Reproduce “Recurrent Neural Networks Learn to Store and Generate Sequences using Non-Linear Representations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Róbert Csordás, Christopher Potts, Christopher D. Manning, Atticus Geiger
Small Gated Recurrent Units (GRUs) learn a non-linear, magnitude-based "onion representation" to store sequential information for a repeat task, requiring autoregressive decoding. This discovery challenges the prevalent Linear Representation Hypothesis in mechanistic interpretability by demonstrating alternative encoding strategies when model capacity is limited.
Key insights
- Neural networks can learn non-linear "onion representations" where token identity is encoded by direction and position by magnitude, requiring an autoregressive "unpeeling" process for decoding.
- The choice between linear and non-linear representations is influenced by model capacity: smaller GRUs primarily leverage non-linear magnitude-based encodings, while larger GRUs prefer linear subspace representations.
- Autoregressive decoding is crucial for models employing onion representations, as their ability to solve the task significantly degrades without this iterative feedback mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.