# SARCOS and TensorFlow I had taken a break from Machine Learning: a Probabilistic Perspective in Python for some time after I got stuck on figuring out how to train a neural network. The problem Nonlinear regression for inverse dynamics was predicting torques to apply to a robot arm to reach certain points in space. There are 7 joints, but we just focus on 1 joint.

The problem claims that Gaussian process regression can get a standardized mean squared error (SMSE) of 0.011. Part (a) asks for a linear regression. This part was easy, but only could get an SMSE of 0.0742260930.

Part (b) asks for a radial basis function (RBF) network. Basically, we come up with $K$ prototypical points from the training data with $K$-means clustering. $K$ was chosen with by looking at the graph of reconstruction error. I chose $K = 100$. Now, each prototype can be though of as unit in a neural network, where the activation function is the radial basis function:

\begin{equation} \kappa(\mathbf{x}, \mathbf{x}^\prime) = \exp\left(-\frac{\lVert\mathbf{x}-\mathbf{x}^\prime\rVert^2}{2\sigma^2}\right), \end{equation}

where $\sigma^2$ is the bandwidth, which was chosen with cross validation. I just tried powers of 2, which gave me $\sigma^2 = 2^8 = 256$.

Now, the output of these activation functions is our new dataset and we just train a linear regression on the outputs. So, neural networks can pretty much be though of as repeated linear regression after applying a nonlinear transformation. I ended up with an SMSE of 0.042844306703931787.

Finally, after months I trained a neural network and was able to achieve an SMSE of 0.0217773683, which is still a ways off from 0.011, but it's a pretty admirable performance for a homework assignment versus a published journal article.

## Thoughts on TensorFlow

I had attempted to learn TensorFlow several times but repeatedly gave up. It does a great job at abstracting out building your network with layers of units and training your model with gradient descent. However, getting your data into batches and feeding it into the graph can be quite a challenge. I eventually decided on using queues, which are apparently being deprecated. I thought these were a pretty good abstraction, but since they are tied to your graph, it makes evaluating your model a pain. It makes it so that validation against a different dataset must be done in separate process, which makes a lot of sense for an production-grade training process, but it is quite difficult for beginners who sort of imagine that there should be model object that you can call like model.predict(...). To be fair, this type of API does exist with TFLearn.

It reminds me a lot of learning Java, which has a lot of features for building class hierarchies suited for large projects but confuses beginners. It took me forever to figure out what public static thing meant. After taking the time, I found the checkpointing system with monitored sessions to be pretty neat. It makes it very easy to stop and start training. In my script sarco_train.py, if you quit training with Ctrl + C, you can just start again by restarting the script and pointing to the same directory.

TensorBoard was pretty cool espsecially after I discovered that if you point your processes to the same directory, you can get plots on the same graph. I used this to plot my batch training MSE together with the MSE on the entire training set and test set. All you have to do is evaluate the tf.summary tensor as I did in the my evaluation script. MSE time series: the orange is batches from the training set. Purple is the whole training set, and blue is the test set.

Of course, you can't see much from this graph. But the cool thing is that these graphs are interactive and update in real time, so you can drill down and remove the noisy MSE from the training batches and get: MSE time series that has been filtered.

From here, I was able to see that the model around training step 190,000 performed the best. And thanks to checkpointing, that model is saved so I can restore it as I do in Selecting a model.

I thought one of the cooler features was the ability to visualize your graph. The training graph is show in the title picture and its quite complex with regularization and the Adams Optimizer calculating gradients and updating weights. The evaluation graph is much easier to look at on the other hand. It's really just a subset of the much larger training graph that stops at calculating the MSE. The training graph just includes extra nodes for regularization and optimizing.

All in all, while learning this TensorFlow was pretty frustrating, it was rewarding to finally have a model that worked quite well. Honestly, though, there is still much more to explore like training on GPUs, embeddings, distributed training, and more complicated network layers. I hope to write about doing that someday.

# Factor Analysis and Matrix Calculus Enough about my abs. Back to more important stuff, you know, like math. I've been slowly working my way through Machine Learning: a Probabilistic Perspective.

Since my last post on this topic Nearest Neighbors and Discriminant Analysis, I've gotten to do some cool problems on spam classification, Spam classification using logistic regression and Spam classification with naive Bayes. It's somewhat surprising how logistic regression performs much better than Naive Bayes with less parameters (5.8% versus 11% misclassification rate). Of course, logistic regression is a discriminative model, while Naive Bayes is generative. Estimating the conditional distribution is in some sense a smaller, and hence, easier problem. Generative models do have some advantages, though, especially when there is missing data.

Other problems have been a bit of drag. Now that I'm reading about Latent linear models and Sparse linear models, I've been getting killed by matrix calculus. I've decided to write down some of the more useful identities as a reference to myself. As evidence of how tedious these exercises are, the derivations in the textbooks and solutions manual are riddle with errors.

Some resources:

## Factor Analysis

The basic idea factor analysis model actually seemed quite intuitive when I first saw it. The idea is that underlying data is just a vector independent standard normals, that is, \begin{equation} \mathbf{z}_i \sim \mathcal{N}(\mathbf{0}, \mathbf{I}_L). \end{equation} However, we actually observe \begin{equation} \mathbf{x}_i \mid \mathbf{z}_i \sim \mathcal{N}(\boldsymbol\mu + \mathbf{W}\mathbf{z}_i, \boldsymbol\Psi), \end{equation} where $\boldsymbol\Psi$ is diagonal.

In general, I'll denote observed values with $\mathbf{x}_i$ and hidden variables with $\mathbf{z}_i$. Intuitively, I think of $\mathbf{z}_i$ as the "genetics." A single gene may affect many different traits eye color, height, hair color, bicep size, and intelligence. So if we observed 2 genes that affect those 5 traits, the entry $\mathbf{W}_{ij}$ is the effect that gene $j$ has on trait $i$. We'll denote the number of latent factors, the length of $\mathbf{z}_i$, as $L$, the number of observed factors, the length of $\mathbf{x}_i$ as $D$, and finally, the number of observations as $N$. Thus, $\mathbf{W}$ is a $D \times L$ matrix.

