The last years have seen many exciting new developments to train spiking neural networks to perform complex information processing. This online workshop brings together researchers in the field to present their work and discuss ways of translating these findings into a better understanding of neural circuits. Topics include artificial and biologically plausible learning algorithms and the dissection of trained spiking circuits toward understanding neural processing. We have a manageable number of talks with ample time for discussions.
The workshop is being organised by Dan Goodman and Friedemann Zenke.
Registration
Registration is now closed. You can see a replay of the talks on Crowdcast or the talks and one of the recorded discussions on YouTube. Note that Claudia Clopath’s talk and the day 2 discussion were not recorded.
Agenda
Talks will be 45m (30m + 15m questions/discussion). Hover over the talk titles to see abstracts (if one is available).
August 31st
| Time (CET) | Session | Local date/time |
|---|---|---|
| 14:00 | Welcome by organizers | Mon Aug 31 2020 12:00:00 GMT+0000 (Coordinated Universal Time) |
| 14:10 | Sander Bohte (CWI) Effective and Efficient Computation with Multiple-timescale Spiking Recurrent Neural Networks |
Mon Aug 31 2020 12:10:00 GMT+0000 (Coordinated Universal Time) |
| 14:55 | Iulia M. Comsa (Google Research) On temporal coding in spiking neural networks with alpha synaptic function |
Mon Aug 31 2020 12:55:00 GMT+0000 (Coordinated Universal Time) |
| 15:40 | Break (30mins) | Mon Aug 31 2020 13:40:00 GMT+0000 (Coordinated Universal Time) |
| 16:10 | Franz Scherr (TUG) E-prop: A biologically inspired paradigm for learning in recurrent networks of spiking neurons |
Mon Aug 31 2020 14:10:00 GMT+0000 (Coordinated Universal Time) |
| 16:55 | Emre Neftci (UC Irvine) Synthesizing Machine Intelligence in Neuromorphic Computers with Differentiable Programming |
Mon Aug 31 2020 14:55:00 GMT+0000 (Coordinated Universal Time) |
| 17:40 | Break (30mins) | Mon Aug 31 2020 15:40:00 GMT+0000 (Coordinated Universal Time) |
| 18:10 | Discussion (can continue as long as needed) Current technical constraints and bottlenecks. We can train spiking neural networks. What now? |
Mon Aug 31 2020 16:10:00 GMT+0000 (Coordinated Universal Time) |
September 1st
| Time (CET) | Session | Local date/time |
|---|---|---|
| 14:00 | Welcome by organizers | Tue Sep 01 2020 12:00:00 GMT+0000 (Coordinated Universal Time) |
| 14:10 | Timothee Masquelier (CNRS Toulouse) Back-propagation in spiking neural networks |
Tue Sep 01 2020 12:10:00 GMT+0000 (Coordinated Universal Time) |
| 14:55 | Claudia Clopath (Imperial College) Training spiking neurons with FORCE to uncover hippocampal function |
Tue Sep 01 2020 12:55:00 GMT+0000 (Coordinated Universal Time) |
| 15:40 | Break (30mins) | Tue Sep 01 2020 13:40:00 GMT+0000 (Coordinated Universal Time) |
| 16:10 | Richard Naud (U Ottawa) Burst-dependent synaptic plasticity can coordinate learning in hierarchical circuits |
Tue Sep 01 2020 14:10:00 GMT+0000 (Coordinated Universal Time) |
| 16:55 | Julian Goeltz (Uni Bern) Fast and deep neuromorphic learning with time-to-first-spike coding |
Tue Sep 01 2020 14:55:00 GMT+0000 (Coordinated Universal Time) |
| 17:40 | Break (30mins) | Tue Sep 01 2020 15:40:00 GMT+0000 (Coordinated Universal Time) |
| 18:10 | Discussion (can continue as long as needed) Why spiking? |
Tue Sep 01 2020 16:10:00 GMT+0000 (Coordinated Universal Time) |
Discussion topics
Monday, August 31
We will discuss two topics:
Current technical constraints and bottlenecks.
How large do we need SNNs to be and how do we train them to do our research? What are the next steps the field should take to swiftly move forward?
We can train spiking neural networks. What now?
How do we leverage this methodological advance to gain a better understanding of how the brain processes information? What constitutes a conceptual advance? And how do we compare trained spiking neural networks to biology?
Tuesday, September 1
Why spiking?
Neurons communicate via precisely timed, discrete pulses rather than by analogue signals. Why? Is there a computational advantage to this mode of communication, or is just to save energy? With the recent advances in our ability to train spiking neural networks discussed in this workshop, can we throw new light on this age old discussion, and outline a programme to resolve it?