← All workshopsRead Transfer Learning

Computer Vision with Neural Networks

Apply a new head to fixed leaf features

The features stay fixed while the head changes

Botie already has fixed feature vectors for leaves. A new head can interpret those same measurements differently without altering the feature rows.

BOTIE

The leaf features are identical to the first case, but the weights change. Why can the class decisions change?
The inputs we are comparing

Fixed leaf features

Same in both cases

10
01
11

New head weights

Three leaves

  1. 02
  2. 1-1

A different head

  1. 0-1
  2. 12

New head bias

Same in both cases

-0.5
Result for three leaves
  1. 01
  2. 10
  3. 21
Choose an explanation to test
A new head combines three fixed leaf features

This check uses the pieces from “Attach a supplied classifier to fixed leaf features”.

BOTIE

Each leaf has three already-computed features. Apply the supplied head weights and bias, then classify scores at or above zero as 1. Return one decision per leaf without changing the feature extractor.
Fixed leaf features
012
200
010
New head weights
  1. 01
  2. 1-2
  3. 20.5
New head bias
1

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.

What does the exactly-on-the-boundary case reveal about the threshold convention?

A reused representation and a new prediction head have different responsibilities and can be changed separately.

The feature vectors and head are invented. This construction demonstrates inference from a supplied head, not backbone training or a transfer-performance claim.

Work through the Python challenges →
Continue the courseBounding Boxes and Overlap →
← Image Augmentation