Research map / Architectures and efficiency / Pruning and efficient tuning
Memory-Efficient Transfer Learning with Fading Side Networks via Masked Dual Path Distillation
Yutong Zhang, Jiaxin Chen, Honglin Chen, Kaiqi Zheng, Shengcai Liao, Hanwen Zhong, Weixin Li, Yunhong Wang
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 dual-path (logits plus feature) distillation instead of logits-based S to B only (ablation) | MDPD (UniPT-based) · Flickr30K ITR | Rsum: 529.7 -> 538.1 | Improved |
| add dual-path distillation instead of feature-based B to S only (ablation) | MDPD (UniPT-based) · Flickr30K ITR | Rsum: 531.5 -> 538.1 | Improved |
| use backbone-to-side feature direction instead of side-to-backbone (ablation) | MDPD (UniPT-based) · Flickr30K ITR | Rsum: 518.2 -> 531.5 | Improved |
| fade the side network at inference (ablation) | MDPD (UniPT-based) · Flickr30K ITR | QPS: 7.98 -> 14.71 | Improved |
| fade the side network at inference (ablation) | MDPD (UniPT-based) · Flickr30K ITR | Rsum: 541.9 -> 541.2 | Hurt |
| fade the side network at inference (ablation) | MDPD (SHERL-based) · Flickr30K ITR | Rsum: 543.9 -> 543.0 | Hurt |
| use convolutional generation block instead of self-attention (ablation) | MDPD (UniPT-based) · Flickr30K ITR | Rsum: 537.4 -> 541.2 | Improved |
| distill features across all layers instead of layer 1 only (ablation) | MDPD (UniPT-based) · Flickr30K ITR | Rsum: 536.8 -> 541.2 | Improved |
| apply MDPD on top of UniPT | UniPT · ITR BERT+BUTD | QPS: 12.75 -> 24.05 | Improved |
| apply MDPD on top of SHERL | SHERL · Flickr30K ITR | QPS: 10.24 -> 14.83 | Improved |
| apply MDPD on top of SHERL | SHERL · GLUE (T5-base) | inference time (ms): 350.4 -> 234.8 | Improved |
| apply MDPD on top of SHERL | SHERL · GLUE (T5-base) | mean score: 84.3 -> 84.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- UniPT — UniPT: Universal Parallel Tuning for Transfer Learning with Efficient Parameter and Memory (CVPR 2024) · better on 7 of 7 shared results
- SHERL — SHERL: Synthesizing High Accuracy and Efficient Memory for Resource-Limited Transfer Learning (ECCV 2024) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DTL: Disentangled Transfer Learning for Visual Recognition (AAAI 2024 · similarity 0.86)
- Progressively Knowledge Distillation via Re-parameterizing Diffusion Reverse Process (AAAI 2024 · similarity 0.85)
- Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network (NeurIPS 2026 · similarity 0.85)
- DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts (NeurIPS 2026 · similarity 0.84)
- Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks (ICLR 2025 · similarity 0.84)
- Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation (ICML 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.
