Demystifying Deep Neural Networks
summary
Shows how networks learn weights and biases with gradient descent and backprop, uses ReLU and other activation functions, and runs on GPUs with large labelled datasets and libraries like Tensorflow and ConvNetJS. Explains convolution as a shared kernel filter over image arrays, cites ImageNet error dropping from about 25% to 15.3% in 2012 and surpassing humans by 2015. Demonstrates style transfer (Gatys / Prisma), DeepDream hallucinations, adversarial examples that flip a truck to an ostrich, and warns about algorithmic bias.
AI-written summary · about
chapters (9)
- 0:00Introduction to neural networks
- 3:38How neural networks differ
- 8:16Building a simple neuron
- 10:22Training neural networks
- 16:42Using TensorFlow for neural networks
- 17:02Convolutional neural networks
- 20:32Applications of convolutional networks
- 25:30Bias and ethical concerns
- 26:55Summary and closing
related talks
28mDemystifying Deep Neural Nets
54mIntroduction to Understanding Deep Learning
35mDeconstructing Deep Learning
41mReally Deep Neural Networks with PyTorch
01h 02mTutorial on Deep Learning I
51mDeep Learning - What is it and What It Can Do For You
01h 04mWhat you always wanted to know about Deep Learning, but were afraid to ask
45mIntroduction To Deep Learning