CLIP and Multimodal Models

CLIP in one diagram

Train on ~400M (image, caption) pairs from the web with a symmetric InfoNCE loss over the batch: each image must pick out its own caption among all captions in the batch (and vice versa) — literally the in-batch sampled softmax, cross-modally. Result: a shared space where text and images are directly comparable.

Superpowers

1. Zero-shot classification: classify by embedding label descriptions (“a photo of a swastika”, “a screenshot of gambling content”) and taking the nearest. New policy category → write text prompts, ship today, fine-tune later. This is the killer feature for content moderation at Roblox-like platforms (new harm categories appear weekly).
2. Cross-modal retrieval: text→image and image→image search over an ANN index of CLIP embeddings — find all UGC similar to a confirmed-bad asset (avatar/clothing reuse is an account-matching signal too: AM-02 Signals and Features).
3. Universal visual features for downstream heads (linear probe / small MLP on frozen CLIP often beats training a CNN from scratch on limited labels).

Beyond CLIP (one-liners)

• BLIP-2 / LLaVA-class VLMs: connect a vision encoder to an LLM (via a projection/Q-Former) → captioning, VQA, “explain why this image violates policy” — useful for reviewer assistance, not just classification.
• Hash matching (not ML, but say it): PhotoDNA/PDQ perceptual hashes catch known bad content exactly and cheaply; ML handles novel content. Moderation = hash layer + CLIP/classifier layer + VLM/human layer — defense in depth again (Anatomy of a Fraud Detection System).

Moderation system sketch (Roblox-flavored)

Asset uploaded → perceptual-hash check against known-bad DB → cheap CNN/CLIP triage score → if uncertain band: heavy VLM + text/audio models (multimodal fusion) → thresholds per harm severity (child-safety thresholds set for high recall, with human review absorbing the FP cost) → reviewer queue ranked by severity×uncertainty → decisions feed training. Adversarial notes: perturbation attacks vs hashes, text-in-image (OCR leg), policy drift → continuous re-labeling.