Nov 10
aside
homework
... for next week is posted.
Also posted new neural net library on
Nov 8 page:
Brain, a neural net API in javascript
neural networks
Discuss my attached code ... which doesn't work.
Ah - I see this morning that I'm not
using the 0th offset weights in either
the forward or backward propagation.
I'll fix that and re-test this weekend.
But use it anyway to explain
- the details of how this particular form of the neural net works
- forward propogation algebra
- back-prop weight adjustments
Discuss AND, OR, XOR ..., linear vs non-linear and why I picked those for (what I'd hoped) would be a simple example.
Even if my code is correct (probably not yet), there are unfortunately many things that can go wrong. Several are discussed in the
Willamette notes
Also mention why you would never use a multi-layer linear network.
Cross-validation in neural nets:
- crude: try different topologies & see if things get better.
- AIMA: "there is not good theory ..."
Go over textook AIMA 18.11 example: handwriting recognition
unsupervised learning
Discuss the general ideas, what sorts of problems, what sorts of math involved.
Mention the Green River project I did in this direction last summer.
Files in greeriver/2011/hei_summer/analysis/results & results_DRG
- social_health_weighted.png
- cluster_municipality_nilmedian.csv
- etc
Their sites:
example techniques :
![[paper clip]](/courses/source/wiki_images/paper_clip_tilt.png)
name last modified size