Description
Policies
Coding: You must write your own code. You may use any numerical linear algebra package, but you may
not use machine learning libraries (e.g. sklearn, tensorflow) unless otherwise specified. In addition, each
student must write and submit their own code in the programming part of the assignment (we may run your
code).
Acknowledgments: We expect you to make an honest effort to solve the problems individually. As we
sometimes reuse problem set questions from previous years, covered by papers and webpages, we expect the
students not to copy, refer to, or look at the solutions in preparing their answers (referring to unauthorized
material is considered a violation of the honor code). Similarly, we expect to not to google directly for
answers (though you are free to google for knowledge about the topic). If you do happen to use other
material, it must be acknowledged here, with a citation on the submitted solution.
Readings
Read the required material.
Required HW submission format:
The following is the required HW submission format:
• Submit all the answers to the HW as one single typed pdf document (not handwritten). This document
must contain all plots. It is encouraged you latex all your work, though you may use another comparable
typesetting method.
• Additionally, submit your code as a separate archive file (a .tar or .zip file). All code must be submitted.
The code should be in a runnable file format like .py files or .txt files. Jupyter notebooks are not
acceptable.
• List the names of all people you collaborated with and for which question(s). Please mark this as
Question 0. Each student must understand, write, and hand in their own answers. Also, each student
must understand, write, and hand in their own code.
1
0 Collaboration and Acknowledgements
List the names of all people you collaborated with and for which question(s). Each student must understand,
write, and hand in their own answers. Also, each student must understand, write, and hand in their own
code.
1 Manual calculation of one round of EM for a GMM [15 points]
(Extended version of: Murphy Exercise 11.7) In this question we consider clustering 1D data with a mixture
of 2 Gaussians using the EM algorithm. You are given the 1-D data points x = [1 10 20].
1.1 M step [10 points]
Suppose the output of the E step is the following matrix:
R =
1 0
0.4 0.6
0 1
where entry Ri,c is the probability of observation xi belonging to cluster c (the responsibility of cluster c for
data point i). You just have to compute the M step. You may state the equations for maximum likelihood
estimates of these quantities (which you should know) without proof; you just have to apply the equations
to this data set. You may leave your answer in fractional form. Show your work.
1. [2 points] Write down the likelihood function you are trying to optimize.
2. [2 points] After performing the M step for the mixing weights π1, π2, what are the new values?
3. [3 points] After performing the M step for the means µ1 and µ2, what are the new values?
4. [3 points] After performing the M step for the standard deviations σ1 and σ2, what are the new values?
1.2 E step [5 points]
Suppose the output of the M step is your previous answer. Let us compute the subsequent E step.
1. [2 points] Write down the formula for the probability of observation xi belonging to cluster c.
2. [3 points] After performing the E step, what is the new value of R?
2 Neural Nets and Backprop [45 points]
Now we will try our hands on deep learning on the MNIST dataset. Let us use the square loss for the loss
function at the top layer.
We will be using 2 layer neural networks throughout, with either tanh hidden units or ReLu hidden units.
Note that should you compare your test errors to those of the neural network results in http://yann.lecun.com/exdb/mnist/
(for comparable network architectures), you should be able to significantly improve upon them.
As before, project each image onto the top 50 PCA directions. This reduces the dimension of each image
from 784 to 50. This will speed up the computation.
2
2.1 With tanh hidden units [15 points]
The input layer should have 50 dimensions (as the image is 50 dimensions, after PCA). Let us use 500 nodes
for the hidden layer, with a tanh transfer function. The output layer can simply consist of the predictions
of the network. Let us make the output nodes to be linear in the hidden layer.
1. (2 points) Specify all your parameter choices: your learning rate (or learning rate scheme), mini-batch
size, initialization scheme.
2. (6 points) Plot the squared loss after every half epoch (starting with your initial squared error). Please
label your axes in a readable way. Plot the loss of both the training and test losses on the same plot.
For the 0/1 loss, do the same, except start your plots a little later (e.g. when the 0/1 loss is below 7%)
so they are more readable.
3. (4 points) What is your final squared loss and 0/1 loss for both the training and test sets?
4. (3 points) Choose 10 hidden layer nodes at random and display the learned weights (these are the 50
dimensional weights, visualized back into image space).
2.2 With ReLu hidden units [15 points]
The input layer should have 50 dimensions (as the image is 50 dimensions, after PCA). Let us use 500 nodes
for the hidden layer, with a rectified linear (ReLu) transfer function. The output layer can simply consist of
the predictions of the network. Let us make the output nodes to be linear in the hidden layer.
