humanoid robots, which need to keep acquiring new skills over years of operation without losing earlier ones, without storing private data, and without being told when one task ends and another begins. The two common fixes in the field each carry a hidden dependency. Experience replay keeps a buffer of old training data and mixes it back in, which means retaining raw data forever. Elastic Weight Consolidation runs an importance calculation and locks certain weights at task boundaries, which means someone has to tell the system where tasks end. Fuze removes both dependencies. It stores no data and needs no task labels. The core mechanism is a small gate on each synapse, about 55 lines of code. Each connection tracks three numbers: a consolidation reserve and two running statistics of its own surprise, meaning how unexpected its recent errors were. When a synapse is predicting well, it consolidates and locks itself. When it gets hit by genuinely novel error, it unlocks and becomes able to learn again. The system discovers task shifts from the data stream itself, with no external signal. The repository includes a benchmark called the Amnesia Test, where four simulated robots learn the same four locomotion skills in sequence. Fuze retains all four skills, while replay and EWC each retain three. The authors then run an audit: they strip replay of its stored data and EWC of its task labels, and both collapse to plain forgetting. Fuze stays at four because it has nothing to remove. The README is candid about limitations. On streams of compatible, non-conflicting tasks, replay wins outright. Fuze trades raw accuracy for stability, losing on average error while winning on forgetting metrics. The full README is longer than what was shown.