Loading slide

Why Shallow Networks Failed

  1. 01Why Shallow Networks Failed
  2. 02Memory card
  3. 03What "deep" actually means
  4. 04The vanishing gradient
  5. 05A small step inside each neuron
  6. 06Why bother bending the number?
  7. 07The bend they chose
  8. 08How the flat ends starve the signal
  9. 09The absurdly simple fix
  10. 10Take a breath
  11. 11Other fixes
  12. 12Depth vs. width
  13. 13Memory card
  14. 14The research gap
  15. 15Reinforce your understanding
  16. 16Question: The vanishing gradient
  17. 17Question: Why does a simple fix work?
  18. 18Question: Why does depth matter?
  19. 19Quiz: answer
  20. 20The one idea to keep
  21. 21The gradient problem was solved. What was still missing was scale.
  22. 22Want to go deeper?
7 / 21
BackNext

The bend they chose

Look at the two pictures. Each one shows what a rule does to a number: the number you feed in runs along the bottom, the number that comes out rises up the side.

On the left is a plain multiply, "times 2." Its picture is a straight ramp. The more you put in, the more comes out, climbing forever with nothing to stop it. A straight line, never a bend. That is the collapse problem from a moment ago, drawn out: stack these and you just get one steeper straight line.

On the right is the rule the early networks actually chose. It was called , and it has one firm habit: the output can never go above 1 or below 0. No matter how huge the input, the answer is capped at 1.

That cap is what creates the bend. A line that is not allowed to climb past 1 cannot keep running straight up. It is forced to level off near the top, and level off near the bottom too, with a steep rise in between. That is the shape on the right, a lazy letter S lying on its side. So "squash everything into 0 to 1" and "bend the line" are not two separate ideas. They are the same fact: the cap is the squash, and the squash is the bend.

For passing a signal forwards, that bend works nicely. The trouble is hiding in those two flat ends, and that is where the vanishing gradient comes from.

Citations(2)↓
  1. 1. deeplearningbook.org
  2. 2. ibm.com
Citations(2)↓
  1. 1. deeplearningbook.org
  2. 2. ibm.com