Search papers, labs, and topics across Lattice.
This paper investigates the emergence of geometric continuity in deep neural networks, finding that residual connections and symmetry-breaking nonlinearities are key drivers. Residual connections induce cross-layer gradient coherence, aligning weight updates, while symmetry-breaking nonlinearities prevent rotational drift, stabilizing weight structure. Through controlled experiments on MLPs and transformers, the authors demonstrate that symmetry breaking, rather than nonlinearity alone, is crucial for geometric continuity, and that different projections in transformers exhibit distinct continuity properties based on their interaction with the residual stream.
Geometric continuity in deep networks isn't just a byproduct of depth, but an actively sculpted property arising from the interplay of residual connections and symmetry-breaking activations.
Weight matrices in deep networks exhibit geometric continuity -- principal singular vectors of adjacent layers point in similar directions. While this property has been widely observed, its origin remains unexplained. Through experiments on toy MLPs and small transformers, we identify two mechanisms: residual connections create cross-layer gradient coherence that aligns weight updates across layers, and symmetry-breaking nonlinearities constrain all layers to a shared coordinate frame, preventing the rotation drift that would otherwise destabilize weight structure. Crucially, a nonlinear but rotation-preserving activation fails to retain continuity, isolating symmetry breaking -- not nonlinearity itself -- as the active ingredient. Activation and normalization play distinct roles: activation concentrates continuity in the leading singular direction, while normalization distributes it across multiple directions. In transformers, continuity is projection-specific: Q, K, Gate, and Up (which read from the residual stream) develop input-space ($\mathbf{v}_1$) continuity; O and Down (which write to it) develop output-space ($\mathbf{u}_1$) continuity; V alone, lacking an adjacent nonlinearity, develops only low continuity.
