Research map / Architectures and efficiency / RNNs and normalization
Dynamic Symmetric Point Tracking: Tackling Non-ideal Reference in Analog In-memory Training
Quan Xiao, Jindan Li, Zhaoxian Wu, Tayfun Gokmen, Tianyi Chen
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 |
|---|---|---|---|
| estimate SP dynamically during training (E-RIDER) instead of fixed reference | TT-v2 with zero reference · LeNet-5 MNIST, ref mean 0.4, std 1.0 | test accuracy: 67.08 -> 88.02 | Improved |
| add residual learning with moving-average SP tracking, chopping and mixed-weight gradient (E-RIDER) over AGAD | AGAD · FCN MNIST, ref mean 0.4, std 1.0 | test accuracy: 90.43 -> 95.14 | Improved |
| raise reference std from 0.05 to 1.0 (nonzero SP offset) (ablation) | TT-v2 · FCN MNIST, ref mean 0.4 | test accuracy: 72.00 -> 57.79 | Hurt |
| raise reference std from 0.05 to 1.0 (nonzero SP offset) (ablation) | E-RIDER · FCN MNIST, ref mean 0.4 | test accuracy: 96.07 -> 95.14 | Hurt |
| estimate SP with ZS using fewer pulses (N) before training (ablation) | two-stage residual learning with ZS · LeNet-5 MNIST (TT-v1) training loss | training loss | Hurt |
| use ZS with N = 4000 pulses on a 2000-state device (ablation) | two-stage ZS · LeNet-5 MNIST | reaching training loss 0.2 | Hurt |
| use E-RIDER instead of two-stage ZS | TT-v2 with ZS · LeNet-5 MNIST | total pulse cost to reach loss 0.2 | Improved |
| decrease response granularity (higher-precision device) for ZS calibration (ablation) | ZS algorithm · ReRAM array | pulses for 1% error | Hurt |
| use small chopper probability p > 0 (ablation) | E-RIDER · MNIST FCN | test accuracy | Improved |
| evaluate E-RIDER on ResNet-18 CIFAR-100 with large SP offset | TT-v2 and AGAD · CIFAR-100 ResNet-18 (80 epochs) | test accuracy | Improved |
Closest papers by meaning
- Analog In-memory Training on General Non-ideal Resistive Elements: The Impact of Response Functions (NeurIPS 2025 · similarity 0.89)
- Analog Foundation Models (NeurIPS 2025 · similarity 0.85)
- Towards Exact Gradient-based Training on Analog In-memory Computing (NeurIPS 2024 · similarity 0.84)
- Two-stream Beats One-stream: Asymmetric Siamese Network for Efficient Visual Tracking (AAAI 2025 · similarity 0.83)
- Reshape and Adapt for Output Quantization (RAOQ): Quantization-aware Training for In-memory Computing Systems (ICML 2024 · similarity 0.83)
- RTracker: Recoverable Tracking via PN Tree Structured Memory (CVPR 2024 · similarity 0.82)
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.
