Dan Goodman
@neural-reckoning
Computational neuroscientist at Imperial College. I like spikes and making science better (Neuromatch, Brian spiking neural network simulator, SNUFA annual workshop on spiking neurons). 🧪 📷
Modulation (top curve, red) also lets us reduce firing rates by orders of magnitude without hurting performance, unlike unmodulated networks (bottom curve, blue). Fewer spikes means less energy. Important both for real brains and neuromorphic devices!
In a neuromorphic or machine learning context, neuromodulation is very parameter efficient. This figure shows every variant of the modulated and unmodulated models in our paper, comparing parameter count versus accuracy. You can see that the frontier of modulated networks is much higher.
Our model allows for setting a number of neuromodulator types, that can have interactions between each other, and making neuromodulator release diffuse in space and time. This actually turns out to further improve performance!
And we can understand how it's doing it: it dials up the sensitivity when the noise level is low, and dials it down when the noise is high. This sort of dynamic gain control is the "listening in the dips" strategy that has been hypothesised to be used by humans. We didn't put this in, it learned it!
The performance enhancement is particularly large in a noisy background, in the range where human speech recognition is much better than state-of-the-art automatic speech recognition systems.
With this model, performance at a variety of tasks goes up (in the picture, an auditory task). A small network with modulation performs much better than a much larger network without modulation. We can also see here that modulation at medium timescales is best. Spatial scale not very important here.
Our model is simple at the core: we let a modulatory neuron modify the underlying parameters of another neuron (like threshold or time constant), and train end to end. We can make this more realistic in various ways (different types of neuromodulators, spatial release and diffusion) - see later.
The hardware version gets higher throughput (4x better) and energy efficiency (5x better) at comparable area.
We see that across a range of tasks we get state-of-the-art performance at a lower memory cost (M) compared to other models (look for the stars in the figure).
We think of this as a biologically inspired fast-slow structure where the SNNs are acting fast, and the memory buffer (a Legendre memory unit) acts as a slow working memory. We co-designed this to work as a neural model and in neuromorphic hardware.
Second joyful nature photo for today to celebrate election result in Hungary. #photography
The kids had fun making paths on the ground in the local wood. #photography
Not normally into bird #photography but quite pleased that I managed to get a shot of a woodpecker! Even managed to get a video of it pecking.
In this sort of case I often think of Kent Anderson's hilarious article on "102 things journal publishers do" (to add value). One of which is to have a fancy office. That's not needed for science.
Today may not have been the best day to go off piste on my run...
Some days you just wish you had your big fancy camera with you. #photography