Demystifying Deep Neural Nets
summary
Covers how convolutional neural nets learn weights with gradient descent and backprop and that libraries like Tensorflow and ConvNetJS handle the heavy math so you can build models quickly. Gives numbers: ImageNet error falls from ~25% to 15.3% in 2012 and convnets beat humans by 2015. Shows practical outcomes: style‑transfer apps and Google DeepDream hallucinate images, tiny adversarial perturbations flip a truck into an ostrich, and simple four‑layer nets often perform surprisingly well.
AI-written summary · about
chapters (19)
- 0:00Start
- 2:05History
- 3:02Recognising a cat
- 4:14Conventional computing vs Neural nets
- 5:35Trivial example
- 6:00A single neuron
- 8:16A Deep Neural Net
- 8:36Activation functions
- 9:57Training the network: the short version
- 11:17Training the network: the long version
- 12:42Minimising the loss function
- 16:48Images are 2D arrays of numbers
- 17:24Convolution
- 17:59Examples
- 19:42Object recognition
- 20:14ImageNet Challenge Results
- 20:54Inside a ConvNet
- 22:06Separating style from content
- 24:24Google's Deep Dream
