Profile cover photo
Profile photo
Fabio Stefanini
196 followers
196 followers
About
Fabio's posts

Post has attachment

“Today, neuromorphic hardware is limited by memory capacity, which can be catastrophically low when these systems are designed to learn autonomously,” said Dr. Fusi. “Creating a better model of synaptic memory could help to solve this problem, speeding up the development of electronic devices that are both compact and energy efficient — and just as powerful as the human brain.”

If you want a nice circuit drawing exercise that may end up being very useful for a physical memory system (as opposed to simulated or numerical) here is a nice work from a colleague of mine in the Fusi lab, Marcus Benna. He shows a possible explanation to the reason why biological synapses are such good devices for storing patterns for crazy long times despite their non floating point precision (such precision is required to store a huge number of patterns as expressed in the well known Hopfield model with its unbounded synapses). The hypothesis is that the myriads of interacting molecular mechanisms that happen in biological synapses are influenced by the long-term potentiation (LTP) and depression (LTD) events imposed by the incoming stimuli and effectively they behave as part of the memory system as a whole. In this way, the actual synaptic efficacy at any point in time is not only the result of the previous LTP and LTD events but also of the "internal states" of the synapse stored as the state of these many interacting molecular mechanisms. By distributing the memory to the internal mechanisms the synaptic efficacy (what we usually call 'weight') can nicely live in its low-precision world without compromising the memory capacity of the system. A simple physical model is described and analytically described to explain why this happens. This is a kind of "holy grail" for memory systems with limited precision... It shouldn't be too hard to draw circuits behaving in this way or even better to to update our good old hydroneuron with some nice plasticity!


[ if you wonder what the hydroneuron is: https://www.youtube.com/watch?v=yNt9sYhasnI ]


Post has attachment
An interesting analysis of power consumption for spiking neuromorphic chips by Marti, Rigotti, Seok, Fusi.

Post has attachment
We are glad to announce our latest success, a Python-based modular framework for your nasty neuromorphic geekness ;)
PyNCS is in GitHub https://github.com/inincs/pyNCS  and is waiting for users, developers, hackers... We wish it will speed-up the exchange and integration of hardware and software modules between groups.

Post has attachment

Comments are welcome.
Fabio

Post has attachment
What can Neuromorphic Engineering do for these guys?

Post has attachment
A big plus for Schmidhuber's actor skills in the first few minutes :)
Wait while more posts are being loaded