Research map / Label-efficient and robust learning / Recognition and long-tail
To Learn or Not to Learn, That is the Question — A Feature-Task Dual Learning Model of Perceptual Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a hierarchical dual-learning model (fixed Gabor feature extraction, slow Hebbian feature-based learning, fast CNN task-based learning) that reproduces specificity, transfer, and the transfer-to-specificity transition in Vernier perceptual learning experiments.
Problem
Perceptual learning shows both specificity (gains restricted to trained location/feature) and transfer (gains generalize to untrained conditions), and no unified computational model explains both.
Approach
Attributes specificity and transfer to two interacting processes with different learning rates: fast task-based readout learning that transfers via translation-invariant max pooling, and slow feature-based Hebbian learning that induces location-specific changes when the same stimulus is repeated excessively.
Key results
- Task-based learning alone yields accuracy gains at untrained locations, while feature-based learning alone drops untrained-location accuracy to near chance (50%).
- Combined learning reached 90% accuracy in 81.3 epochs vs 84.7 for task-based only (t = 3.90, p < 0.001, 50 repetitions).
- The model reproduces the four classical experiments qualitatively, including the transition from transfer to specificity as training sessions increase (Table A4 p-values all significant).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add feature-based learning to task-based learning (ablation) | Dual-learning model · Vernier discrimination, trained location | epochs to 90% accuracy: 84.7 -> 81.3 | Helped |
| freeze feature-based learning, leaving task-based learning enabled (ablation) | Dual-learning model · Vernier discrimination, trained and untrained locations | discrimination accuracy | Helped |
| freeze task-based learning, leaving feature-based learning enabled (ablation) | Dual-learning model · Vernier discrimination, trained and untrained locations | discrimination accuracy | Mixed |
| enable both feature-based and task-based learning (ablation) | Dual-learning model · Vernier discrimination, trained and untrained locations | discrimination accuracy | Mixed |
| increase feature-based learning rate tenfold (ablation) | Dual-learning model · Specificity experiment, untrained locations | perceptual performance | Hurt |
| increase feature-based learning rate tenfold (ablation) | Dual-learning model · Varied-condition transfer experiment | transfer | Hurt |
| increase feature-based learning rate tenfold (ablation) | Dual-learning model · Double-training experiment | transfer | Hurt |
| decrease task-based learning rate tenfold (ablation) | Dual-learning model · Vernier discrimination experiments | learning effectiveness | Hurt |
| decrease task-based learning rate tenfold (ablation) | Dual-learning model · Varied-condition transfer experiment | transfer | Helped |
| decrease task-based learning rate tenfold (ablation) | Dual-learning model · Double-training experiment | transfer | Hurt |
Surprising findings
- Accelerating feature-based learning tenfold degrades transfer and breaks the double-training transfer effect, showing the slow rate is essential.
- Slowing task-based learning tenfold removes the transfer difference between conditions, indicating both learning speeds are necessary.
Limitations
- Very simple network model lacking many biological details of the visual pathway.
- Only the Vernier discrimination task is simulated; claims about other tasks are not tested.
- Outlier removal was needed for the double-training result, and several hyperparameters (learning rates, thresholds) are set by hand with inconsistent values across experiments.
Experimental setup
Datasets: synthetic Vernier discrimination stimuli
Benchmarks: simulated replications of psychophysical experiments (Vernier specificity, orientation transfer, transfer-to-specificity
Metrics: discrimination accuracy difficulty threshold at 80% accuracy percent improvement t-test p-values feature representation similarity
Baselines: HyperBF reweighting model CNN-based perceptual learning model (Li et al.)
Builds on: HyperBF reweighting model two-stage model theory reverse hierarchy theory competitive Hebbian learning
Closest papers by meaning
- Flexible Context-Driven Sensory Processing in Dynamical Vision Models (NeurIPS 2024 · similarity 0.86)
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.85)
- Subliminal Learning Is Steering Vector Distillation (NeurIPS 2026 · similarity 0.84)
- Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning (CVPR 2024 · similarity 0.84)
- Perceptual Scales Predicted by Fisher Information Metrics (ICLR 2024 · similarity 0.83)
- Understanding Task Transfer in Vision-Language Models (CVPR 2026 · similarity 0.83)
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.