Reproduce “Stochastic Gradient Descent-Induced Drift of Representation in a Two-Layer Neural Network” 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
Farhad Pashakhanloo, Alexei Koulakov
Research by Pashakhanloo and Koulakov establishes that the inherent noise of Stochastic Gradient Descent (SGD) drives representational drift in a two-layer linear neural network during continual online learning. Their analysis shows that representations of more frequent stimuli drift at slower rates, aligning with experimental findings in biological systems like the piriform cortex.
Key insights
- Stochastic Gradient Descent (SGD) sampling noise inherently drives representational drift in neural networks, even after achieving optimal task performance.
- Representational drift is a second-order phenomenon, meaning it is directly proportional to temporary deviations (fluctuations) from the minimum-loss manifold in parameter space.
- The rate of representational drift is stimulus-dependent, with more frequently presented stimuli exhibiting both smaller fluctuations and slower drift rates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.