Archive
Batch normalization, visually: the reset that keeps deep networks steady
Why did deep networks stop learning?
Why does every AI start out random?
Why slow down near the finish line?
What if every dial had its own step size?
How does a rolling ball skip small bumps?
Why train on a few examples at a time?
Too big, it explodes. Too small, it crawls.
How does PyTorch find every gradient by itself?
Who's to blame when a network gets it wrong?
How does one nudge ripple through ten layers?
How do you find a valley blindfolded?
This is what a confident mistake looks like
How does a network know it's wrong?
How do raw scores become probabilities?
Without one bend, 100 layers act like one
How do hidden layers bend a straight line?
Why can't one neuron learn XOR?
The perceptron: the 1958 machine that learned from its mistakes
How does a computer know this is a 3?