TL;DR
This paper introduces "Spontaneous Neurons" (SPON), a novel mechanism to combat accuracy degradation in large language models (LLMs) when using activation sparsity for faster inference. SPON injects lightweight, input-independent activation vectors that act as stable anchors, preserving performance and generalization without sacrificing efficiency.
What Problem Does This Paper Address?
Large Language Models (LLMs) are computationally expensive, especially during inference. Activation sparsity, which selectively suppresses hidden activations, offers a promising way to speed up inference. However, existing methods for achieving high activation sparsity often lead to a significant drop in accuracy. The authors argue this is due to "representational instability": sparsity disrupts the input-dependent activations learned during pretraining, causing shifts in the model's internal states. This makes it hard for the model to retain knowledge and generalize. The paper aims to solve this by stabilizing these representations under sparsity.
Methods
The core contribution is Spontaneous Neurons (SPON). SPON introduces a small set of learnable, input-independent activation vectors into each module of an LLM. These vectors act as "representational anchors," providing a stable baseline that complements the input-dependent activations.
Key aspects of SPON:
- Biologically Inspired: Modeled after spontaneous neural activity in biological systems.
- Lightweight: Adds minimal trainable parameters.
- Input-Independent: These activations are not directly influenced by the input sequence, providing a consistent anchor.
- Distribution Matching: SPON vectors are trained to align the sparse model's output distribution with that of the dense (non-sparse) model. This is achieved by minimizing the KL divergence between their output logits.
- Inference-Neutral: After training, the SPON activations can be folded into the bias terms of the linear layers, incurring no additional inference overhead.
The authors also provide a theoretical analysis showing that SPON acts as a Fisher-weighted error correction mechanism, aligning the sparse model's residual with the dense model's output distribution.
Datasets
- Calibration/Training: Wikitext-raw-v2 is used for training the spontaneous neurons.
- Evaluation:
- Language Modeling: WikiText.
- Zero-Shot Generalization: CommonsenseQA, TruthfulQA, OpenBookQA, MedMCQA, MMLU, and MathQA.
- LLM Backbones: Llama3 (1B, 8B, 70B), Mistral-7B, and Qwen3 (8B, 32B).
Key Insights and Ideas
- Representational Instability is the Root Cause: The paper strongly posits that the accuracy drop in sparse LLMs is not due to information loss per se, but due to the disruption of stable, input-dependent representations.
- Spontaneity as a Stabilizer: Borrowing from neuroscience, the idea of "resting" or "spontaneous" activity provides a powerful analogy for how to maintain stability in a dynamic, sparse system.
- Anchors for Sparse Computation: SPON vectors act as persistent anchors, preventing the representational drift that occurs when too many neurons are deactivated.
- Bias Terms are Not Enough: The paper demonstrates that simply learning bias terms (like in BitFit) is less effective than learning spontaneous activations in the activation space. This is because learning in the activation space allows for a more direct alignment with the pretrained feature subspace.
- Orthogonal to Existing Methods: SPON is presented as a complementary technique that can be combined with other sparsity methods (like TEAL) or even pruning and quantization for further gains.
Key Results
- Performance Restoration: SPON consistently restores performance across various LLM backbones (Llama3, Mistral, Qwen3) and sparsity levels (up to 60%), significantly outperforming baselines like TEAL. For example, at 50% sparsity, SPON achieved an average improvement of 1.178% across tasks.
- Preserves Efficiency: SPON introduces negligible inference overhead, maintaining the speed-ups achieved by activation sparsity.
- Stabilizes Representations: Analysis using Centered Kernel Alignment (CKA) and t-SNE visualizations shows that SPON significantly reduces representational drift and improves alignment with dense models.
- Scalability: SPON shows consistent gains even on larger models (Llama3-70B, Qwen3-32B).
- Robustness: SPON also demonstrates benefits when combined with quantization and network pruning, suggesting its utility in broader model compression scenarios.
Limitations and Future Work
- Computational Cost of Training: While SPON is inference-neutral, the training process for SPON involves an additional calibration step.
- Optimal Placement: While the paper explores where to inject SPON (MLP down-projection layers being most effective), further research could refine optimal placement strategies.
- MoE Integration: The paper presents preliminary studies on Mixture-of-Experts (MoE) models, suggesting SPON complements expert sparsity. However, more advanced integration of SPON with MoE routing during pretraining is left as future work.
- Specialized Kernels: Realizing the full hardware potential of combined quantization, pruning, and SPON would require specialized kernels.
Summary updated