But why stop there? To generalize this model, we can consider a mixture of factor analysis models. Now, the underlying data is $(\mathbf{z}_i, q_i),$ where \begin{equation} q_i \sim \operatorname{Cat}(\pi_1,\pi_2,\ldots,\pi_K), \end{equation} and for each category $k \in \{1,2,\ldots,K\}$, we a separate $\mathbf{W}_k$ and $\boldsymbol\mu_k$, so that \begin{equation} \mathbf{x}_i \mid (\mathbf{z}_i, q_i = k) \sim \mathcal{N}(\boldsymbol\mu_k + \mathbf{W}_k\mathbf{z}_i, \boldsymbol\Psi). \end{equation}

One helpful way to view this is a graphical model, which I've included in the title picture. One can easily sort out the dependencies. The observed variables are shaded. The deterministic parameters are in diamonds. The latent factors are the unshaded circles with thin borders. The parameters that we are trying to estimate are given thick borders, so we want to estimate $\boldsymbol\theta = \left(\mathbf{W}_k, \boldsymbol\mu_k,\boldsymbol\Psi,\pi_k\right)$ in this case.

### Fitting the model with the EM Algorithm

From the graphical model, one can write down the probability or likelihood of the data, \begin{equation} L(\boldsymbol\theta) = p\left(\mathcal{D} \mid \mathbf{\theta}\right) = \prod_{i=1}^N \mathcal{N}\left(\mathbf{z}_i \mid \mathbf{0}, \mathbf{I}_L\right) \prod_{k^\prime=1}^K\left(\pi_{k^\prime} \mathcal{N}\left(\mathbf{x}_i \mid \mathbf{W}_k\mathbf{z}_i + \boldsymbol\mu_k, \boldsymbol\Psi\right) \right)^{I(q_i = k)}. \end{equation}

Typically, one fits models by maximizing this likelihood, or equivalently, the log-likelihood. The problem is that we can't evaluate the function if we don't know $\mathbf{z}_i$ and $q_i$. This is where the expectation–maximization (EM) algorithm comes in. We replace the unknown values with their expectation and then maximize. We do this iteratively until we achieve convergence.

Use the graphical model as a reference, we can plug in the values that are in diamonds or shaded, we are taking the expectation of the terms that involve variables in circles with a thin border, and we are choosing the values for the variables in thick borders such that the likelihood is maximized.

I won't go into the mathematical and convergence properties of this algorithm, but I'll show how it's performed. To take the expectation, we need some value for $\boldsymbol\theta$, so we choose some initial $\boldsymbol\theta_0$. At each iteration, we use $\boldsymbol\theta_l = \left(\pi_k^{(l)},\boldsymbol\mu_k^{(l)},\mathbf{W}_k^{(l)}, \boldsymbol\Psi^{(l)}\right)$ to create a better estimate $\boldsymbol\theta_{l + 1} = \left(\pi_k^{(l+1)},\boldsymbol\mu_k^{(l+1)},\mathbf{W}_k^{(l+1)}, \boldsymbol\Psi^{(l+1)}\right).$

Let's write the log-likelihood \begin{align} l(\boldsymbol\theta) &= \sum_{i=1}^N\log\mathcal{N}\left(\mathbf{z}_i \mid \mathbf{0}, \mathbf{I}_L\right) + \sum_{i=1}^N\sum_{k = 1}^K I(q_i = k)\left[\log\pi_{k} + \log\mathcal{N}\left(\mathbf{x}_i \mid \mathbf{W}_k\mathbf{z}_i + \boldsymbol\mu_k, \boldsymbol\Psi\right)\right] \label{eqn:loglikelihood}\\ &= \sum_{i=1}^N \left[-\frac{L}{2}\log 2\pi -\frac{1}{2}\mathbf{z}_i^\intercal\mathbf{z}_i \right] + \nonumber\\ &\sum_{i=1}^N\sum_{k = 1}^K I(q_i = k)\left[ \log\pi_{k} -\frac{N}{2}\log 2\pi -\frac{1}{2}\log |\boldsymbol\Psi| -\frac{1}{2}\left(\mathbf{x}_i - \mathbf{W}_k\mathbf{z}_i - \boldsymbol\mu_k\right)^\intercal{\boldsymbol\Psi}^{-1}\left(\mathbf{x}_i - \mathbf{W}_k\mathbf{z}_i- \boldsymbol\mu_k\right) \right]. \nonumber \end{align}

We need to eliminate anything with $q_i$ and $\mathbf{z}_i$ by taking expectation. Let's first start with $I(q_i = k).$

\begin{align} \mathbb{E}\left[I(q_i = k) \mid \mathbf{x}_i, \boldsymbol\theta_l\right] &= p(q_i = k \mid \mathbf{x}_i, \boldsymbol\theta_l) \nonumber\\ &= \frac{p(\mathbf{x}_i \mid q_i = k, \boldsymbol\theta_l)p(q_i = k \mid \boldsymbol\theta_l)}{\sum_{k^\prime = 1}^K p(\mathbf{x}_i \mid q_i = k^\prime, \boldsymbol\theta_l)p(q_i = k^\prime \mid \boldsymbol\theta_l)n} \label{eqn:Iq_initial} \end{align} by Bayes' rule.

Recall that $p(q_i = k \mid \boldsymbol\theta_l) = \pi_k^{(l)}$, and \begin{align} p(\mathbf{x}_i \mid q_i = k, \boldsymbol\theta_l) &= \int p(\mathbf{x}_i,\mathbf{z}_i \mid q_i = k, \boldsymbol\theta_l)d\mathbf{z}_i \nonumber\\ &= \int \mathcal{N}\left(\mathbf{x}_i \mid \mathbf{W}_k^{(l)}\mathbf{z}_i + \boldsymbol\mu_k^{(l)}, \boldsymbol\Psi^{(l)}\right)\mathcal{N}\left(\mathbf{z}_i \mid \mathbf{0}, \mathbf{I}_L\right)d\mathbf{z}_i \nonumber\\ &= \mathcal{N}\left( \mathbf{x}_i \mid \boldsymbol\mu_k^{(l)}, \boldsymbol\Psi^{(l)} + \mathbf{W}_k^{(l)}\left(\mathbf{W}_k^{(l)}\right)^\intercal\right) \label{eqn:px} \end{align} by Equation 4.126 of Murphy's textbook. Plugging in Equation \ref{eqn:px} into Equation \ref{eqn:Iq_initial}, we find that \begin{equation} r_{ik}^{(l)} = \mathbb{E}\left[I(q_i = k) \mid \mathbf{x}_i, \boldsymbol\theta_l\right] = \frac{\pi_k^{(l)}\mathcal{N}\left( \mathbf{x}_i \mid \boldsymbol\mu_k^{(l)}, \boldsymbol\Psi^{(l)} + \mathbf{W}_k^{(l)}\left(\mathbf{W}_k^{(l)}\right)^\intercal\right)} {\sum_{k^\prime=1}^K\pi_{k^\prime}^{(l)}\mathcal{N}\left( \mathbf{x}_i \mid \boldsymbol\mu_{k^\prime}^{(l)}, \boldsymbol\Psi^{(l)} + \mathbf{W}_{k^\prime}^{(l)}\left(\mathbf{W}_{k^\prime}^{(l)}\right)^\intercal\right)}. \label{eqn:Iq} \end{equation}

