Hello, if you have any need, please feel free to consult us, this is my wechat: wx91due
CAP4611 Homework 5
Preliminaries
- Visit https://playground.tensorflow.org
- Turn on “Show test data” check box.
- Set the batch size to 10.
- Familiarize yourself with the four training sets/ problems for classification
– P-GAUSS: Gaussian– P-XOR: Exclusive or– P-CIRCLE: Circle– P-SPIRAL: Spiral
- Experiment with choosing the ratio of training to test data (which changes the amount of the training data), and the noise. You need to push “regenerate to create a new dataset”. We will call these training data quality.
– Q-BEST: 90% training data, 0 noise– Q-WORST: 10% training data, 50 noise– Q-OK: 70% training data, 20 noise
Four sample models
Consider the following four models. For all of them, the regularization is None.
- Linear regression (LR)
– 0 hidden layers– x1 , x2 inputs– activation linear
- Logistic regression (LO)
– 0 hidden layer– x1 , x2 inputs– activation sigmoid– regularization none
- 1980s neural net (NN)
– 1 hidden layer (5 neurons)– x1 , x2 inputs– activation sigmoid– regularization none
- Engineered features (EF)
– 0 hidden layers– x1 , x1x2 inputs– activation linear
Problem 1 - Evaluation (50 pts):
For each of the four sample models (LR, LO, NN, EF), and all four problems (P-GAUSS, P-XOR, P-CIRCLE, P-SPIRAL) and all three quality settings Q-BEST, Q-WORST, Q-OK do the following:
- Try to train the model to learn the task.
- You might want to vary the learning rate. Don’t let any of the models train more than 30 seconds.
- Record the best training loss and test loss obtained.
- Record the time it took to finish the learning.
- Judge whether the learning was success and whether the system overfit or not. Use your own criteria of success.
Write it up like this:
EF + P-XOR + Q-BEST
Successful, not overfit, <1 sec L_train = 0.018, L_test = 0.035, rate 0.00001
Write two paragraphs of conclusions. Did any of the models solve all problems?
Were some models better for some problems? Did some models train faster than others?
Problem 2 - Network design (50 pts):
Develop three new models, to improve learning, as follows
- MYENG:
– any input features (try to be creative about this)– any number of hidden layers– activation linear
- DEEP
– x1 , x2 input– sigmoid activation– any number of hidden layers
- MYBEST
– x1 , x2 input– any activation– any number of hidden layers– feel free to use regularization if needed or change the batch size.
Your objective is to find the best solution within the set constraints.
Repeat the evaluation process for the resulting networks.
Write up the conclusions of the work in two paragraphs. Did you manage to create a network which solves all problems in all quality settings? Which settings were the hardest? Did some models train faster than others?