Dec 1
Start discussion of last two chapters
of neural nets online book :
- why "deep" nets : layers of abstraction
- but : problems with deep training (different rates, differing convergence)
- one approach: feature detection in images ("convolutional networks")
- but : requires an innate understanding of that problem built into network
aside: convolution & image processing
- the math concept itself
- applications to sensors (e.g. astronomy and telescopes)
- applications to image feature analysis - blurring, line detection, etc
recent results :
- successes
- image classification
- street numbers
- failures
- "blind spots" i.e. "adversarial" images
- "white noise problem" : looks random to human, not to deep trained network
appendix : "simple" general AI "
for Thursday :
Sam :
