## Dimensionality reduction 101: linear algebra, hidden variables and generative models

Suppose you are faced with a high dimensional dataset and want to find some structure in the data: often there are only a few causes, but lots of different data points are generated due to noise corruption. How can we infer these causes? Here I’m going to cover the simplest method to do this inference: we will assume the data is generated by a linear transformation of the hidden causes. In this case, it is quite simple to recover the parameters of this transformation and therefore determine the hidden (or latent) variables which represent their cause.

## An introduction to the metropolis method with python

I already talked about MCMC methods before, but today I want to cover one of the most well known methods of all, Metropolis-Hastings. The goal is to obtain samples according to to the equilibrium distribution of a given physical system, the Boltzmann distribution. Incidentally, we can also rewrite arbitrary probability distributions in this form, which is what allows for the cross pollination of methods between probabilistic inference and statistical mechanics (look at my older post on this). Since we don’t know how to sample directly from the boltzmann distribution in general, we need to use some sampling method. Continue reading “An introduction to the metropolis method with python”

## Simple pattern formation with cellular automata

A cellular automaton is a dynamical system where space, time and dynamic variable are all discrete. The system is thus composed of a lattice of cells (discrete space), each described by a state (discrete dynamic variable) which evolve into the next time step (discrete time) according to a dynamic rule.

x_i^{t+1} = f(x_i^t, \Omega_i^t, \xi)

This rule generally depends on the state of the target cell $x_i^t$, the state of its neighbors $\Omega_i^t$, and a number of auxiliary external variables $\xi$. Since all these inputs are discrete, we can enumerate them and then define the dynamic rule by a transition table. The transition table maps each possible input to the next state for the cell. As an example consider the elementary 1D cellular automaton. In this case the neighborhood consists of only the 2 nearest neighbors $\Omega_i^t = \{x_{i-1}^t, x_{i+1}^t\}$ and no external variables.

In general, there are two types of neighborhoods, commonly classified as Moore or Von Neumann. A Moore neighborhood of radius $r$ corresponds to all cells within a hypercube of size $r$ centered at the current cell. In 2D we can write it as $\Omega_{ij}^t = \{x^t_{kl}:|i-k|\leq r \wedge |j-l|\leq r\}\setminus x^t_{ij}$. The Von Neumann neighborhood is more restrictive: only cells within a manhattan distance of $r$ belong to the neighborhood. In 2D we write $\Omega_{ij}^t = \{x^t_{kl}:|i-l|+|j-k| \leq r\}\setminus x^t_{ij}$.

Finally it is worth elucidating the concept of totalistic automata. In high dimensional spaces, the number of possible configurations of the neighborhood $\Omega$ can be quite large. As a simplification, we may consider instead as an input to the transition table the sum of all neighbors in a specific state $N_k = \sum_{x \in \Omega}\delta(x = k)$. If there are only 2 states, we need only consider $N_1$, since $N_0 = r – N_1$. For an arbitrary number $m$ of states, we will obviously need to consider $m-1$ such inputs to fully characterize the neighborhood. Even then, each input $N_k$ can take $r+1$ different values, which might be too much. In such cases we may consider only the case when $N_k$ is above some threshold. Then we can define as an input the boolean variable

P_{k,T}=\begin{cases}
1& \text{if $N_k \geq T$},\\
0& \text{if $N_k < T$}.
\end{cases}

In the simulation you can find here, I considered a cellular automaton with the following properties: number of states $m=2$; moore neighborhood with radius $r=1$; lattice size $L_x \times L_y$; and 3 inputs for the transition table:

• Current state $x_{ij}^t$
• Neighborhood state $P_{1,T}$ with $T$ unspecified
• One external input $\xi$
\xi_{ij}=\begin{cases}
1& \text{if $i \geq L_x/2$},\\
0& \text{if $i < L_x/2$}.
\end{cases}
• Initial condition $x_{ij} = 0 \; \forall_{ij}$

For these conditions a deterministic simulation of these conditions yields only a few steady states: homogeneous 1 or 0, half the lattice 1 and the other 0, and oscillation between a combination of the previous.

One possibility would be to add noise to the cellular automaton in order to provide more interesting dynamics. There are two ways to add noise to a cellular automaton:

The most straightforward way is to perform the following procedure at each time step:

• Apply the deterministic dynamics to the whole lattice
• For each lattice site $ij$, invert the state $x_{ij}$ with probability $p$

This procedure only works of course for $m=2$. In the case of more states there is no obvious way to generalize the procedure and we need to use a proper monte carlo method to get the dynamics.

