Not all neurons in LLMs work hard - and that's actually good!
This paper studies how to make LLMs naturally sparse without compromising performance. Understanding why some LLM neurons naturally go silent during training
A systematic study showing what makes LLMs develop efficient neuron patterns
https://arxiv.org/abs/2411.02335
🎯 Original Problem:
LLMs have significant activation sparsity (many neurons contribute weakly to outputs), which could be leveraged for efficiency and interpretability. However, we lack understanding of how this sparsity scales with model size, training data, and architecture choices.
-----
🔧 Methods used in this Paper:
→ Introduced PPL-p% sparsity - a new metric that measures activation sparsity while considering model performance impact
→ Studied how sparsity changes with training data volume, activation functions (ReLU vs SiLU), and model architecture
→ Analyzed models from 0.1B to 1.2B parameters, tracking sparsity patterns during training
-----
💡 Key Insights:
→ ReLU activation leads to increasing sparsity with more training data, while SiLU shows opposite trend
→ Deeper models (lower width-depth ratio) achieve better sparsity at fixed parameter scale
→ Sparsity limit varies weakly with model scale when width-depth ratios are similar
→ Smaller models reach their sparsity limit faster during training
-----
📊 Results:
→ ReLU models achieve higher sparsity than SiLU while maintaining comparable performance
→ PPL-1% sparsity setting shows no significant performance degradation across tasks
→ Linear increase in activation ratio with width-depth ratio up to a bottleneck point
→ Models show convergent power-law relationships between activation ratio and training data
------
Are you into AI and LLMs❓ Join me on X/Twitter with 48K+ others, to remain on the bleeding-edge of AI every day.
𝕏/🐦 https://x.com/rohanpaul_ai
Share this post