Semi-supervised
97.18%
CNN with pseudo-labeling under strict low-label conditions.
SSL and Semi-Supervised Learning on MNIST
Computer Vision · December 2024
Low-label learning case study with pseudo-labeling and SimCLR representations.
Impact: 98.55% SimCLR probe
Constraint: 100 labels
Stack: PyTorch · CNN · SimCLR
TL;DR
- Problem: Train strong classifiers with only 100 labeled samples across 60,000 images.
- Approach: Combined pseudo-labeling and self-supervised representation learning in a reproducible protocol.
- Outcome: High-quality results under strict label constraints, with clear error-analysis workflow.
What I Built
Semi-supervised branch: CNN baseline with pseudo-label refresh loops and targeted augmentation.
Self-supervised branch: SimCLR pretraining with downstream linear-probe and MLP-head evaluation.
Protocol controls: Stable splits, repeatable seeds, and aligned evaluation criteria.
Analysis: Confidence profiles and failure-sample inspection for ambiguous digits.
Results
Self-supervised
98.55%
SimCLR linear probe with only 100 labeled samples.
Self-supervised
98.44%
SimCLR with MLP classification head.
Links
Demo link coming soon GitHub