Reproduce “Separating the what and how of compositional computation to enable reuse and continual learning” on OpenResearch

About this paper

Haozhe Shan, Sun Minni, Lea Duncker

ContextRNN learns task epochs from inputs and target responses, then uses inferred context to select or blend low-rank components in a separate recurrent network. On MemoryAnti, after pretraining on three tasks, it reached 83% test accuracy within an average of 40 trials, while the downstream network remained fixed. This computational result concerns a designed task family, not evidence about brains.

Key insights

• Explicitly separating the inference of "what" computation to perform (context) from "how" to perform it (RNN dynamics) is key for effective continual learning.
• An online, incremental probabilistic model can successfully learn and infer the compositional structure of tasks, including shared latent task epochs.
• Dynamically modulating RNN weights with inferred contextual beliefs allows the network to activate specific computational modules, preventing catastrophic forgetting and enabling skill reuse.

Summary by alphaXiv.