On Learning Expand-and-Sparsify Representations

Sparse autoencoders (SAEs) decompose internal activations of neural networks into sparse linear combinations of learned features by fitting an overcomplete dictionary $\mathbf{W}\in\mathbb{R}^{m\times n}$ with $m<n$, and inferring a sparse code $\mathbf{x}\in\mathbb{R}^n$ from $\mathbf{h}\approx\mathbf{W}\mathbf{x}$. This inference problem closely resembles the canonical setup of compressed sensing, but dense decoders requires $O(mn)$ learned values, which becomes costly at large feature counts. We introduce Expander SAEs: TopK SAEs whose decoder and tied encoder are supported on a left-$d$-regular expander mask with $d\ll m$, learning only $dn$ decoder values while keeping the sparse-coding problem $(m,n,k)$ fixed. The same structure reduces storage and turns the matching-pursuit correlation step $\mathbf{W}^\top \mathbf{r}$ in OMP into an $O(dn)$ gather-and-reduce operation. Our experiments show that across Pythia-70M/160M, Qwen2.5-3B, and Llama-3.2-1B residual-stream activations, varying $d$ traces a consistent storage--fidelity frontier, and that at the most compressed modern-LM setting, Qwen2.5-3B with $d=7$ uses $293\times$ fewer learned decoder values than the full dense decoder while retaining $84$% of dense CE-loss recovered. Control experiments show that the improved storage--fidelity tradeoff is driven by sparse, diverse decoder support structure rather than by fewer learned decoder values, and that when sparse and dense decoders are compared at matched parameter count, part of the remaining gap comes from encoder amortisation. On the theoretical side, we show that expansion and column flatness are sufficient for identifiability of noiseless $k$-sparse codes, and we derive complementary sufficient conditions under which OMP recovers the support exactly.

Predictive systems increasingly span heterogeneous modalities such as graphs, language, and tabular records, but sparsity and efficiency remain modality-specific (graph edge or neighborhood sparsification, Transformer head or layer pruning, and separate tabular feature-selection pipelines). This fragmentation makes results hard to compare, complicates deployment, and weakens reliability analysis across end-to-end KDD pipelines. A unified sparsification primitive would make accuracy-efficiency trade-offs comparable across modalities and enable controlled reliability analysis under representation compression. We ask whether a single representation-level mechanism can yield comparable accuracy-efficiency trade-offs across modalities while preserving or improving probability calibration. We propose L0-Gated Cross-Modality Learning (L0GM), a modality-agnostic, feature-wise hard-concrete gating framework that enforces L0-style sparsity directly on learned representations. L0GM attaches hard-concrete stochastic gates to each modality's classifier-facing interface: node embeddings (GNNs), pooled sequence embeddings such as CLS (Transformers), and learned tabular embedding vectors (tabular models). This yields end-to-end trainable sparsification with an explicit control knob for the active feature fraction. To stabilize optimization and make trade-offs interpretable, we introduce an L0-annealing schedule that induces clear accuracy-sparsity Pareto frontiers. Across three public benchmarks (ogbn-products, Adult, IMDB), L0GM achieves competitive predictive performance while activating fewer representation dimensions, and it reduces Expected Calibration Error (ECE) in our evaluation. Overall, L0GM establishes a modality-agnostic, reproducible sparsification primitive that supports comparable accuracy, efficiency, and calibration trade-off analysis across heterogeneous modalities.

Sparse autoencoders (SAEs) are typically trained to reconstruct the entire residual stream through a sparse dictionary, implicitly assuming that all activation content is amenable to sparse, monosemantic decomposition. We question this assumption and hypothesize that activations contain a low-rank, dense component that is computationally important to the model yet inherently unsuitable for sparse representation, which serves as a major source of the persistent dense latents widely observed in trained SAEs. To test this, we add a small rank-$r$ linear bottleneck in parallel with standard SAEs (BatchTopK and Matryoshka), allowing dense structure to be absorbed before sparse reconstruction. On Gemma-2-2B layer 12, a rank-24 bottleneck reduces dense latent count by up to 84% while improving sparse probing and targeted probe perturbation on both architectures at matched sparsity. The absorbed component is (i) structurally identifiable as the top principal components and outlier dimensions; (ii) causally necessary, with removing it raising next-token cross-entropy by 7.5$\times$, far exceeding the 2.8$\times$ from removing the geometrically near-identical top-24 PCA directions; and (iii) redundantly encoded by sparse dictionaries, with ablating 787 maximally aligned sparse features raising cross-entropy by only 2.9$\times$ and ablating 2,048 topic-aligned features leaving MMLU topic classification virtually unchanged, whereas removing the scaffold drops it from 98.7% to chance. Together, our findings identify a compact, semantically informative and causally important component of residual stream activations (which we term a computational scaffold) that standard sparse dictionaries represent inefficiently, suggesting that the scope of sparsity-based interpretability methods warrants careful re-examination.