Assignment 3
Neural Networks and Deep Learning
CSCI 5922
Spring 2017
Assigned Thu Sep 28, 2017
Due: Thu Oct 12, 2017
Assignment submission
Goal
The goal of this assignment is to use
tensorflow to build some neural networks, and to experiment with
the options and flexibility that tensorflow offers.
Data Set
For Part 1 of the assignment, you will use
the room occupancy data set that you used for Assignment 2.
For Part 2 of the assignment, you will use the data set
available
here from
me. This data set consists of 104 relationships described by the
Hinton Family Trees network.
Here is some
very embarrassing
python code I wrote to
read in the data set and build
an input/output representation
that can be fed to tensorflow.
Part 1
Use tensorflow to build a feedforward neural
net to predict occupancy. This net should do exactly the same
thing that your code in Part 2 of Assignment 2 does.
(1a) Run a simulation using tensorflow that is identical to
Assignment 2, part 2f, in which you vary the number of hidden
units and make a plot. Superimpose the plot you made from
Assignment 2, part 2f.
(1b) DIscuss the results: are they the same as with your own
code? If one works better than the other, explain why you think
that is.
(1c) Add a second hidden layer, and train a few architectures
with 2 hidden layers. Report what architectures you tried
(expressed as 5-h1-h2-1, i.e., 5 input, h1 hidden
in first layer, h2 hidden in second layer, and one
output unit), and which ones, if any, outperform your
single-hidden-layer network.
It will not be a big deal to add the second hidden layer once
you have all the rest of your code in place.
Part 2
Replicate the Hinton Family Trees
architecture in tensorflow. My notes describe the
architecture and number of neurons in each layer. You can also
refer to the
original
paper.
(2a) Randomly split the data set into 89 examples for training
and 15 for testing. Train and evaluate 20 such random splits
of the data and report the mean and standard deviation of
the test set accuracy. (Report accuracy not squared error.
A response should be counted as correct if the most active unit
is the target person2.)
(2b) Train a network on all 104 examples and examine the weights
from the one-hot person1 input representation to the distributed
person1 representation. Figure out a sensible way to graphically
display the weights in the 6 hidden units.
(2c) For at least 2 of the hidden units, interpret what the
network has learned in its mapping from inputs. You'll have to
refer to my create_dataset.py code to determine the
interpretation of the person1 input neurons. (I tried to keep
the same ordering as Hinton uses.)
Part 3 (Extra Credit)
Compare the achitecture Hinton describes for
Family trees with a generic feedforward architecture with one
hidden layer consisting of 12 neurons.
(3a) Conduct an experiment like the one in (2a) using the
generic architecture. Report the mean and standard deviation of
the test set accuracy.
(3b) Do you see any difference in performance between the
structured net (2a) and the generic net (3a)?
(3c) It wasn't difficult to interpret at least some of the
hidden units in the structured net. Can you interpret what any
of the hidden units are doing for the generic net?
Explain.