How to Effortlessly Automate File Recognition with AutoML
summary
Pipeline splits PDF bundles into images, uses simple image features and perceptual hashing to pre-group pages, sends ambiguous groups to a human feedback loop, then trains a Google AutoML image classifier on ~1,500–2,000 labeled pages. Training uses 8 nodes to finish in about an hour (≈$3.15/node/hr) and yields about 96% accuracy on classification; deployed models incur a baseline ~$30/day while inference costs roughly $1.25–$4 per million predictions, and results post back via webhook.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:42What were we trying to do
- 2:05Challenges
- 3:28The Challenge
- 4:40The Opportunity
- 5:14The Problem
- 6:45The Advantages
- 7:48Google Cloud Vision
- 11:32Machine Learning
- 11:46Setup
- 12:14Google AutoML
- 13:48Google AutoML UI
- 14:11Training and Deployment
- 19:55AutoML Downsides
- 23:42Python Notebooks
- 25:31Start with AutoML
- 26:32Google Cloud Vision API
- 27:39Use Multiple Tools
- 27:54Know Your Data
- 28:19Questions
related talks
24mAutoML: No Code No Problem
47mBuilding smarter solutions without machine learning expertise
38mReal-world Machine Learning with TensorFlow and Cloud AI
57mBuilding smarter solutions with Machine Learning, from magic to reality
15mAI as a Service: Quick and Scalable Machine Learning POCs with No-Code/Low-Code Tools
50mAzure AI, Power new possibilities for every organization
31mBuild threat detection models using ML
22mAutoML Vision: the PaaS of Machine Learning