search

ctn-waterloo / modelling_ideas

Ideas for models that could be made with Nengo if anyone has time
9 stars 1 forks source link

Noisy inputs with long time constants may make LIF neurons act *less* like low-pass filters #6

Open tcstewar opened 9 years ago

tcstewar commented 9 years ago

Figure 3 in this paper http://galton.uchicago.edu/~nbrunel/pdfs/brunel01PRL.pdf shows a very weird result.

It's a plot of the firing rate of an LIF neuron with noisy inputs that are filtered by an exponential PSC (i.e. exactly what we do in nengo, only with a gaussian approximation for the input spikes). If the PSC has tau=0 (delta function), then there's a low-pass filter effect (which I believe is due to the membrane time constant). If we increase tau, the low-pass filter effect goes away. This very much surprises me, and makes me suspect that I'm interpreting something wrong.

However, it might be worth checking whether we get such effects. It might be more noticeable with lower firing rates (so the membrane time constant starts having time to have an effect) or with injected noise.

tcstewar commented 9 years ago

Ah, this is less bizarre than I thought. After chatting with Nicholas Brunel a bit, it turns out that the step response in this case was injected with no PSC, even though the random spiking input had a PSC. So that graph is for a neuron with random spiking input where someone directly injects a step function current.

But still, the idea here is that if we have synapses with two very different time constants, it can still be the case that the presence of the longer time constant (but poisson spiky) input can improve the system's response to the shorter time constant input.