mardgui

mardgui / SEM

Public

[CVPR Findings 2026] SEM: Sparse Embedding Modulation for Post-Hoc Debiasing of Vision-Language Models

19
0
100% credibility
Found Mar 26, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Shell
AI Summary

Research codebase evaluating post-hoc debiasing methods for CLIP vision-language models on fairness benchmarks.

How It Works

1
📖 Discover Fair AI Tool

You stumble upon a helpful project that fixes biases in AI matching pictures to words, like avoiding unfair stereotypes.

2
🖥️ Set Up Your Workspace

You prepare a simple space on your computer with everyday tools to explore fairness in AI.

3
🖼️ Gather Test Photos

You collect everyday photo albums of faces and birds to check how AI treats them fairly.

4
🔄 Smart-ify the Photos

You turn the photos into understandable hints so the AI can learn from them without bias.

5
✅ Run Fairness Checks

You launch tests and watch the AI improve its matches, reducing unfair pulls toward stereotypes.

6
📥 Add Fairness Boosters

You grab ready-made helpers that make the AI even better at fair picture-word matching.

🎉 See Fair Results

You enjoy clear charts proving the AI now matches pictures and words much more fairly.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is SEM?

SEM debias-es vision-language models like CLIP post-hoc by modulating sparse embeddings from autoencoders, tackling social biases in retrieval and zero-shot classification on datasets like CelebA, FairFace, UTKFace, and Waterbirds. You get shell scripts to precompute CLIP embeddings, run evaluations across four backbones (ViT-B/16, ViT-L/14, RN50, RN101), and reproduce CVPR Findings 2026 tables comparing SEM against baselines like PRISM and ZSDebias. Built in Python with PyTorch, it includes pre-trained weights via GitHub releases for quick starts.

Why is it gaining traction?

Unlike dense-space methods that entangle bias and semantics, SEM pinpoints bias neurons in disentangled sparse latents for precise interventions without retraining. Developers dig the zero-shot fairness boosts—lower KL divergence and skew in stereotype queries—plus one-command repro of CVPR findings track results, saving weeks on baselines. It's a fresh take from cvpr 2026 papers github, standing out in the cvpr findings workshop crowd.

Who should use this?

ML researchers auditing VLM fairness on faces or birds, especially those replicating CVPR 2024/2025/2026 github repos for debiasing papers. Vision devs integrating CLIP into search/retrieval apps needing gender/race/background bias fixes without fine-tuning. CVPR poster github hunters prepping rebuttals or predictors.

Verdict

Grab it if debiasing VLMs is your jam—docs and repro scripts are solid despite 19 stars and 1.0% credibility score signaling early maturity. Low test coverage means tweak cautiously, but precomputes make it dev-ready now.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.