Writing a Neural Net from Scratch
summary
Builds a handwritten-digit recognizer entirely in C# and LINQPad with no libraries, implementing perceptrons, leaky ReLU/softmax activators, backpropagation and stochastic gradient descent. Trains on 50,000/10,000 images, starts with a 30-neuron net that reaches ~96.5% training and ~95.6% test accuracy, scales to 200 neurons and two hidden layers to hit ~98.3% test accuracy, shows data centering, augmentation, learning-rate tuning and why XOR needs multiple neurons.
AI-written summary · about
chapters (18)
- 0:00Introduction to ML
- 2:05Neural network basics
- 4:41Learning from scratch
- 5:46The artificial neuron
- 8:02Training and activation
- 10:37Loss functions
- 15:03Adjusting weights
- 17:50Gradient descent
- 20:34Mathematics of training
- 26:06Coding in LINQPad
- 30:03Training the logic gate
- 33:33Multi-layer networks
- 37:00Implementing the network
- 39:38Digit recognition results
- 41:53Interactive drawing demo
- 45:21Improving performance
- 50:39Advanced CNN concepts
- 52:52Conclusion and Q&A
