Research map / Architectures and efficiency / RNNs and normalization
Backprop with Approximate Activations for Memory-efficient Network Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Store approximate activations at 8-bit precision for backpropagation | ResNet-164 · CIFAR-10 | Test Set Error: 5.36%±0.15 -> 5.48%±0.13 | No clear effect |
| Store approximate activations at 4-bit precision for backpropagation | ResNet-164 · CIFAR-10 | Test Set Error: 5.36%±0.15 -> 5.49%±0.16 | No clear effect |
| Store approximate activations at 8-bit precision for backpropagation | ResNet-164 · CIFAR-100 | Test Set Error: 23.44%±0.26 -> 23.63%±0.32 | No clear effect |
| Store approximate activations at 4-bit precision for backpropagation | ResNet-164 · CIFAR-100 | Test Set Error: 23.44%±0.26 -> 23.58%±0.30 | No clear effect |
| Store approximate activations at 8-bit precision for backpropagation | ResNet-152 · ImageNet | Val Set Top-5 Error: 7.20% -> 7.70% | Hurt |
| Store approximate activations at 4-bit precision for backpropagation | ResNet-152 · ImageNet | Val Set Top-5 Error: 7.20% -> 7.72% | Hurt |
| Reduce feature channels to 1/4 per layer | ResNet-164 · CIFAR-10 | Test Set Error: 5.36%±0.15 -> 9.49%±0.12 | Hurt |
| Reduce feature channels to 1/4 per layer | ResNet-164 · CIFAR-100 | Test Set Error: 23.44%±0.26 -> 33.47%±0.50 | Hurt |
| Use quantized activations directly during the forward pass | ResNet-164 · CIFAR-10 | Test Set Error: 5.36%±0.15 -> 75.49%±9.09 | Hurt |
| Use quantized activations directly during the forward pass | ResNet-164 · CIFAR-100 | Test Set Error: 23.44%±0.26 -> 95.41%±2.16 | Hurt |
| Use 4-bit approximate activations | ResNet-1001 (4x) · CIFAR-10 | Maximum Batch-size: 26 -> 182 | Helped |
| Use 4-bit approximate activations | ResNet-1001 · CIFAR-10 | Maximum Batch-size: 134 -> 876 | Helped |
Closest papers by meaning
- Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation (ICML 2024 · similarity 0.91)
- Backpropagation-Free Deep Learning with Recursive Local Representation Alignment (AAAI 2023 · similarity 0.89)
- Learning with Auxiliary Activation for Memory-Efficient Training (ICLR 2023 · similarity 0.88)
- meProp: Sparsified Back Propagation for Accelerated Deep Learning with Reduced Overfitting (ICML 2017 · similarity 0.88)
- Accelerated training through iterative gradient propagation along the residual path (ICLR 2025 · similarity 0.87)
- Overcoming Rank Collapse in Feedback Alignment (NeurIPS 2026 · 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.