Next, we need to take care of the terms with $\mathbf{z}_i$. To do this, we find the conditional distribution for $\mathbf{z}_i$. Note that we can condition on both $\mathbf{x}_i$ and $q_i$ since we only care about the $\mathbf{z}_i$ terms multiplied by $I(q_i = k)$, for the other $\mathbf{z}_i$ terms disappear when we take the derivative with respect $\pi_k$, $\mathbf{W}_k$, $\boldsymbol\mu_k$, or $\boldsymbol\Psi$.

If we note that $\mathbf{x}_i \mid (\mathbf{z}_i, q_i = k, \boldsymbol\theta_l) \sim \mathcal{N}\left(\mathbf{W}_k^{(l)}\mathbf{z}_i + \boldsymbol\mu_k^{(l)}, \boldsymbol\Psi^{(l)}\right)$, and $\mathbf{z}_i \mid (q_i = k, \boldsymbol\theta_l) \sim \mathcal{N}\left(\mathbf{0}, \mathbf{I}_L\right),$ \begin{align} p(\mathbf{z}_i \mid \mathbf{x}_i, q_i = k,\boldsymbol\theta_l) &= \mathcal{N}\left(\mathbf{z}_i \mid \mathbf{m}_{ik}^{(l)}, \boldsymbol\Sigma_{ik}^{(l)}\right), \\ \text{where}~\boldsymbol\Sigma_{ik}^{(l)} &= \left(\mathbf{I}_L + \left(\mathbf{W}_k^{(l)}\right)^\intercal\left(\boldsymbol\Psi^{(l)}\right)^{-1}\mathbf{W}_k^{(l)}\right)^{-1} \nonumber\\ \text{and}~\mathbf{m}_{ik}^{(l)} &= \boldsymbol\Sigma_{ik}^{(l)} \mathbf{W}_k^{(l)}\left(\boldsymbol\Psi^{(l)}\right)^{-1}\left(\mathbf{x}_i - \boldsymbol\mu_k^{(l)}\right)\nonumber \end{align} by Equation 4.125 in Murphy's textbook.

Now, let's do a couple of things to clean up notation. First, to simply Equation \ref{eqn:loglikelihood}, we'll drop all terms that aren't functions of $\boldsymbol\theta$, so we have \begin{equation} \tilde{l}(\boldsymbol\theta) = \sum_{i=1}^N\sum_{k = 1}^K I(q_i = k)\left[ \log\pi_{k} -\frac{1}{2}\log |\boldsymbol\Psi| -\frac{1}{2}\left(\mathbf{x}_i - \mathbf{W}_k\mathbf{z}_i - \boldsymbol\mu_k\right)^\intercal{\boldsymbol\Psi}^{-1}\left(\mathbf{x}_i - \mathbf{W}_k\mathbf{z}_i- \boldsymbol\mu_k\right) \right]. \label{eqn:loglikelihood1} \end{equation}

In the next step, we define \begin{equation} \tilde{\mathbf{z}}_i = \begin{pmatrix} \mathbf{z}_i \\ 1 \end{pmatrix},~\text{and}~ \tilde{\mathbf{W}}_k = \begin{pmatrix} \mathbf{W}_k & \boldsymbol\mu_k \end{pmatrix}. \end{equation} Now $\boldsymbol\theta = (\tilde{\mathbf{W}}_k, \boldsymbol\Psi, \pi_k)$, and we can rewrite Equation \ref{eqn:loglikelihood1} as \begin{align} \tilde{l}(\boldsymbol\theta) &= \sum_{i=1}^N\sum_{k = 1}^K I(q_i = k)\left[ \log\pi_{k} -\frac{1}{2}\log |\boldsymbol\Psi| -\frac{1}{2}\left(\mathbf{x}_i - \tilde{\mathbf{W}}_k\tilde{\mathbf{z}}_i\right)^\intercal{\boldsymbol\Psi}^{-1}\left(\mathbf{x}_i - \tilde{\mathbf{W}}_k\tilde{\mathbf{z}}_i\right) \right] \nonumber\\ &= \sum_{i=1}^N\sum_{k = 1}^K I(q_i = k)\left[ \log\pi_{k} -\frac{1}{2}\log |\boldsymbol\Psi| -\frac{1}{2}\left( \mathbf{x}_i^\intercal\boldsymbol\Psi^{-1}\mathbf{x}_i - 2\mathbf{x}_i^\intercal\boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\tilde{\mathbf{z}}_i + \tilde{\mathbf{z}}_i^\intercal\tilde{\mathbf{W}}_k^\intercal\boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\tilde{\mathbf{z}}_i \right) \right] \nonumber\\ &= \sum_{i=1}^N\sum_{k = 1}^K I(q_i = k)\left[ \log\pi_{k} -\frac{1}{2}\log |\boldsymbol\Psi| -\frac{1}{2}\left( \mathbf{x}_i^\intercal\boldsymbol\Psi^{-1}\mathbf{x}_i - 2\mathbf{x}_i^\intercal\boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\tilde{\mathbf{z}}_i + \operatorname{tr}\left(\tilde{\mathbf{W}}_k^\intercal\boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\tilde{\mathbf{z}}_i\tilde{\mathbf{z}}_i^\intercal \right)\right) \right] \end{align} by cyclic property the trace.

