What's new in Deep Learning?
summary
Survey of concrete advances and how to use them: Glorot/Xavier and He initialization for stable signal flow, BatchNorm applied per-mini‑batch with running averages to speed training and allow larger learning rates, residual networks that train to 110–1000 layers and cut ImageNet top‑5 error (ResNet‑152 ~3.57%), CIFAR-10 experiments showing plain deep nets degrade while ResNets converge, Keras code sketch using Adam and categorical cross-entropy that actually trains.
AI-written summary · about
chapters (10)
- 0:00Deep learning at Zalando
- 1:27Core deep learning principles
- 8:06Training neural network challenges
- 13:27Weight initialization
- 14:50Batch normalization
- 20:12Deep learning in computer vision
- 23:22Deeper networks and residuals
- 29:50Residual networks with batch norm
- 31:42Implementing residual networks with Keras
- 35:17Summary and Q&A
related talks
56mTutorial on Deep Learning II
45mIntroduction To Deep Learning
01h 02mTutorial on Deep Learning I
54mIntroduction to Understanding Deep Learning
27mDemystifying Deep Neural Networks
35mDeconstructing Deep Learning
01h 00mDeep Learning Revolution: Rethinking Machine Learning Pipelines
51mDeep Learning - What is it and What It Can Do For You