Loading slide

A Network That Learns

  1. 01A Network That Learns
  2. 02Stacking neurons
  3. 03Quick recall
  4. 04What stacking does
  5. 05Hidden layers
  6. 06The problem of credit
  7. 07Backpropagation
  8. 08Sharing the blame
  9. 09The search problem
  10. 10Fog on a hill
  11. 11Gradient descent
  12. 12Just to be clear
  13. 13What the hidden layer learns
  14. 14What the hidden layer discovers
  15. 15Did it actually learn?
  16. 16Generalization
  17. 17Limits in the 1990s
  18. 18Reinforce your understanding
  19. 19Question: Backpropagation
  20. 20Question: What the hidden layer learns
  21. 21Question: Memorizing vs. generalizing
  22. 22Quiz: answer
  23. 23The method existed. What was missing was scale.
  24. 24Want to go deeper?
8 / 23
BackNext

Sharing the blame

The whisper game gives the feeling of it. Now the detail, because this is the engine the whole field runs on.

A network is more tangled than a line of people. Each neuron does not listen to just one neighbour. It receives signals from many connections at once, and passes its own signal on to many more. So when the final answer is wrong, the blame cannot go to a single weight. It has to be split across all of them, and split fairly.

So backpropagation works out a single number for every weight in the network: if this one weight had been a little higher, or a little lower, would the final answer have been closer to correct, and by how much? A weight that pushed hard toward the mistake gets a large number. A weight that barely mattered gets one near zero. That number is the weight's exact share of the blame, and it doubles as an instruction: this is the direction to nudge, and this is roughly how far.

The remarkable part is that it computes this for every weight at the same time, in one backward sweep. A network might have billions of weights. Backpropagation does not test them one by one, trying each in turn. It sends the error backward through the network once, and out the other side comes a precise correction for every weight at once. That efficiency is the only reason training enormous networks is possible at all.

Two things had to come together for this to work. The error has to travel back along the very connections that carried the signal forward, so each weight is judged on the path it actually influenced. And the size of the correction has to be proportional to the size of the mistake: a small error, a barely-there nudge; a large error, a firmer push.

Backpropagation solved the credit problem. But knowing which weights were responsible, and by how much, is only half the picture.

Citations(2)↓
  1. 1. doi.org
  2. 2. 3blue1brown.com