Now, we note that \begin{equation} \tilde{\mathbf{z}}_i \mid \left(\mathbf{x}_i, q_i = k,\boldsymbol\theta_l\right) \sim \mathcal{N}\left( \begin{pmatrix} \mathbf{m}_{ik}^{(l)} \\ 1 \end{pmatrix}, \begin{pmatrix} \boldsymbol\Sigma_{ik}^{(l)} & \mathbf{0} \\ \mathbf{0}^\intercal & 0 \end{pmatrix} \right). \end{equation} Using this, we'll have that \begin{align} \mathbf{b}_{ik}^{(l)} &= \mathbb{E}\left[\tilde{\mathbf{z}}_i \mid \mathbf{x}_i ,q_i = k, \boldsymbol\theta_l\right] =\begin{pmatrix} \mathbf{m}_{ik}^{(l)} \\ 1 \end{pmatrix} \\ \mathbf{C}_{ik}^{(l)} &= \mathbb{E}\left[\tilde{\mathbf{z}}_i\tilde{\mathbf{z}}_i^\intercal \mid \mathbf{x}_i ,q_i = k, \boldsymbol\theta_l\right] = \begin{pmatrix} \boldsymbol\Sigma_{ik}^{(l)} & \mathbf{b}_{ik}^{(l)} \\ \left(\mathbf{b}_{ik}^{(l)}\right)^\intercal & 1 \end{pmatrix}. \end{align}

Thus, E-step becomes writing our objective function as \begin{equation} Q_{\boldsymbol\theta_l}(\boldsymbol\theta) = \sum_{i=1}^N\sum_{k = 1}^K r_{ik}^{(l)}\left[ \log\pi_{k} -\frac{1}{2}\log |\boldsymbol\Psi| -\frac{1}{2}\left( \mathbf{x}_i^\intercal\boldsymbol\Psi^{-1}\mathbf{x}_i - 2\mathbf{x}_i^\intercal\boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\mathbf{b}_{ik}^{(l)} + \operatorname{tr}\left(\tilde{\mathbf{W}}_k^\intercal\boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\mathbf{C}_{ik}^{(l)} \right)\right) \right]. \label{eqn:estep} \end{equation}

We want to maximize $Q$ to obtain \begin{equation} \boldsymbol\theta_{l+1} = \operatorname*{arg\,max}_{\boldsymbol\theta} Q_{\boldsymbol\theta_l}(\boldsymbol\theta) \end{equation} for the M-step. This can be done by taking derivatives.

For $\pi_k$ this, is not that hard. Let $\pi_K = 1 - \sum_{k=1}^{K-1}\pi_k$, which gives us that \begin{equation} \frac{\partial}{\partial\pi_k}Q_{\boldsymbol\theta_l}(\boldsymbol\theta) = \sum_{i=1}^N \left(\frac{r_{ik}^{(l)}}{\pi_k} - \frac{r_{iK}^{(l)}}{\pi_K}\right) \end{equation} for $k < K$. Setting this equal to $0$, we can solve for $\hat{\boldsymbol\pi}$. \begin{align*} \hat\pi_K\sum_{i=1}^N r_{ik}^{(l)} &= \hat\pi_k\sum_{i=1}^N r_{iK}^{(l)} \\ \hat\pi_K\sum_{i=1}^N\left(r_{i1}^{(l)} + r_{i2}^{(l)} + \cdots + r_{i,K-1}^{(l)}\right) &= \left(\hat\pi_1 + \hat\pi_2 + \cdots + \hat\pi_{K-1}\right)\sum_{i=1}^N r_{iK}^{(l)} \\ \hat\pi_K\sum_{i=1}^N\left(1 - r_{iK}^{(l)}\right) &= \left(1 - \hat\pi_K\right)\sum_{i=1}^N r_{iK}^{(l)} \\ N\hat\pi_K - \hat\pi_K\sum_{i=1}^N r_{iK}^{(l)} &= \sum_{i=1}^N r_{iK}^{(l)} - \hat\pi_K\sum_{i=1}^N r_{iK}^{(l)} \\ \hat\pi_K &= \frac{1}{N}\sum_{i=1}^N r_{iK}^{(l)}. \end{align*}

By symmetry, we have that \begin{equation} \pi_k^{(l + 1)} = \hat\pi_k = \frac{1}{N}\sum_{i=1}^N r_{ik}^{(l)} \label{eqn:mpi} \end{equation} for any $k$.

For $\boldsymbol\Psi$, we need several matrix identities. The first one is \begin{equation} \frac{\partial}{\partial \mathbf{A}} \log |\mathbf{A}| = \left(\mathbf{A}^{-1}\right)^\intercal. \label{eqn:mat_det} \end{equation}

To see this, note that we can write $|\mathbf{A}| = \sum_{i=1}^N \left(-1\right)^{i+j}\mathbf{A}_{ij}\left|\mathbf{A}_{-i,-j}\right|$ for any $j$, so we have that \begin{align*} \frac{\partial}{\partial \mathbf{A}_{ij}} \log |\mathbf{A}| &= \frac{1}{|\mathbf{A}|}(-1)^{i+j}\left|\mathbf{A}_{-i,-j}\right| \\ &= \frac{1}{|\mathbf{A}|}\mathbf{C}_{ij} \\ &= \left(\mathbf{A}^{-1}\right)^\intercal_{ij}, \end{align*} where we have used the definition of the matrix inverse in terms of the adjugate matrix, and $\mathbf{C}$ is the cofactor matrix.

