Computer Vision with Neural Networks

Keep the strongest local signal

What the pooled map cannot tell us

Botie's feature map marks stroke responses. Keeping the strongest response in each region can reduce dependence on its exact position inside that region.

BOTIE

Botie may later need either coarse stroke presence or precise pixel boundaries. Which information would you preserve for that goal?

The inputs we are comparing

Feature map

Example 1

1	3	2	4
0	2	1	3
5	2	6	1
4	3	2	0

Example 2

2	0	1	1
1	4	3	2
0	1	5	2
3	2	1	4

Result for example 1

Choose an explanation to test

Retain the detailed map when precise boundaries matterUse pooled responses when coarse local presence is enough

Every feature response is negative

This check uses the pieces from “Keep one strongest response from each region”.

BOTIE

Take the maximum in each non-overlapping 2-by-2 region, using stride two and no padding. Return the four pooled values in row-major order. Do not insert a zero that was absent from a region.

Feature map: -1 -3 -2 -4
-5 -2 -1 -6
-7 -2 -9 -3
-4 -8 -5 -6

What result should your machine produce for this case?

Use a number, text, or an array matching the requested result. For a matrix, put each row in brackets. Saving your prediction runs your current construction on these inputs and checks the earlier examples too.

Build a construction in the workspace above first. This case will test those pieces.

Could two different arrangements within a pooling window produce the same output? Explain with its maximum.

Your notes for later. These stay with this journey.

A useful summary is also an information-losing operation.

This is a small construction problem. The lesson explains the assumptions and the parts of the full method that this example leaves out.

Work through the Python challenges →

Continue the courseConvolutional Networks →

← Convolution