Improving Photomath’s math problem scanning with deep learning
summary
Photomath replaces legacy on-device OCR with a cloud-deployed deep learning pipeline that moves from 2016 object-detection OCR to an end-to-end CNN+BiLSTM and finally to a transformer-based image-to-expression model trained in a six-month, 12-engineer sprint. Deployment runs on multi-cluster Torch servers (EU/US) with about one million solutions every 20 minutes and ~1 TB GPU memory; release lifts reported system accuracy from 64% to 90%, cuts error rate ~10→8, and increases premium-content clicks by ~55%. Annotation is handled by an internal team of ~90–100 people.
AI-written summary · about
chapters (4)
related talks
20mGoogle Photos Magic Editor: GenAI Under the Hood of a Billion-User App
26mWhere can I park? Deep Learning is here to help you!
42mUsing Deep Learning to Understand Documents
19mDocument Reader using Deep Learning with a Small Dataset
01h 04mFrom Pixels to Words: Exploring Multi-Modal AI with Python
36mDeep Learning - Segmentation, Identification and Classification
10mSeeing the World With AI: How Deep Learning Enables a New Era of Imaging Technologie
30mIntroduction to Computer Vision with OpenCV