Next, we prove that \begin{equation} \frac{\partial}{\partial \mathbf{A}}\mathbf{x}^\intercal \mathbf{A}\mathbf{y} = \mathbf{x}\mathbf{y}^\intercal. \label{eqn:mat_quad} \end{equation} To see this we rewrite \begin{equation*} \mathbf{x}^\intercal \mathbf{A}\mathbf{y} = \operatorname{tr}\left(\mathbf{x}^\intercal \mathbf{A}\mathbf{y}\right) = \operatorname{tr}\left(\mathbf{A}\mathbf{y}\mathbf{x}^\intercal\right) = \sum_{i=1}^N\sum_{k=1}^N \mathbf{A}_{ik}y_kx_i, \end{equation*} which implies that \begin{equation*} \frac{\partial}{\partial \mathbf{A}_{ij}}\mathbf{x}^\intercal \mathbf{A}\mathbf{y} = x_iy_j. \end{equation*}

Now, the last trick is to rewrite $\log|\boldsymbol\Psi| = -\log\left|\boldsymbol\Psi^{-1}\right|$, and note that the MLE is preserved by parametrization. So, using Equations \ref{eqn:mat_det} and \ref{eqn:mat_quad}, we can take the derivative of Equation \ref{eqn:estep} with respect to $\boldsymbol\Psi^{-1}$ to get \begin{align*} \frac{\partial}{\partial\boldsymbol\Psi^{-1}}Q_{\boldsymbol\theta_l}(\boldsymbol\theta) &= \sum_{i=1}^N\sum_{k = 1}^K r_{ik}^{(l)}\left[ \frac{1}{2}\boldsymbol\Psi -\frac{1}{2}\left(\mathbf{x}_i\mathbf{x}_i^\intercal - 2\tilde{\mathbf{W}}_k\mathbf{b}_{ik}^{(l)}\mathbf{x}_i^\intercal + \tilde{\mathbf{W}}_k\mathbf{C}_{ik}^{(l)}\tilde{\mathbf{W}}_k^\intercal\right) \right]. \end{align*} If we set this equal to $0$, we find that \begin{align*} \hat{\boldsymbol\Psi} = \frac{1}{N}\sum_{i=1}^N\sum_{k = 1}^K r_{ik}^{(l)}\left( \mathbf{x}_i\mathbf{x}_i^\intercal - 2\tilde{\mathbf{W}}_k\mathbf{b}_{ik}^{(l)}\mathbf{x}_i^\intercal + \tilde{\mathbf{W}}_k\mathbf{C}_{ik}^{(l)}\tilde{\mathbf{W}}_k^\intercal \right), \end{align*} where we have used the symmetry of $\boldsymbol\Psi$.

We have two issues. $\boldsymbol\Psi$ is suppose to be diagonal, and we need to choose a value for $\tilde{\mathbf{W}}_k$. To enforce the diagonal constraint, we just take the diagonal of $\hat{\boldsymbol\Psi}$ since that we could have just taken derivatives component-wise. For $\tilde{\mathbf{W}}_k$, we use $\tilde{\mathbf{W}}_k^{(l+1)}$, which turns out to not depend on $\boldsymbol\Psi,$ so we have \begin{equation} \boldsymbol\Psi^{(l+1)} = \frac{1}{N}\sum_{i=1}^N\sum_{k = 1}^K r_{ik}^{(l)}\operatorname{diag}\left( \mathbf{x}_i\mathbf{x}_i^\intercal - 2\tilde{\mathbf{W}}_k^{(l+1)}\mathbf{b}_{ik}^{(l)}\mathbf{x}_i^\intercal + \tilde{\mathbf{W}}^{(l+1)}_k\mathbf{C}_{ik}^{(l)}\left(\tilde{\mathbf{W}}_k^{(l+1)}\right)^\intercal \right). \label{eqn:mpsi} \end{equation}

Now, we need to deal with the various $\tilde{\mathbf{W}}_k$. I'm not going to prove this identity, but we have that \begin{equation} \frac{\partial}{\partial \mathbf{X}}\operatorname{tr}\left(\mathbf{X}^\intercal \mathbf{B} \mathbf{X}\mathbf{C}\right) = \mathbf{B}\mathbf{X}\mathbf{C} + \mathbf{B}^\intercal \mathbf{X} \mathbf{C}^\intercal, \label{eqn:trw} \end{equation} which is Equation 117 in the The Matrix Cookbook. Taking the derivative of Equation \ref{eqn:estep} with respect to $\tilde{\mathbf{W}}_k$, we get \begin{align} \frac{\partial}{\partial\tilde{\mathbf{W}}_k}Q_{\boldsymbol\theta_l}(\boldsymbol\theta) &= -\frac{1}{2}\sum_{i=1}^N r_{ik}^{(l)} \left( 2\boldsymbol\Psi^{-1}\mathbf{x}_i\left(\mathbf{b}_{ik}^{(l)}\right)^\intercal - \boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\mathbf{C}_{ik}^{(l)} - \boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\mathbf{C}_{ik}^{(l)} \right) \nonumber\\ &= \sum_{i=1}^N r_{ik}^{(l)} \left(\boldsymbol\Psi^{-1}\tilde{\mathbf{W}}_k\mathbf{C}_{ik}^{(l)} - \boldsymbol\Psi^{-1}\mathbf{x}_i\left(\mathbf{b}_{ik}^{(l)}\right)^\intercal \right), \end{align} where we have used the symmetry of $\boldsymbol\Psi$ and $\mathbf{C}_{ik}^{(l)}$. Setting this equal to $0$, we find that \begin{equation} \tilde{\mathbf{W}}_k^{(l+1)} = \left(\sum_{i=1}^N r_{ik}^{(l)}\mathbf{x}_i\left(\mathbf{b}_{ik}^{(l)}\right)^\intercal\right) \left(\sum_{i=1}^N r_{ik}^{(l)}\mathbf{C}_{ik}^{(l)} \right)^{-1}. \label{eqn:mw} \end{equation}

All in all, we combine Equations \ref{eqn:mpi}, \ref{eqn:mpsi}, and \ref{eqn:mw} to get $\boldsymbol\theta_{l+1}$, so our M-step is \begin{align} \pi_k^{(l + 1)} &= \hat\pi_k = \frac{1}{N}\sum_{i=1}^N r_{ik}^{(l)} \\ \tilde{\mathbf{W}}_k^{(l+1)} &= \left(\sum_{i=1}^N r_{ik}^{(l)}\mathbf{x}_i\left(\mathbf{b}_{ik}^{(l)}\right)^\intercal\right) \left(\sum_{i=1}^N r_{ik}^{(l)}\mathbf{C}_{ik}^{(l)} \right)^{-1} \\ \boldsymbol\Psi^{(l+1)} &= \frac{1}{N}\sum_{i=1}^N\sum_{k = 1}^K r_{ik}^{(l)}\operatorname{diag}\left( \mathbf{x}_i\mathbf{x}_i^\intercal - 2\tilde{\mathbf{W}}_k^{(l+1)}\mathbf{b}_{ik}^{(l)}\mathbf{x}_i^\intercal + \tilde{\mathbf{W}}^{(l+1)}_k\mathbf{C}_{ik}^{(l)}\left(\tilde{\mathbf{W}}_k^{(l+1)}\right)^\intercal \right) . \end{align}

## Principal Component Analysis

Now, $\boldsymbol\Psi$ begin diagonal is already a pretty strong assumption. Principal Component Analysis (PCA) makes the even stronger assumption that $\boldsymbol\Psi = \sigma^2 \mathbf{I}_D$. For this reason it's much easier to compute. All it takes is a singular value decomposition (SVD).

