Code
Discuss
What kind of confusion matrix would be ideal, given that not all 'lasing' images lased?
Can we trust model? Data comes from different runs?
Are you seeing big swings in the confusion matrix?
Big swings in the accuracy? What might help?
Exercises
Try different hyper parameter selection in the model
- learning rate
- optimizer momentum
- minibatch size
- validation size
- relu vs. other