The predictive coding model provides a biologically-plausible account of how organisms infer stimuli from noisy inputs. Introduced by Rao and Ballard in 1999. Friston extended the model in 2005 to also learn uncertainty associated with different features (e.g. attentional mechanism). Friston’s model can also be viewed as approximate Bayesian inference based on a minimization of free energy.
This paper provides a tutorial that aims to be broadly accessible to a technical audience.
- Conditions for a model to be biological plausible:
- Local computation - each neuron need only know about its inputs and outputs.
- Local plasticity - localized changes can be used to train the model.
We start with a problem: a single organism is trying to infer the diameter of a food item on the basis of observed light intensity from one noisy light receptor. There exists a non-linear function \(g\) relating average light intensity with size.
Interesting to note: once distributions aren’t standard (e.g. normal or otherwise well known) you can’t represent them compactly with summary statistics. It seems likely there has to be some sort of approximation going on in the brain.
Also interesting: calculating the bayesian normalization term (the denominator) seems difficult for neural systems (also, not super straightforward for computer systems, either).
Suggests that instead of find the whole posterior, we just find the most likely size of the food item given the sensor reading. This is claimed to be much more plausible to implement in neural circuits.
- Importantly, the value that maximizes the likelihood (\(\phi\)) does not depend on the denominator so we can not consider it. Take the natural log of the numerator \(p(u | \phi) * p(\phi)\) to get \(ln(p(u | \phi)) + ln(p(\phi))\).
Note posted on by Nick Jalbert