A nice application of these methods is Latent Semantic Indexing. Here, we've represented 9 documents as a word count vector. There are nearly 500 words. Using only 2 latent factors, we can cluster the documents. Documents 1, 2, and 3 are about alien abductions.

Unfortunately, the assumption that the variance does not differ between the dimensions can be a bad one. In Probabilistic Principal Component Analysis versus Factor Analysis, we see how PCA fails to capture the relationship between $x_1$ and $x_2$ since it reduces the variance by focusing on $x_3$.

# Nearest Neighbors and Discriminant Analysis To avoid facing the ticking clock, I've decided to read a textbook and start a new project Machine Learning: a Probabilistic Perspective in Python. I've been meaning to learn more about the math behind machine learning and the Python ecosystem for scientific computing, so I've decided to code up various solutions for the textbook in Jupyter notebooks. Here are a couple of the more interesting exercises.

## Cross Validation and Nearest Neighbors

A classic problem in machine learning that is often used to benchmark classification algorithms is digit recognition. The standard dataset is The MNIST database of handwritten digits. A digit is represented as $26 \times 26$ grid of grayscale pixels like the following examples. Surprisingly, a simple nearest neighbor approach works well despite the curse of dimensionality. When using $1$ neighbor and euclidean distance, we find that the error rate is only 3.8% on 1,000 test cases. Here are some correct predictions. Of course we get some predictions wrong. Here are some. The notebook to generate these figures is here, KNN classifier on shuffled MNIST data.

### Model Selection

How should we know how many neighbors to use? For a parametric models, we can pick our models using a maximum likelihood estimate (MLE) or a Bayesian approach, using either the posterior mean or maximum a posteriori probability (MAP) estimate. However, nearest neighbors is a nonparametric model since the number of parameters increases with the amount of data.

To measure model performance, we use empirical measures, that is, we directly measure prediction accuracy. There are a few ways of going about this.

1. We can test the accuracy on the data that we trained on. This is called the training set error. Generally, it's not a very good metric since it's prone to overfitting. For instance, using 1 neighbor leads to 0% error.
2. We can withhold some of our data and call it the test set. We fit our model on the data not in the test set and calculate the misclassification rate on the test set. This is the test set error. It generally gives a good idea how the model will perform with new data, but it cannot be used when you have a limited number of test cases and cannot afford to throw away data.
3. A compromise is the cross validation error. Here, we split our data into $k$ folds. For each fold, we fit our model on the remaining $k - 1$ folds and use the chosen fold as our test set. In this way, we compute $k$ different error rates. Then, we take the average of those error rates.

All three approaches were tried and compared in this notebook, CV for KNN. The test set error and cross validation error are quite close, which indicates that the cross validation error is a good proxy for the test set error when faced with limited data. The training set error will almost always underpredict the true error rate and is often vulnerable to overfitting. In this rare case, all three error rates actually agree that using 1 neighbor is optional, but usually, model that achieves the best training set error is overfitted to the data and predicts future test cases badly.

## Discriminant Analysis

Some of the simplest models are the various flavors of discriminant analysis. Suppose we have classes $k = 1,2,\ldots,K$. We make $N$ obervations $(\mathbf{x}_i, y_i)$, where $i \in \{1,2,\ldots, N\}$ and $y_i \in \{1,2,\ldots,K\}$. $\mathbf{x}_i$ is our data and $y_i$ is our class label. Based off some new data $\mathbf{x}^\prime$, we would want to predict its class label $y^\prime$.

Discriminant analysis makes the strong assumption that $\mathbf{x}$ is generated from a multivariate normal distribution. The distribution has different parameters for each class, so if $\mathbf{x}$ belongs to class $y = k$, $\left(\mathbf{x} \mid y = k\right) \sim \mathcal{N}(\boldsymbol\mu_k,\Sigma_k)$. An example would be trying to predict the gender of a person after observing their height and weight. Using the MLE, we can fit a multivariate normal to males as I do in the notebook Whitening versus standardizing. We can do a similar thing for our female observations. Thus, given an observation $\mathbf{x}$, we have estimated the distribution of $\mathbf{x} \mid y = k$. Now, the distribution of $y$ is just a multinomial distribution, with parameters $\boldsymbol\pi \in \left\{(\pi_1,\pi_2,\ldots, pi_K) \in \mathbb{R}^K : \sum_{k=1}^Kx_k = 1,~0 < x_k < 1\right\}$. The MLE estimate for $\boldsymbol\pi$ is just \begin{equation} \boldsymbol{\hat{\pi}} = \left(\frac{N_1}{N}, \frac{N_2}{N},\ldots,\frac{N_K}{N}\right), \end{equation} where $N_k$ is the number of our observations of class $k$.

Let $f(\mathbf{x} \mid \boldsymbol\mu,\Sigma)$ be the probability density function of a multivariate normal with mean $boldsymbol\mu$ and covariance $\Sigma$. Now, we apply Bayes' rule which gives us that \begin{align} p(y = k \mid \mathbf{x}) &= \frac{p(\mathbf{x} \mid y = k)p(y=k)}{\sum_{k^\prime = 1}^Kp(\mathbf{x} \mid y = k^\prime)p(y=k^\prime)}\nonumber\\ &= \frac{f(\mathbf{x} \mid \boldsymbol{\hat{\mu}}_k, \hat{\Sigma}_k)\frac{N_k}{N}}{\sum_{k^\prime = 1}^K f(\mathbf{x} \mid \boldsymbol{\hat{\mu}}_{k^\prime}, \hat{\Sigma}_{k^\prime})\frac{N_{k^\prime}}{N}} \nonumber\\ &= \frac{f(\mathbf{x} \mid \boldsymbol{\hat{\mu}}_k, \hat{\Sigma}_k)N_k}{\sum_{k^\prime = 1}^K f(\mathbf{x} \mid \boldsymbol{\hat{\mu}}_{k^\prime}, \hat{\Sigma}_{k^\prime})N_{k^\prime}}. \end{align}

This gives us the quadratic disriminant analysis (QDA) classifier. Now, we can fix $\Sigma_k = \Sigma$ for all $k$, which leads to further cancellations. This gives the linear discriminant analysis (LDA) classifier. The names quadractic and linear come from the shape of boundaries as seen in the notebook LDA/QDA on height/weight data. In this case, they lead to the same misclassification rate.

