Recitation for 6.034 Artificial Intelligence at MIT, Fall 2016
Subtopics covered:
1. Forward propagation (evaluating output)
2. Choosing neural net parameters by hand to classify small datasets
- Each neuron in first layer draws a line and shades one side
- Remaining layers combine line shadings with logic function(s)
3. Training with backward propagation (adjusting weights and thresholds)
- Threshold trick (representing threshold as a weight, with input -1)
- Threshold functions (stairstep and sigmoid)
- Measuring performance (accuracy)
- Gradient ascent intuition
- Applying quick formulas to compute delta values and weight updates