1 of 8

Neural Networks for Feature Analysis

Daksh, Mayank, Katie

2 of 8

Back to square one

  • Last week’s results were not useful ☹️
  • Why?
    • Mistake in training data
    • 80% accuracy because of bias
    • (irrelevant) learning happened
  • Let’s investigate🕵️

​

3 of 8

What did we learn?

We can’t progress towards feature analysis if we don’t know how the implementation of

model works

​

  • Learning how Alexnet(CNN) works
  • How our model processes the inputs
  • Testing by running simple datasets

Convolutional Neural Network

4 of 8

Class 1

Class 2

5 of 8

Should is a magic word

Model should be able to tell difference

Model should pick up this feature

Model should get 100% accuracy

No.

  • 66.6% accuracy (exactly ⅔)
  • why?

6 of 8

ML Software Stack

  • Same as Beehive Monitoring
  • 2 parts
    • Dataset prep
    • Training

​

7 of 8

Class 1

Class 2

8 of 8

Goals for next week

  • Get 100% accuracy on black/white squares
  • Project Webpage
  • Run controls
    • 50% on randomly mislabeled classes
  • Then try the plots
    • entropy vs accuracy, etc.

1

2

2

2

1

1

2

1

1

2

2

1