How is the biological error signal in predictive coding computed?

Question

I am quite inspired by the results obtained by PredNet, which implements a predictive coding model using artificial neural networks. They compute the prediction error as a simple subtraction, and then split the positive and negative into two different feature maps.

Within the predictive coding framework it is believed that biological brains also uses the prediction error to communicate between different areas. How is this error signal computed in biological brains?.

Update

I am especially interested in the mechanical aspects, such as how doe does the brain compare a prediction with the true signal?. Synthetic approaches generally just use a pixel-wise error, and I am wondering if the brain does something analogously.

Answer 1

I provided an answer to a similar question here that limitedly deals with the role of biological prediction errors.

Here's an excerpt of that answer:

...to answer this properly, we must first make it clear that there are potentially dozens, hundreds, or an arbitrarily high number of other "prediction error types" in use by the brain. Here are just a few major ways, hypothetically:

• Lots of different neurotransmitters (e.g. dopamine)
• The opening/closing of various ion channel species that regulate the membrane potential
• Synaptic vescicles/receptors
• Neuronal firing rates (as in bursting, a rapid succession of action potentials)
• Temporal coding (relative firing times to the firing of other neurons)
• And I can think of 10 other more-subtle and harder to explain possibilities, but that are just as important, off the top of my head

Keep in mind that each neuron also seems to have its own differentiated mechanisms for, both, interpreting and signaling prediction error. This complicates things further. For instance, one neurotransmitter may communicate prediction error to one particular neuron, but has no effect (or a different effect) on a different neuron. It may even be that neurotransmitter X must be present while temporal code Y happens for the event to be interpreted as a prediction error.

The study of biological mechanisms for prediction error is a very complicated thing that has no simple interpretation, as opposed to what you find in artificial approaches (presumably, like your PredNet example). While we have not yet uncovered how the brain computes or uses these mechanisms to encode and communicate prediction errors across neurons, what is obvious is that the brain has to be doing some kind of prediction error coding. However, if we try to oversimplify what the brain is doing, we are likely to not have a very intelligent model. The fragility of predictions, susceptibility to error from data complications, and limited nature of current artificial intelligence implementations lends to the idea that there is a lack of good ideas about how to implement and fully utilize prediction error.

The work I do is of a theoretical nature so your question is right up my alley. I have some unique ideas on how various biological prediction errors may work, but it requires a lot of background to understand. Unfortunately, I also have not published my ideas so they are definitely not peer reviewed. That makes me somewhat reluctant to mention my personal ideas as an answer.

It's possible that a kind of prediction error is used in all the listed mechanisms to fine tune the respective properties. Each mechanism likely has a unique role that is central to intelligence- from not only prediction errors encountered in the environment, but even to intrinsic behaviors and being able to predict the outcome of its own actions.

Answer 2

TL;DR: We don't know whether the brain really uses predictive coding or not. But neurally computing an error signal on a small scale is possible (see below).

Predictive coding is an hypothesis for a putative signal-processing mechanism used in vertebrate brains. As things stand presently (2017), mapping the hypothesis of predictive coding onto known neural structures and responses is a matter of active research (e.g. Choi et al. 2016, Zmarz & Keller 2016, Roth et al 2016).

On a small-scale neural level, an error signal could be partially computed through converging an excitatory "prediction" signal and an inhibitory "evidence" signal onto a single neuron. This is only a partial error signal, since it is rectified by the spiking threshold of the post-synaptic neuron.

The main issue is that the forward and backwards pathways need to follow very precise wiring patterns, and be quite tightly aligned, for the simple formulations of predictive coding to work. That doesn't mesh well with intuitions of what is biologically reasonable.

There are some approaches to map a predictive coding framework onto other mechanisms that may be more biologically plausible (e.g. Spratling 2008).