Research map / Continual learning / Online and replay-based
How Well Do Unsupervised Learning Algorithms Model Human Real-time and Life-long Learning?
Chengxu Zhuang, Violet Xiang, Yoon Bai, Xiaoxuan Jia, Nicholas Turk-Browne, Kenneth Norman, James J. DiCarlo, Daniel LK Yamins
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 SimCLR-style negative sampling to BYOL (BYOLNeg) (ablation) | BYOL · Real-time and life-long benchmarks | mismatch score / trajectory-averaged Mini-ImageNet accuracy | Improved |
| add SimCLR-style negative sampling to DINO (DINONeg) (ablation) | DINO ViT-S · Life-long benchmark | trajectory-averaged Mini-ImageNet accuracy | Improved |
| add SimCLR-style negative sampling to DINO (DINONeg) (ablation) | DINO ViT-S · Real-time benchmark | mismatch score to human | No clear effect |
| increase within-batch diversity (longer replay window, lower current-memory ratio) | all SSL algorithms · Life-long benchmark | trajectory-averaged Mini-ImageNet accuracy | Improved |
| use algorithms without negative sampling (SwAV, BYOL, SimSiam) instead of SimCLR/MoCo v2 in low-diversity streams | life-long continual learning · Life-long benchmark, low diversity | trajectory-averaged Mini-ImageNet accuracy | Hurt |
| use MAE instead of contrastive algorithms | life-long continual learning · Life-long benchmark across diversity conditions | trajectory-averaged Mini-ImageNet accuracy | No clear effect |
| make the current-memory mix ratio extreme: R=1:0 (only current context) | SimCLR-More-MLPs · Real-time benchmark | mismatch score to human | Hurt |
| increase current-context share up to R=7:1 from 1:1 | SimCLR-More-MLPs · Real-time benchmark | mismatch score to human | No clear effect |
| lower the current-context share (R=1:7, 1:15, 0:1) | SimCLR-More-MLPs · Real-time benchmark | mismatch score to human | Hurt |
| lengthen the current-context replay window to 40m or 80m from 20m | SimCLR-More-MLPs · Real-time benchmark | mismatch score to human | Hurt |
| lengthen aggregation time T from 0.2s to 1.0s | SimCLR-More-MLPs · Real-time benchmark | mismatch score to human: T=0.2s -> T=1.0s | Hurt |
| lengthen aggregation time T | SimCLR-More-MLPs · Life-long benchmark | trajectory-averaged Mini-ImageNet accuracy | No clear effect |
Closest papers by meaning
- The CLEAR Benchmark: Continual LEArning on Real-World Imagery (NeurIPS 2021 · similarity 0.89)
- Scaling and Benchmarking Self-Supervised Visual Representation Learning (ICCV 2019 · similarity 0.88)
- Online Continual Learning With Natural Distribution Shifts: An Empirical Study With Visual Data (ICCV 2021 · similarity 0.88)
- Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online Continual Learning (NeurIPS 2024 · similarity 0.87)
- Probing the Mid-level Vision Capabilities of Self-Supervised Learning (CVPR 2025 · similarity 0.87)
- RTify: Aligning Deep Neural Networks with Human Behavioral Decisions (NeurIPS 2024 · similarity 0.86)
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.
