ML Algorithms for Trust and Safety
summary
YouTube removes 3.75 million videos and 1.26 billion comments in Q4 2021 and claims roughly 92% of video removals and >99% of comment removals come from automated flagging. Systems use modular multimodal pipelines (video pixels, audio, ASR text, OCR, metadata) with precomputed video vectors, sliding-segment localization and T5 for text. Models face scarce end-task labels and adversarial circumvention so policies use pseudo-labels, sampling for measurement and a three-strike system (94% with one strike never get a second) plus a semi-supervised graph to find coordinated violative creators.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:26Video removals by numbers - Q4 2021
- 2:41How do we use machine learning systems?
- 4:14Challenges
- 6:54Multi-modal content classification
- 8:48Video representation
- 9:42Text representation: Text-To-Text Transfer Transformer
- 10:45Speech representation
- 13:43Semi-Supervised Graph Building (GRALE)
related talks
42mOnline Safety Bill: How Global Platforms Use MLOps to Keep People Safe
30mML Automoderation System for LLM-Augmented Online Platforms
30mData Driven Security
16mProduction MLOps for Spam Detection: Real-Time Multi-Modal AI
24mEngaging new and diverse audiences with ML
30mVideo Quality Assessment of User Generated Contents
39mSecure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
26mModern NLP for Proactive Harmful Content Moderation