## 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.