A second way is to implement a probabilistic cellular automaton. In this case the transition table is generalized to a markov matrix: each input is now mapped not to a specific state but rather to a set of probabilities for a transition to each state ($m$ probabilities). Naturally for each input these sum to one. In this case we have $m$ times more parameters than before.

## The link between thermodynamics and inference

In recent blog posts I talked a bit about how many aspects of maximum entropy were analogous to methods in statistical physics. In this short post, I’ll summarize the most interesting similarities. In bayesian inference, we are usually interested in the posterior distribution of some parameters $\theta$ given the data d. This posterior can be written as a boltzmann distribution: $$P(\theta|d)=\frac{P(\theta,d)}{P(d)}=\left.\frac{e^{-\beta H(\theta,d)}}{Z}\right|_{\beta=1}$$ with $H(\theta,d) = -\log P(\theta,d)/\beta$ and $Z=\int d\theta\;e^{-\beta H(\theta,d)}$. I’ll note that we are working with units such that $k_B=1$ and thus $\beta=1/T$.

The energy is just the expectation value of the hamiltonian H (note that the expectation is taken with respect to $P(\theta|d)$): $$E = \langle H \rangle = -\frac{\partial \log Z}{\partial \beta}$$

And the entropy is equal to $$S=-\int d\theta\;P(\theta|d)\log P(\theta|d)=\beta\langle H \rangle – \log Z$$

We can also define the free energy, which is $$F=E\, – \frac{S}{\beta}=-\frac{\log Z}{\beta}$$

A cool way to approximate Z if we can’t calculate it analytically (we usually can’t calculate it numerically for high dimensional problems because the integrals take a very long time to calculate) is to use laplace’s approximation: $$Z=\int d\theta\;e^{-\beta H(\theta,d)}\simeq\sqrt{\frac{2\pi}{\beta|H”(\theta^*)|}}e^{-\beta H(\theta^*)}$$ where $|H”(\theta^*)|$ is the determinant of the hessian of the hamiltonian (say that 3 times real fast) and $\theta^*$ is such that $H(\theta^*)=\min H(\theta)$ (minimum because of the minus sign). Needless to say this approximation works best for small temperature ($\beta\rightarrow\infty$) which might not be close to the correct value at $\beta=1$. $\theta^*$ is known as the maximum a posteriori (MAP) estimate. Expectation values can also be approximated in a similar way: $$\langle f(\theta) \rangle = \int d\theta \; f(\theta) P(\theta|d) \simeq\sqrt{\frac{2\pi}{\beta|H”(\theta^*)|}} f(\theta^*)P(\theta^*|d)$$

So the MAP estimate is defined as $\text{argmax}_{\theta} P(\theta|d)$. The result won’t change if we take the log of the posterior, which leads to a form similar to the entropy: \begin{align}\theta_{\text{MAP}}&=\text{argmax}_{\theta} (-\beta H – \log Z)\\&=\text{argmax}_{\theta} (-2\beta H + S)\end{align} Funny, huh? For infinite temperature ($\beta=0$) the parameters reflect total lack of knowledge: the entropy is maximized. As we lower the temperature, the energy term contributes more, reflecting the information provided by the data, until at temperature zero we would only care about the data contribution and ignore the entropy term.

(This is also the basic idea for the simulated annealing optimization algorithm, where in that case the objective function plays the role of the energy and the algorithm walks around phase space randomly, with jump size proportional to the temperature. The annealing schedule progressively lowers the temperature, restricting the random walk to regions of high objective function value, until it freezes at some point.)

Another cool connection is the fact that the heat capacity is given by $$C(\beta)=\beta^2\langle (\Delta H)^2 \rangle=\beta^2\langle (H-\langle H \rangle)^2 \rangle=\beta^2\frac{\partial^2 \log Z}{\partial \beta^2}$$

In the paper I looked at last time, the authors used this fact to estimate the entropy: they calculated $\langle (\Delta H)^2 \rangle$ by MCMC for various betas and used the relation $$S = \, \int_{1}^{\infty} d\beta\; \frac{1}{\beta} C(\beta)$$

## Review of ‘Searching for Collective Behavior in a Large Network of Sensory Neurons’

Last time I reviewed the principle of maximum entropy. Today I am looking at a paper which uses it to create a simplified probabilistic representation of neural dynamics. The idea is to measure the spike trains of each neuron individually (in this case there are around 100 neurons from a salamander retina being measured) and simultaneously. In this way, all correlations in the network are preserved, which allows the construction of a probability distribution describing some features of the network.

