AI Snips
Chapters
Transcript
Episode notes
Coffee Cup Example Shows Generalization
- They use the coffee cup example: training on many labeled coffee-cup images lets the net generalize to unseen cups.
- Feed many images labeled 'coffee cup' so the network adjusts weights to recognize new cup pictures it hasn't seen.
Hidden Units Are Simple Weighted Functions
- Neural network hidden units are simple mathematical functions that take many inputs, weight them, sum them, and apply a nonlinearity.
- Each hidden unit reads many pixels (or upstream outputs), multiplies by learned weights, sums, then uses a logistic-like function to decide its output.
Training Changes Weights Not Functions
- Training a neural net does not change the form of the neuron functions; it adjusts the numeric weights on connections between units.
- The same equations stay fixed while many connection weights move to change contributions from different inputs.