The second paper put out by Dr. Terence Sanger in the likelihood calculus paper series is Neuro-mechanical control using differential stochastic operators. Building on the probabalistic representation of systems through differential stochastic operators presented in the last paper (Controlling variability, which I review here) Dr. Sanger starts exploring how one could effect control over a system whose dynamics are described in terms of these operators. Here, he specifically looks at driving a population of neurons described by differential stochastic operators to generate the desired system dynamics. Neural control of a system requires that several phenomena outside the realm of classical control theory be addressed, including the effects of variability in control due to stochastic firing, large partially unlabeled cooperative controllers, bandlimited control due to finite neural resources, and variation in the number of available neurons.

The function of a neuron in a control system can be completely described in terms of 1) the probability of it spiking due to the current system state, , and 2) the effect of its response on the change in the state. Due to the inherent uncertainty in these systems, each individual neuron’s effect on the change in state is captured by a distribution, . And because the effect of each neuron is only a small part of a large dynamical system that includes the dynamics of the system being controlled and the effects of all the other neurons, these distributions tend to be very broad.

Rephrasing the above description, neurons are mapping a given state to a change in state . Instead of using two conditional densities to describe this mapping, and , we can rewrite this more compactly as

,

which can be read as the probability of a change in state given the current state is equal to the probability of that change in state occurring if the neuron spikes, , multiplied by the probability of that neuron spiking given the current state, , plus the probability of that state occurring if the neuron doesn’t spike, , multiplied by the probability of that neuron not spiking, .

**Differential stochastic operators**

We want to capture the mapping in such a way that if we have a description of a bunch of controllers (such as neurons) and the change in system state they effect individually, we can combine them in a straightforward way to get the overall change in state resulting from all the controllers operating in parallel. To do this we can use the linear operators developed in the previous paper, which allows us combine the effects of multiple components through simple summation to determine the overall change in system state. We’ll go over it again here, as I find reading through several different versions of an idea very helpful for solidifying understanding.

Let denote a class of linear operators that act on time-varying probability densities, , such that . Because these operators need to preserve the properties of valid probability density (specifically that and ), for a given operator where we require that:

- 1) for all ,
- 2) whenever ,

which respectively enforce the aforementioned constraints.

So, first thing’s first. Let’s read out . This says that our change in the probability density, , is found by taking our function that tells us what the change in density is for system state given our current state , which is , and weighting that by the probability of currently being in state , which is , then summing that all up, which is the integral.

Now the constraints. The first constraint reads out as the integral of the changes of the probability density at each point for a given state must be equal to 0. This means that the area of the probability density over the states after updating them is the same. So, assuming we start out with a valid density whose sum equals 1, we always have a density whose sum equals 1.

The second constraint reads out as the change in probability density for state given a current state must be greater than zero whenever . This means the only time that the change in the probability density can be negative is if there is a probability of being in that state; it enforces that all , because can’t be negative when is zero.

Dr. Sanger defines the linear operators that satisfy these two conditions to be “differential stochastic operators”. The discrete time versions are matrices, dubbed “difference stochastic operators”.

**Superposition of differential stochastic operators**

Differential stochastic operators can be derived in different ways, here we’ll go through the derivation from the ‘master’ equation defining , and from a stochastic differential equation. They each have their insights, so it’s helpful to work through both.

Derivation from master equation

The equation for written out above,

,

determines the probability flow on the state. By employing the Kramers-Moyal expansion we can capture this probability flow through a partial differential equation describing change in probability density. In other words, instead of capturing the time evolution of the system state with a probability density over possible changes in state, , we capture it through the changing in probability of each state, . The Kramers-Moyal expansion looks like:

,

where . Truncating this expansion at the first two terms we get the Fokker-Planck equation, where the first term describes the drift of the density, and the second term the diffusion. These two terms are sufficient for describing Gaussian conditional densities, which capture many physical phenomena. In the case where does not have a Gaussian distribution, higher-order terms from the Kramers-Moyal expansion will need to be included.

Now, imagine we have a simple system of two neurons, where the change in state is defined as . If these neurons have conditionally independent variability, i.e. , then we can sum the Kramers-Moyal expansion of each of these terms to describe the evolution of the overall system state:

,

as long as the neurons have conditionally independent variability. This means that they can’t be connected (directly or indirectly) such that a spike in one neuron causes a spike in the other. While this might seem a poor assumption for modeling networks of spiking neurons, in large populations with many input, the effects of any single input neuron tends to be small enough that the assumption holds approximately.

We can rewrite the previous equation now, taking advantage of linearity of and the Kramers-Moyal coefficients, to get

,

which means that by describing neurons with the differential stochastic operators, , we can determine the cumulative effect on the dynamics of the system through simple summation. Which we all remember from the last paper, but hey, good to review.

Now, in the event that we want to model the effect of a group of strongly interconnected neurons, we can instead consider the effect of the group as a function of the possible firing patterns (spike or no spike from each of the neurons). So where before was written in terms of the two cases and , it would now be written:

,

where each is a different spike pattern. This group on neurons and their effect on the system dynamics is then considered as a single unit, and the differential stochastic operator describing them can then be combined with the operator from another group of neurons, provided that the two groups exhibit conditionally independent variability.

If there are no independent groups in the network then it’s fully connected and this is not for you go home.

Derivation from a stochastic differential equation

For this derivation, imagine a set of controllers operating in parallel. Each controller has a stochastic differential equation that defines how a control signal input affects the system dynamics,

,

where and are arbitrary equations and are random functions of time, or noise terms. Let for be the controller equations, and be the unforced, or passive, dynamics of the system, which define how the system behaves without controller input. We can write the equation for the whole system as

,

where are constant or slowly varying control inputs. The reason we would choose to denote the unforced (passive) dynamics of the system as is because we can now define , and rewrite the above equation as

.

