What training leaves behind

This module followed one idea across four decades.

In the 1940s, researchers described simplified mathematical neurons.
In 1957, Frank Rosenblatt built learning around adjustable connections.
In 1969, a clear limit of a single-layer Perceptron helped cool enthusiasm.
In 1986, a widely read backpropagation paper showed how errors could guide weights across several layers.

The lesson was never that one person suddenly invented learning. Each step reused earlier mathematics, machinery, and questions.

Training changes weights and biases inside a fixed structure. Saving those values preserves what the network learned. Loading them into the matching structure lets inference use that learning on new input.

One problem still remained. In a deep network, the useful correction could weaken before it reached the earliest layers.

The next module begins with that obstacle.

The whole chapter, simply

Learning can take a long time, but we do not want to repeat it for every new example.

So we save the numbers the machine learned and reuse them when it sees something new.

This module followed one idea across four decades.

In the 1940s, researchers described simplified mathematical neurons.
In 1957, Frank Rosenblatt built learning around adjustable connections.
In 1969, a clear limit of a single-layer Perceptron helped cool enthusiasm.
In 1986, a widely read backpropagation paper showed how errors could guide weights across several layers.

The lesson was never that one person suddenly invented learning. Each step reused earlier mathematics, machinery, and questions.

One problem still remained. In a deep network, the useful correction could weaken before it reached the earliest layers.

The next module begins with that obstacle.

The whole chapter, simply

Learning can take a long time, but we do not want to repeat it for every new example.

So we save the numbers the machine learned and reuse them when it sees something new.