Research map / Architectures and efficiency / RNNs and normalization
Towards Memory-Efficient Neural Networks via Multi-Level In Situ Generation
Jiaqi Gu, Hanqing Zhu, Chenghao Feng, Mingjie Liu, Zixuan Jiang, Ray T. Chen, David Z. Pan
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 orthogonal regularization (ablation) | in situ generation, ResNet-18 · CIFAR-10 (B_i=2, B_c=44, r=0.05) | accuracy: 92.46% -> 92.88% | Improved |
| add l2 initialization from the teacher (ablation) | in situ generation, ResNet-18 · CIFAR-10 (B_i=2, B_c=44, r=0.05) | accuracy: 92.46% -> 93.18% | Improved |
| use l2 initialization instead of truncated SVD initialization (ablation) | in situ generation, ResNet-18 · CIFAR-10 (B_i=2, B_c=44, r=0.05) | accuracy: 93.05% -> 93.18% | Improved |
| add knowledge distillation to l2+Ortho (ablation) | in situ generation, ResNet-18 · CIFAR-10 (B_i=2, B_c=44, r=0.05) | accuracy: 93.17% -> 93.29% | Improved |
| add the full recipe (l2 init, Ortho, KD) over baseline (ablation) | in situ generation, ResNet-18 · CIFAR-10 (B_i=4, B_c=28, r=0.05) | accuracy: 91.98% -> 93.19% | Improved |
| add the full recipe over baseline at higher compression (ablation) | in situ generation, ResNet-18 · CIFAR-10 (B_i=3, B_c=17, r=0.025) | accuracy: 90.62% -> 91.52% | Improved |
| add orthogonal regularization with weight 0.01-0.05 (ablation) | in situ generation, ResNet-18 · CIFAR-10 | accuracy: 0.5%-1% improvement | Improved |
| add orthogonal regularization to l2 init (ablation) | in situ generation, ResNet-18 · CIFAR-10 (B_i=3, B_c=17, r=0.025) vs (B_i=8, B_c=8, r=0.025) | accuracy: 91.32% -> 91.40%; 88.85% -> 88.65% | Mixed |
| use mixed-precision bitwidths (8,8,8) instead of 32-bit bases | in situ generation, ResNet-18 · CIFAR-10 | accuracy: 93.29% -> 93.79% | Improved |
| use relatively balanced rather than extremely imbalanced mixed precision (ablation) | in situ generation, ResNet-18 · CIFAR-10 | accuracy: (2,5,3), (4,5,6) outperform (5,1,8), (2,4,8) | Improved |
| replace BSConv with multi-level in situ generation (3-6-3 bits) | ResNet-18 · CIFAR-10 | accuracy: 93.10% -> 93.72% | Improved |
| replace Conv with in situ generation | DenseNet-121 · CIFAR-100 | accuracy: 76.51% -> 75.59% | Hurt |
Closest papers by meaning
- Harnessing Manycore Processors with Distributed Memory for Accelerated Training of Sparse and Recurrent Models (AAAI 2024 · similarity 0.85)
- Analog Foundation Models (NeurIPS 2025 · similarity 0.85)
- MEET: Towards Memory-Efficient Temporal Sparse Deep Neural Networks (CVPR 2025 · similarity 0.85)
- FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · similarity 0.85)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.84)
- SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation (NeurIPS 2024 · similarity 0.84)
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.