See more notebooks as I add them at Machine Learning: a Probabilistic Perspective in Python.

# 2016 Republican Candidates on TV in 2015 Happy New Years, everyone! As a way to start off the year, I thought that it would be interesting to write about something that has evolved a lot over the past year: the Republican field of presidential candidates.

At Penn, I've been working on Snapstream Searcher, which searches through closed captioning television scripts. I've decided to see how often a candidate is mentioned on TV has changed over the year. Check out the chart and do your own analysis here.

As you can see in the title picture, Donald Trump has surged in popularity since he announced his candidicy in June. In general every candidate, experiences a surge in mentions upon announcing his or her candidacy. Usually the surge is not sustained, though.

Many candidates lost popularity over the course of 2015. Jeb Bush lost quite a bit of ground, but perhaps no one has suffered as much as Chris Christie. Other candidates like Ben Carson are passing fads with a bump from Octorber to November before fading away: Some cool features I added are the ability to zoom. The D3 brush calculates the coordinates, and then, I update the scales and axes. The overflow is hidden with SVG clipping. To illustrate the usefulness of this feature, we can focus in on the September debate. Here, we see Carly Fiorina's bump in popularity due to her strong debate performance.

Another cool feature that I added was the ability to see actual data points. If one holds the Control key and hovers over the point, we can see a tooltip. Play around with it here, and let me know what you think!

# Last Visit to $0$ of a Simple Random Walk One of the more interesting problems that I've come across recently is to calculate the distribution of the last time a simple random walk is at $0.$

Let $X_1,X_2,\ldots,X_{2n}$ be independent, indentically distributed random variables such that $P(X_i = 1) = P(X_i = -1) = 1/2.$ Define $S_k = \sum_{i=1}^k X_i.$ Then, we have a path $$(0,0) \rightarrow (1,S_1) \rightarrow (2,S_2) \rightarrow \cdots \rightarrow (2n,S_{2n}).$$ Define the random variable $L_{2n} = \sup\{ k \leq 2n : S_k = 0\}.$ We want the distribution of $L_{2n}.$

Note that we have that \begin{equation} \mathbb{P}(S_{2n} = 0) = 2^{-2n}{2n \choose n} \end{equation} since we have $n$ positive steps and $n$ negative steps.

Let $\displaystyle N_{n,x} ={n \choose (n+x)/2}$ denote the number of paths from $(0,0)$ to $(n,x)$ since $(n+x)/2$ positive steps implies there are $(n-x)/2$ negative steps, and $(n+x)/2 - (n-x)/2 = x$. Note that $n + x$ must be even for this to be well defined. If $n + x$ is not even, then $N_{n,x} = 0$ since $x$ must have the same parity as $n.$ First, we prove the reflection principle.

### Reflection Principle

If $x,y > 0,$ the number of paths that from $(0,x)$ to $(n,y)$ that are $0$ at some time, that is, they touch the $x$-axis is equal to the total number of paths from $(0,-x)$ to $(n,y),$ which is $N_{n,y+x}.$ Therefore, the number of paths from $(0,x)$ to $(n,y)$ that do not touch $0$ is $N_{n,|y-x|} - N_{n,y+x}.$

We can establish a one-to-one correspondence between the set $A$, the paths from $(0,x)$ to $(n,y)$ that are $0$ at some time and the set $B$ the paths from $(0,-x)$ to $(n,y)$.

Consider any path $P$ in $A$. $P$ must include the point $(m,0),$ where $0 < m < n$. Fix $m$ to be the greatest such integer. We construct a path $Q_1$ from $(0,-x)$ to $(m,0)$ by going in the opposite direction as $P.$ We construct a path $Q_2$ from $(m,0)$ to $(n,y)$ by mirroring $P$. Thus, we have that $Q = Q_1 \cup Q_2 \in B.$

Now consider any path $Q$ in $B$. Since paths are continuous, $Q$ must cross the $x$-axis, so $Q$ includes a point $(m,0)$, where $0 < m < n.$ Fix $m$ to be the greatest such integer. We construct $P_1,$ a path from $(0,x)$ to $(m,0)$ by going in the opposite direction as $Q$. We construct $P_2$ by mirroring $Q$. Thus, we have that $P = P_1 \cup P_2 \in A.$

So, we have established a one-to-one correspondence, and therefore, we have proven $|A| = |B|.$

### Symmetry of Zeroes

$\mathbb{P}(S_1 \neq 0, S_2 \neq 0,\ldots,S_{2n} \neq 0) = \mathbb{P}(S_{2n} = 0).$

First note that $$\mathbb{P}(S_1 \neq 0,\ldots,S_{2n} \neq 0) = \mathbb{P}(S_1 > 0,\ldots,S_{2n} > 0) + \mathbb{P}(S_1 < 0,\ldots,S_{2n} < 0)$$ since we can never have the path touch $0.$ Also note that the two terms are equal, so \begin{equation} \mathbb{P}(S_1 \neq 0, S_2 \neq 0,\ldots,S_{2n} \neq 0) = 2\mathbb{P}(S_1 > 0, S_2 > 0,\ldots,S_{2n} > 0). \end{equation} Now, note that \begin{equation} \mathbb{P}(S_1 > 0, S_2 > 0,\ldots,S_{2n} > 0) = \sum_{r=1}^{n}\mathbb{P}(S_1 > 0, S_2 > 0,\ldots,S_{2n} = 2r) \end{equation} since we have taken an even number of steps.