1. (2 points) Specify all your parameter choices: your learning rate (or learning rate scheme), mini-batch
size, initialization scheme.
2. (6 points) Plot the squared loss after every half epoch (starting with your initial squared error). Please
label your axes in a readable way. Plot the loss of both the training and test losses on the same plot.
For the 0/1 loss, do the same, except start your plots a little later (e.g. when the 0/1 loss is below 7%)
so they are more readable.
3. (4 points) What is your final squared loss and 0/1 loss for both the training and test sets?
4. (3 points) Choose 10 hidden layer nodes at random and display the learned weights (these are the 50
dimensional weights, visualized back into image space).
2.3 With ReLu hidden units + ReLu output units [15 points]
The input layer should have 50 dimensions, and let us use 500 nodes for the hidden layer, with a rectified
linear (ReLu) transfer function. Now the output layer — the ten predictions made by the network — should
be a ReLu unit where each output is a ReLu taking as input a linear combination of the hidden layer outputs.
1. (2 points) Specify all your parameter choices: your learning rate (or learning rate scheme), mini-batch
size, initialization scheme.
2. (6 points) Plot the squared loss after every half epoch (starting with your initial squared error). Please
label your axes in a readable way. Plot the loss of both the training and test losses on the same plot.
For the 0/1 loss, do the same, except start your plots a little later (e.g. when the 0/1 loss is below 7%)
so they are more readable.
3. (4 points) What is your final squared loss and 0/1 loss for both the training and test sets?
4. (3 points) Choose 10 hidden layer nodes at random and display the learned weights (these are the 50
dimensional weights, visualized back into image space).
3
3 EM v.s. Gradient Descent [15 points]
Let us gain some intuition as to how EM differs from gradient descent. The setting is where we have a family
of distributions Pr(X, Z|θ) where θ is a vector in R
d
. We think of X as the observable variables (our data),
and Z as the latent or hidden variables.
Given the observed data X, the MLE is the solution to the following optimization problem:
θMLE = arg max
θ
L(θ) where L(θ) = log Pr(X|θ)
The EM algorithm seeks to maximize the log likelihood function, L(θ). Now let us consider gradient descent.
1. (8 points) Show that:
∇L(θ) = EZ∼Pr(Z|X,θ)∇ log Pr(X, Z|θ)
2. (3 points) Suppose starting from θ, Alice does one gradient update. Now suppose Bob, also starting
with θ, performs an E step, and then, instead of doing an exact M-step, Bob does a gradient update
on the objective function in the M-step. Both Alice and Bob use the same learning rates. Are Alice
and Bob doing the same thing? Give a brief justification of your answer.
3. (4 points) Suppose now that you run the EM algorithm to convergence (assume it converged). Do you
reach a critical point of the likelihood function? (i.e. do you reach a point where the gradient of the
log likelihood function is 0). Give a brief justification of your answer.
4 Markov Decision Processes and Dynamic Programming [15 points]
Consider an MDP with a finite set of states and a finite set of action. Denote the reward function by
R(x, a), which is the instantaneous reward at state x upon taking action a. Let Pr(x
0
|x, a) be the transition
probability of moving from state x to x
0 upon taking action a. Let 0 ≤ γ < 1 be the discount factor.
Let V denote a value function, which associates a value V (x) for each state x. Let Bell(V ) be the Bellman
update operator when applied to V . Specifically, it is defined as follow: Ve = Bell(V ) where
Ve(x) = max
a
R(x, a) + γ
X
x0
Pr(x
0
|x, a)V (x
0
)
!
Let us now prove that this update rule converges geometrically to the optimal values.
1. (8 points) Show that the Bellman operator is a contraction mapping. Specifically, show that for any
two value functions
kBell(V1) − Bell(V2)k∞ ≤ γkV1 − V2k∞
where the kZk∞ denotes the infinity norm of a vector Z, i.e. kZk∞ = maxs |Z(s)|.
2. (3 points) Suppose that upon repeated updating, we reach a fixed point, i.e. we find a V such that
Bell(V ) = V . Show that this V is unique. This V is the value of the optimal policy.
3. (4 points) Suppose we find a V such that Bell(V ) = V . Specify the optimal policy Π(x), which specifies
the action to be taken in state x, in terms of V and other relevant quantities.
4