anurag builds things

anurag builds

I build things to understand how they actually work. A question, a system built to answer it, a comparison against real data. Just the code, the curves, and what they actually mean.

Linear probe accuracy by augmentation recipe: 48.3 / 71.9 / 83.7 / 82.1%.

Augmentations Are the Model

self-supervised-learning
contrastive-learning
augmentations
computer-vision

Four augmentation recipes on CIFAR-10, held constant everywhere else. The result isn’t subtle - and the pattern has a cleaner explanation than ‘stronger is better’.

15 Apr 2026
Four pretext tasks: contrastive, masked, predictive, denoising.

Self-Supervised Learning

self-supervised-learning
representation-learning
fundamentals

Core intuition behind self-supervised learning: why it works, when to use it, and how it connects to real systems.

18 Mar 2026
SimCLR diagram: two augmented views, encoder, projection, embedding space.

Learning Representations Without Labels (SimCLR)

self-supervised-learning
contrastive-learning
computer-vision

Implementing SimCLR from scratch to learn visual representations without labels using contrastive learning on CIFAR-10.

18 Mar 2026
Augmentation recipe matrix: minimal, moderate, full, aggressive.

Why Augmentations Matter in Contrastive Learning

self-supervised-learning
contrastive-learning
augmentations

Testing how different augmentation strategies affect representation quality in SimCLR-style contrastive learning.

18 Mar 2026
No matching items