Deep Learning in Real Life
summary
A concrete mobile app (Magritte) ships a convolutional model for fruit/vegetable recognition trained with Keras/TensorFlow on GPUs, but performance drops when phone camera data shifts. The talk says feature engineering still consumes 60–70% of project time so deep nets pay off only at scale, and recommends fixes: data augmentation, adding web images, and transfer learning from ImageNet; it also discusses dropout-based uncertainty and costly Bayesian nets, plus practical monitoring and retraining strategies (including skewed test splits like 95/1/1).
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:07What is deep learning
- 10:08Deep learning applications
- 14:25Why deep learning explodes
- 18:42Using deep learning in practice
- 26:43Performance with little data
- 31:16Model interpretability
- 37:00Handling prediction uncertainty
- 42:13Summary and best practices
- 44:51Q&A
related talks
05mDeep Learning in the Wild
38mDeep Learning for detection on a phone
01h 00mDeep Learning Revolution: Rethinking Machine Learning Pipelines
41mDeep Learning to Solve Challenging Problems
47mDeep Learning: a practical introduction from an Android dev perspective
41mWarm Up Your Smartphone's Neurons with Deep Learning
40mMachines are Learning
01h 04mWhat you always wanted to know about Deep Learning, but were afraid to ask