Naturally, a probability distribution describing the full network dynamics would need a model of the whole network dynamics, which is not what the authors are aiming at here. Instead, they wish to just capture the correct statistics of the network states. What are the network states? Imagine you bin time into small windows. In each window, each neuron will be spiking or not. Then, for each time point you will have a binary word with 100 bits, where each a 1 corresponds to a spike and a -1 to silence. This is a network state, which we will represent by $\boldsymbol{\sigma}$.

So, the goal is to get $P(\boldsymbol{\sigma})$. It would be more interesting to have something like $P(\boldsymbol{\sigma}_{t+1}|\boldsymbol{\sigma}_t)$ (subscript denoting time) but we don’t always get what we want, now do we? It is a much harder problem to get this conditional probability, so we’ll have to settle for the overall probability of each state. According to maximum entropy, this distribution will be given by $$P(\boldsymbol{\sigma})=\frac{1}{Z}\exp\left(-\sum_i \lambda_i f_i(\boldsymbol{\sigma})\right)$$ Continue reading “Review of ‘Searching for Collective Behavior in a Large Network of Sensory Neurons’”

## Maximum entropy: a primer and some recent applications

I’ll let Caticha summarize the principle of maximum entropy:

Among all possible probability distributions that agree with whatever we know select that particular distribution that reflects maximum ignorance about everything else. Since ignorance is measured by entropy, the method is mathematically implemented by selecting the distribution that maximizes entropy subject to the constraints imposed by the available information.

It appears to have been introduced by Jaynes in 57, and has seen a resurgence in the past decade with people taking bayesian inference more seriously. (As an aside, Jayne’s posthumously published book is well worth a read, in spite of some cringeworthy rants peppered throughout.) I won’t dwell too much on the philosophy as the two previously mentioned sources have already gone into great detail to justify the method.

Usually we consider constraints which are linear in the probabilities, namely we constrain the probability distribution to have specific expectation values. Consider that we know the expectation values of a certain set of functions $f^k$. Then, $p(x)$ should be such that $$\langle f^k \rangle = \int dx \; p(x) f^k(x)$$ for all k. Let’s omit the notation $(x)$ for simplicity. Then, we can use variational calculus to find p which minimizes the functional $$S[p]\; – \alpha \int dx\; p\; – \sum_k \lambda_k \langle f^k \rangle$$ The constraint with $\alpha$ is the normalization condition and $S$ is the shannon entropy.

The solution to this is $$p = \frac{1}{Z}\exp\left(-\sum_k\lambda_k f^k \right)$$ with $$Z=\int dx \; \exp \left(-\sum_k\lambda_k f^k \right)$$ the partition function (which is just the normalization constant). Now, we can find the remaining multipliers by solving the system of equations $$-\frac{\partial \log Z}{\partial \lambda_k} = \langle f^k \rangle$$ I’ll let you confirm that if we fix the mean and variance we get a gaussian distribution. Go on, I’ll wait.

## How to do inverse transformation sampling in scipy and numpy

Let’s say you have some data which follows a certain probability distribution. You can create a histogram and visualize the probability distribution, but now you want to sample from it. How do you go about doing this with python?

The short answer:

The long answer:

You do inverse transform sampling, which is just a method to rescale a uniform random variable to have the probability distribution we want. The idea is that the cumulative distribution function for the histogram you have maps the random variable’s space of possible values to the region [0,1]. If you invert it, you can sample uniform random numbers and transform them to your target distribution!

To implement this, we calculate the CDF for each bin in the histogram (red points above) and interpolate it using scipy’s interpolate functions. Then we just need to sample uniform random points and pass them through the inverse CDF! Here is how it looks:

## Gamma distribution approximation to the negative binomial distribution

In a recent data analysis project I was fitting a negative binomial distribution to some data when I realized that the gamma distribution was an equally good fit. And with equally good I mean the MLE fits were numerically indistinguishable. This intrigued me. In the internet I could find only a cryptic sentence on wikipedia saying the negative binomial is a discrete analog to the gamma and a paper talking about bounds on how closely the negative binomial approximates the gamma, but nobody really explains why this is the case. So here is a quick physicist’s derivation of the limit for large k.

## Negative binomial with continuous parameters in python

So scipy doesn’t support a negative binomial for a continuous r parameter. The expression for its pdf is $P(k)=\frac{\Gamma(k+r)}{k!\,\Gamma(r)} (1-p)^rp^k$. I coded a small class which computes the pdf and is also able to find MLE estimates for p and k given some data. It relies on the mpmath arbitrary precision library since the gamma function values can get quite large and overflow a double. It might be useful to someone so here’s the code below.