CS84 Today

0. Discuss Papers that you liked (if any)

1. Reviewing Regression:

http://arachnoid.com/polysolve/

2. Discussing the theory of Back-Propagation:

http://karpathy.github.io/neuralnets/

3. Discussing Features, layers, and the effects of hyperparameters:

first hidden node:
tanh(x1 * w1 + x2 * w2);

http://playground.tensorflow.org/

4. Writing a binary classifier in Keras (DEMO)

5. If we have time, discussing train vs test and crossvalidation. If we don't have time, this will go into next week.

Homework due 2/28:

1a. Pick a novel binary math function, e.g. f(x1,x2,x3,x4) -> boolean, generate data, and train a network to classify this data

1b. Shrink the size of your neural network until it no longer predicts 90+% accuracy -- by removing layers or removing nodes per layer.

1c. Save two versions of your model. One which accurately predicts, and one which doesn't accurately predict.

2a. Find an interesting data set online in csv format.

2b. Train a network to classify a feature on the dataset

2c. Write 1-3 sentences about the performance of your algorithm and what it implies

3a. Do this 3-5 times (please take this part of the assignment seriously... spend at least a couple hours on this!): Find some information on the internet that could be converted into a dataset (but has not currently been formatted as a csv)

3b. Write what the columns (features) might be in your dataset

3c. What are some things you might be able to classify (predict) that are already labeled? What are things you could predict that aren't labeled but could be labeled by hand? What are things that you could predict that would require other data to be labeled (and where could you find that data)?

3d. Be prepared to come to next class to discuss ideas on how you might parse your data set into a csv.