To calculate $\mathbb{P}(S_1 > 0, S_2 > 0,\ldots,S_{2n} = 2r),$ we note that $X_1 = 1$ since $S_1 > 0$. Then, the number of paths from $(1,1)$ to $(2n,2r)$ that do not touch $0$ by the Reflection Principle is $N_{2n-1,2r-1} - N_{2n-1,2r+1}$. Thus, \begin{align*} \mathbb{P}(S_1 > 0, S_2 > 0,\ldots,S_{2n} = 2r) &= \left(\frac{1}{2}\right)\left(\frac{1}{2^{2n-1}}\right)\left(N_{2n-1,2r-1} - N_{2n-1,2r+1}\right) \\ &= \left(\frac{1}{2^{2n}}\right)\left(N_{2n-1,2r-1} - N_{2n-1,2r+1}\right). \end{align*}

So, we have that \begin{align*} \mathbb{P}(S_1 > 0, S_2 > 0,\ldots,S_{2n} > 0) &= \left(\frac{1}{2^{2n}}\right) \sum_{r=1}^n\left(N_{2n-1,2r-1} - N_{2n-1,2r+1}\right) \\ &= \frac{1}{2^{2n}}N_{2n-1,1} = \frac{1}{2^{2n}}{2n - 1 \choose n}\\ &= \frac{1}{2^{2n}}\frac{(2n-1)!}{n!(n-1)!} = \frac{1}{2}\frac{1}{2^{2n}}\frac{(2n)!}{n!n!} \\ &= \frac{1}{2}\mathbb{P}(S_{2n} = 0) \end{align*} by telescoping since $N_{2n-1,2n+1} = 0$ and substituting in the previous equations.

Recalling the earlier equation, we have the desired result, $$\mathbb{P}(S_1 \neq 0, S_2 \neq 0,\ldots,S_{2n} \neq 0) = 2\mathbb{P}(S_1 > 0, S_2 > 0,\ldots,S_{2n} > 0) = \mathbb{P}(S_{2n} = 0).$$

### Distribution of Last Zero Visit

Let $L_{2n}$ be the random variable whose value is the last time the simple random walk visited $0.$ Formally, $$L_{2n} = \sup\left\{2k : k \in \{0,\ldots,n\},~\sum_{i=1}^{2k} X_i = 0\right\}.$$ Then, we have that $$\mathbb{P}(L_{2n} = 2k) = 2^{-2n}{2k \choose k}{2n-2k \choose n-k}.$$

To see this, we have that \begin{align*} \mathbb{P}(L_{2n} = 2k) &= \mathbb{P}(S_{2k} = 0, S_{2k+1} \neq 0, \ldots,S_{2n} \neq 0) \\ &= \mathbb{P}(S_{2k} = 0, S_{2k+1} - S_{2k} \neq 0, \ldots,S_{2n} - S_{2k} \neq 0) \\ &= \mathbb{P}(S_{2k} = 0)\mathbb{P}(S_1 \neq 0, \ldots,S_{2n-2k} \neq 0) \\ &= \mathbb{P}(S_{2k} = 0)\mathbb{P}(S_{2n-2k} = 0), \end{align*} where the last equaility is by Symmetry of Zeroes. Thus, we have that $$\mathbb{P}(L_{2n} = 2k) = \mathbb{P}(S_{2k} = 0)\mathbb{P}(S_{2n-2k} = 0) =2^{-2n}{2k \choose k}{2n-2k \choose n-k}$$ as desired.

### Conculusion

Let's look at what this distribution looks like when $n = 15.$ We have symmetric distribution about $15$, so the mean is $15.$ However, the distribution is U-shaped, and the most likely values are very far from the mean. Thus, to take an analogy from sports, if two evenly matched teams were to played against each other multiple times over the course of a season, the most likely scenario is for one team team to lead the other team the entire season. So, saying that team A led team B the entire season is an almost meaningless statistic.

# CrossFit and Pivot Charts Back when I lived in Boston, I was an avid CrossFitter. For a variety of reasons, mainly financial, I no longer go to a CrossFit box, but I'm still interested in the sport. Being a bit of a data nerd, I've always been curious about what makes an elite CrossFitter and how much height and weight play a role. To satisfy my curiosity, I scraped data on the top 2,000 athletes from CrossFit Games, and created a pivot chart in D3.js, where you can compare statistics on workouts and lifts by different groups of athletes.

Play around with the data yourself at 2015 CrossFit Open Pivot Chart. Be careful. The data may not be that reliable. If there are a lot of outliers, it may be better to use a robust statistic like median instead of mean (in particular, Sprint 400m and Run 5k workouts seem to have this problem). If you don't choose your groups wisely, you may fall into Simpson's Paradox by excluding important data. For example, from the chart below, one may conclude that back squat strength decreases with age.

Mean Back Squat by Age Group But now, when we consider gender, we have an entirely different story:

Mean Back Squat by Gender and Age Group Unsurprisingly, women back squat less than men do. Back squat strength remains stable with age for women, and if anything, back squat strength actually increases slightly with age for men. Whoa, what's going on here? Check this out:

Gender and Age of Top CrossFit Athletes Notice that the 3 rightmost female bars (red) are taller than the 3 rightmost male bars (blue). On the other hand, the 3 leftmost female bars are shorter than the 3 leftmost male bars. Thus, it seems that women age better than men in the CrossFit world. This has the implication that there are more women than men in older age groups, so the average back squat of that group appears to be lower, when in reality, there simply are a greater relative number of women in that group. You can reach the same conclusion from the title picture, where the bars are stacked.

## Height and Weight

There definitely seems to be a prototypical build for an elite CrossFit athlete. For men it's about 5'10" and 200 lb, with a lot of athletes just over 6 feet, too.

Male Height and Weight For women, most athletes seem to be about 5'6" and 145 lb, which happen to be my dimensions. Smaller female athletes that barely break 5 feet are pretty well-represented, too. I was somewhat surprised at the lack of taller women.

Female Height and Weight Some open workouts like 1A, which was a one-rep max clean and jerk favored larger athletes:

Open Workout 1A by Height (One-rep Max Clean & Jerk) Other open workouts like 4, which was an ascending rep ladder of cleans and handstand push-ups, favored smaller athletes:

Open Workout 4 by Height (Cleans and Handstand Push-ups) You can check the other workouts yourself, but overall, this year's open seemed fair with regard to athlete size.

Anyway, feel free to play with the data yourself here. Let me know if you find anything cool and if you have any suggestions for improving usability.