| name | sparse-linear-expert-transformer |
| description | Sparsely gated tiny linear experts (sgatlin) methodology — replacing transformer feedforward layers with networks of sparsely-gated linear neurons for improved compute efficiency and interpretability. Covers isoflop comparison, linear expert removal of nonlinearity, semantic cluster interpretation, and causal factual recall. Activation: sparse linear expert, sgatlin, tiny linear expert, sparsely gated linear, linear MoE, sparse transformer FFN. |
Context
Mixture-of-Experts (MoE) models achieve scalability through sparsity — routing to a subset of experts per token. However, individual experts typically remain large and dense. This paper demonstrates that further increasing sparsity by shrinking each expert to a single linear neuron and selecting a tiny fraction of many available neurons improves both compute efficiency and interpretability.
Source paper: arXiv:2606.07414 "Sparsely gated tiny linear experts" (Simon Schug — Jun 2026).
Core Methodology
1. Sparsely Gated Linear Neurons (sgatlin)
Key Insight: Removing the nonlinearity typically applied to MoE experts, while dramatically increasing the number of experts (each being a single neuron), yields better performance at the same compute budget.
- Replace all transformer feedforward (FFN) layers with sgatlin layers
- Each "expert" is a single linear neuron (no activation function)
- Sparse gating selects a tiny fraction of available neurons per token
- Total parameters scale with number of neurons, not with expert size
2. Isoflop Comparison Protocol
- Match compute budget (FLOPs) between baseline FFN and sgatlin
- Compare perplexity across language models at different compute budgets
- Key finding: sgatlin improves perplexity across all tested compute budgets
- The improvement comes from better parameter utilization through sparsity + linearity
3. Interpretability via Linearity + Sparsity
- Semantic clustering: Feedforward circuits in sgatlin form semantically structured clusters
- Causal implication: Individual linear neurons are causally implicated in factual recall
- No replacement models needed: Unlike standard FFN interpretation, sgatlin circuits can be interpreted directly without training additional replacement models
- The linearity of experts enables direct attribution analysis
4. Sparse Gating Mechanism
- Router computes scores over all available linear neurons
- Top-k selection (very small k relative to total neurons)
- Weighted sum of selected neurons produces the layer output
- Gating function is typically softmax or top-k with temperature
Implementation Steps
- Replace FFN with sgatlin: For each transformer layer, replace
FFN(x) = W2 * activation(W1 * x) with sgatlin(x) = Σ_{i∈topk} g_i(x) * (w_i^T x) * v_i
- Scale neuron count: Increase total neurons to match or exceed original parameter count
- Configure routing: Set top-k value (typically k << total_neurons) for desired sparsity level
- Isoflop matching: Ensure total FLOPs match baseline for fair comparison
- Training: Train with standard language modeling objective
- Interpretability analysis: After training, analyze semantic clusters via neuron weight inspection
Key Results
- sgatlin improves perplexity over standard FFN at matched FLOPs
- Feedforward circuits form semantically structured clusters
- Neurons are causally implicated in factual recall
- Direct interpretation without auxiliary replacement models
- Path toward compute-efficient and interpretable transformer FFN layers
Pitfalls
- Nonlinearity removal is critical: The key finding depends on removing the activation function from experts — adding it back degrades both performance and interpretability
- Routing collapse: With many tiny experts, routing may collapse to a subset — use entropy regularization on gating distribution
- Gradient flow: Linear experts may have different gradient dynamics — monitor training stability
- Interpretability is not automatic: Semantic clusters emerge but require careful analysis (weight inspection, causal intervention)
Verification
- Reproduce isoflop comparison: same FLOPs, different architectures
- Verify perplexity improvement at matched compute budget
- Check semantic clustering via neuron weight analysis
- Perform causal intervention on individual neurons for factual recall
Related Skills
emo-emergent-moe-modularity — modular Mixture-of-Experts design
unipool-shared-expert-moe — globally shared expert pool MoE
hierarchical-moe-detection — hierarchical MoE for object detection
moe-optimal-transport-routing — MoE routing via optimal transport
routing-distraction-multimodal-moe — routing analysis in multimodal MoE
Activation
sparse linear expert, sgatlin, tiny linear expert, sparsely gated linear, linear MoE, sparse transformer FFN, sparse gating interpretability, linear neuron expert, compute-efficient transformer