# Biological Plausibility of FORCE training

In "Supervised Learning in Spiking Neural Networks with FORCE Training" by Wilten Nicola and Claudia Clopath. The authors create a learning rule for learning non-linear dynamics from populations of spiking neurons. The learning rule only seems to depend on:

1. an error signal
2. the firing rates of each neurons
3. an approximate inverse correlation matrix calculated from the firing rates of each neuron

Is it possible to calculate the inverse correlation matrix of using another population of neurons or is there some biological mechanism that could be used to explain this? Additionally, is this calculation sensitive to delay? Often, forgetting to take delay into account is the undoing of many learning rules and cognitive architectures.

• Not biologically plausible – honi May 29 '17 at 15:32
• @honi you're right yet again! – Seanny123 May 30 '17 at 12:47
• lol, sorry for not giving the complete answer. i didn't remember why, i just remembered thinking about that question in depth when i read the paper several years ago. – honi Jun 2 '17 at 1:35

From the paper:

Although FORCE trained networks have dynamics that are starting to resemble those of populations of neurons, at present all top-down procedures used to construct any functional spiking neural network need further work to become biologically plausible learning rules [Sussillo and Abbott, 2009, Boerlin et al., 2013, Eliasmith et al., 2012]. For example, FORCE trained networks require non-local information in the form of the correlation matrix $P(t)$. However, we should not dismiss the final weight matrices generated by these techniques as biologically implausible simply because the techniques are themselves biologically implausible. More work should be done in implementing either FORCE, NEF, or spike-based coding networks using a biologically plausible learning mechanism based on synaptic plasticity or homeostasis [Bi and Poo, 1998, Pfister and Gerstner, 2006, Clopath et al., 2010, Graupner and Brunel, 2012, Babadi and Abbott, 2016, Vogels et al., 2011]. This has been resolved for spike-based coding networks and linear dynamical systems for example [Bourdoukan and Deneve, 2015]

Basically, calculating the correlation matrix $P(t)$ requires every neuron to know what every other neuron is doing, so the FORCE algorithm isn't biologically plausible. I'm not sure what alternative there would be calculating the correlation matrix. Maybe there's some mathematical way to approximate it gradually over time?