Research map / Architectures and efficiency / RNNs and normalization
Large Memory Layers with Product Keys
Guillaume Lample, Alexandre Sablayrolles, Marc'Aurelio Ranzato, Ludovic Denoyer, Herve Jegou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add product-key memory to a 12-layer, 1024-dimension Transformer | Transformer · CC-News test | perplexity: 17.6566 -> 15.6184 | Improved |
| Add product-key memory to a 12-layer, 1600-dimension Transformer | Transformer · CC-News test | perplexity: 14.9994 -> 13.6522 | Improved |
| Add query BatchNorm to a memory of size 16384 (ablation) | Product-key memory · 6-layer transformer memory ablation | perplexity: 22.7972 -> 22.9658 | Hurt |
| Add query BatchNorm to a memory of size 262144 (ablation) | Product-key memory · 6-layer transformer memory ablation | perplexity: 20.4776 -> 19.7805 | Improved |
| Add query BatchNorm to a memory of size 1048576 (ablation) | Product-key memory · 6-layer transformer memory ablation | perplexity: 19.7913 -> 17.9819 | Improved |
| Increase memory size from 16384 to 1048576 (ablation) | Product-key memory · 6-layer transformer memory ablation | perplexity: 22.7972 -> 17.9819 | Improved |
| Place memory at intermediate layer 5 rather than layer 1 (ablation) | 6-layer transformer · 6-layer transformer memory ablation | perplexity: 21.4669 -> 19.7805 | Improved |
| Replace flat keys with product keys (ablation) | Memory layer · Memory size 16384 | perplexity: 23.2152 -> 22.9658 | Improved |
| Replace flat keys with product keys (ablation) | Memory layer · Memory size 65536 | perplexity: 22.5859 -> 21.9114 | Improved |
| Replace flat keys with product keys (ablation) | Memory layer · Memory size 147456 | perplexity: 22.1477 -> 20.7050 | Improved |
| Replace flat keys with product keys (ablation) | Memory layer · Memory size 147456 | inference speed (w/s): 13.9k -> 36.4k | Improved |
| Increase the number of memory heads or selected neighbors (ablation) | Product-key memory · 6-layer transformer memory ablation | perplexity | Improved |
Closest papers by meaning
- Memory Layers at Scale (ICML 2025 · similarity 0.89)
- MemoryFormer : Minimize Transformer Computation by Removing Fully-Connected Layers (NeurIPS 2024 · similarity 0.88)
- Memory-Augmenting Decoder-Only Language Models through Encoders (Student Abstract) (AAAI 2024 · similarity 0.87)
- Search for Efficient Large Language Models (NeurIPS 2024 · similarity 0.86)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.86)
- Memory Decoder: A Pretrained, Plug-and-Play Memory for Large Language Models (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
