Fourier analysis of neural activations

Method used to identify the periodic features and their periods in Llama-3.1-8B's MLP neurons

Neighborhood — ranked by edge-count

concept

Fourier features
implements
Features identified in Llama-3.1-8B that compute sums using periods respecting base-10 addition (2, 5, 10) rather than concept-specific periods

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Disjoint neuron clusters for Fourier periodsconcept0.758
The 28 identified neurons can be partitioned into disjoint clusters each computing a different Fourier period sum
Fourier features with period 10 contribute to base-10 sum computation in the 28-neuron clusterfinding0.748
One of the three base-10 Fourier periods identified in the sparse neuron set
Does the geometric structure of activation space causally shape neural network behavior?question0.742
Central research question driving the work.
Fourier Features for Numerical Computationconcept0.737
Neural Representations of Location Composed of Spatially Periodic Bands (Krupic et al., 2012)concept0.734
Discovery of band cells; TEM-t also recapitulates these representations.
neural substratesconcept0.733
Brain-based physical implementations of consciousness-related functions, assumed by many ToCs to be exclusive.
Neural Annealingframework0.730
Michael Johnson's prior work on how neural networks (and brains) can be 'annealed' to find optimal states.
How do we establish bidirectional causal relationships between neural systems?question0.726
Motivates the bidirectional design of MAS over unidirectional model stitching.