The corresponding Fokker-Planck equation for the evolution of the state probability density is

.

Look familiar? We can rewrite this as a superposition of linear operators

,

and there you go.

**Population model**

So, now we can apply this superposition of differential stochastic equations to describe the effect of a population of neurons on a given system. Dr. Sanger lists several ways that this model can go about being controlled; 1) modifying the tuning curves of the neurons, which specifies how they respond to stimulus; 2) modify the output functions that determines the effect that a neuron has on the dynamics of the system; and 3) modifying the firing threshold of the neurons.

I found the difference between 1 and 3 can be a little confusing, so let’s look at an example tuning curve in 2D space to try to make this clear. Imagine a neuron sensitive to -dimensional input signals, and that it’s tuning curve looks like this:

If we’re changing the tuning curve, then how this neuron responds to its input stimulus will change. For example, in this diagram we show changing the tuning curve of a neuron:

Here, the neuron no longer responds to the same type of input that it responded to previously. We have made a qualitative change to the type of stimulus this neuron responds to.

If we change the firing threshold, however, then what we’re changing is when the neuron starts responding to stimulus that it is sensitive to.

Here we show the neuron becoming more and more sensitive to a its stimulus, respond stronger sooner and sooner. So the type of signal that the neuron responds to isn’t changing, but rather when the neuron starts responding.

Alright, now that we’ve got that sorted out, let’s move on.

Tuning curves (1) and output functions (2) are both modifiable through learning, by changing the incoming and outgoing connection weights, respectively, but for controlling systems on the fly this is going to be too slow, i.e. slower than the speed at which the system moves. So what’s left is (3), modifying the firing threshold of the neurons. So the model then looks like:

where is projected in to a population of neurons, each with a stochastic differential operator that sum together to generate . In this diagram, is the firing threshold of neuron , and denotes the modulation of the firing rate of neuron as a function of the current system state. When the modulation is dependent on the system state we have a feedback, or closed-loop, control system. Dr. Sanger notes that in the case that is heavily dependent on , modulating the firing threshold is indistinguishable from modifying the tuning curve, meaning that we can get some pretty powerful control out of this.

**Conclusions**

‘The theory of differential stochastic operators links the dynamics of individual neurons to the dynamics of a full neuro-mechanical system. The control system is a set of reflex elements whose gain is modulated in order to produce a desired dynamics of the overall system.’

This paper presents a very different formulation of control than classical control theory. Here, the goal is to modulate the dynamics of the system to closely match a desired set of dynamics that achieve some task, rather than to minimize the deviation from some prespecified trajectory through state space. Dr. Sanger notes that this description of control matches well to behavioral descriptions of neural control system, where there are numerous subsystems and circuits that have reflexive responses to external stimuli which must be modulated to achieve desired behavior. The goal of control is to fully define the dynamics of the system’s reaction to the environment.

**Comments and thoughts**

What comes to mind first for me, in terms of using the modulation of reflex elements to effect a desired set of dynamics, is modeling the spinocerebellum. With a ton of projections directly to the spinal cord, and strong implications in locomotor and balance system, it seems like it would be a very good candidate for being modeled with this type of control. The idea being that the cerebellum is projecting modulatory values to the different spinal circuits (reflex networks and central pattern generators, for example) that specify how to respond to changes in the environment to maintain our balance or the rhythm of our walk. How we go about specifying exactly what those modulatory terms need to be is something that Dr. Sanger tackles in the last paper of this series, which I’ll be reviewing in the next couple of months. I’m looking forward to it.

On another note, in my lab we all work with the Neural Engineering Framework, in which populations of neurons are taken to represent vectors, and to perform transformations on these vectors to relay information an perform various functions. To this end, something that interests me about likelihood calculus is its application to populations of neurons representing vectors. Instead of finding by summing the effects of all of the neurons in a population, or defining it in terms of the population spiking patterns, we’re looking at it in terms of the different vectors this population can represent, and the effect of that vector on the system. So we can still have spiking neuron based models, but we can do all the likelihood calculus business one level removed, simplifying the calculation and reducing the number of differential stochastic operators needed.

There are a ton of things going on in this paper, and lots to think about. At several points I deviated from the notation used in this paper because I found it unclear, but aside from that I’ve really enjoyed reading through it and writing it up. Very interesting work.

Sanger TD (2010). Neuro-mechanical control using differential stochastic operators. Conference proceedings : … Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Conference, 2010, 4494-7 PMID: 21095779

Woooooooooosh.

That was the sound of a lot of that going over my head, but I can tell it was very clearly explained for people who know at least a bit about the topic!

Thanks for the comment! I’d be happy to try and further explain it if you have any questions.

This article relates to chapter 2 of my psychology textbook. Chapter 2 mainly goes over the basics of the nervous system and the endocrine system. Both my class and this article discuss how our nervous system effects us in different enviroments and how we control our neurons. I find very fascinating how our brain sends messages (neurons) to the spinal cord to determine our movements and actions. I like the fact that we change when our neuron starts responding to a stimulus, and how we respond faster to certain situations when we become sensitive to our stimulus.

I also find it very fascinating! Learning how to modulate the neural circuits in the spinal cord such that we can generate desired responses to the environment stimuli (for example, by modifying our reflexes) is something that I’m hopeful this kind of research can lead us towards better understanding.