Thomas Nowotny

Finally, our work was done in two sensillum types out of dozens, so we can't be sure that other ORN types may not have different F-I curves and adaptation; however, nonlinear F-I curves are ubiquitous and strong adaptation has been seen in many studies...

So, ephaptic interactions mainly seem to matter when one of the ORNs in the sensillum is adapted due to longer odour exposure, pointing to a more significant role of ephaptic interactions for novelty detection, as previously described.
All our observations are backed up by a computational model.

To our surprise, the non-linearity of frequency-input relations (F-I curve) of spiking olfactory receptor neurons (ORNs) means that the ephaptic inhibition is all but invisible if ORNs are highly activated at odour onset. Adaptation additionally masks the effects.

Happy to share our preprint on ephaptic interactions between olfactory receptor neurons in Drosophila by Lydia Ellison.
doi.org/10.64898/202...
We set out to see how near-instantaneous electrical interactions are good for processing tiny odour onset delays, only to find that they weren't.

Happy to share our preprint with @neworderofjamie.bsky.social , @danakarca.bsky.social and @drtnowotny.bsky.social !

We’ve been working on a neuron position learning algorithm by coupling space and time! See manuscript below 👇

Excited to see the paper fully published. It's an important milestone for training SNNs with exact gradients, replacing our earlier tricks of a "delay line augmentation" to capture temporal relationships. Delays can now be learnt alongside weights naturally. Amazing work @mbalazs98.bsky.social !

If you're interested in doing a PhD at Sussex, lots of exciting PhD project ideas at www.sussex.ac.uk/research/cen... including several Spiking Neural Network projects with myself and @drtnowotny.bsky.social . Funding available via www.sussex.ac.uk/study/fees-f... or www.sussex.ac.uk/study/fees-f...

Excellent paper and great new project. Very interesting how the race of technology between GPU and FPGA shapes up.

Strong results on Speech commands with GLE presented by Paul Haider at #cns2025florence today . @cnsorg.bsky.social

And we are off - excited that #cns2025florence is under way with record attendance - first keynote starts in 5 minutes.

Towards a Better Future: How Sussex Students are Using AI to Change the World! YouTube video by University of Sussex

Very excited by our new Sussex AI promotional video www.youtube.com/watch?v=D5dS...

Posting this after some recent conversations with potential international applicants - still time to apply to our Masters courses and International PhD Academy for 2025 entry - join the diverse and vibrant Neuroscience community on our beautiful campus next to Brighton

... and you can see the great lineup of tutorials and workshops here: ocns.memberclicks.net/cns-2025-mee...
Of course, tutorial and workshop registration can also be added later if you have already registered for the main meeting. @cnsorg.bsky.social

The deadline is tomorrow - last push! @cnsorg.bsky.social

Only 5 days to go to the (extended) deadline. Make them count. A great lineup of keynotes, tutorials and workshops is secured - add your research as an oral or poster by submitting an abstract.

@drtnowotny.bsky.social and I are again participating in Google Summer of Code under the @incforg.bsky.social. We have 3 paid projects involving GeNN for contributors with a range of skills and experience levels. If you're interested, please get in touch via the forums linked from the thread:

Call for Abstracts: 34th Annual Computational Neuroscience Meeting, CNS*2025 🌟

📍 Join us in Florence 🇮🇹 July 5-9, 2025! 🧠✨

🗓️ Abstract Deadline: March 11, 2025
🔗 Submit here: www.cnsorg.org/cns-2025-abs...

📚 Workshops: www.cnsorg.org/cns-2025-cal...
🎓 Tutorials: www.cnsorg.org/cns-2025-cal...

This is awesome work from Balazs! Not only does our Eventprop-based method supports multiple spikes per neuron and recurrent connectivity but uses less than half the memory of the current state-of-the-art delay-learning method and is up to 26x faster.

Wikipedia is one of the last major bastions of verified information. Which, of course, is why the oligarchs want to destroy it.

You can donate to them here.

donate.wikimedia.org/w/index.php?...

Figure displaying the results of computational benchmarking of the algorithm. A) GPU memory use is plotted versus the number of timesteps in the simulation. We compared GeNN versus SpyX and for 256 hidden neurons and 1024 hidden neurons. The graph shows a linear increase of memory requirements for SpyX but essentially flat lines for GeNN. B) A similar graph for the training time as a function of timesteps. All curves increase but the increase for SpyX is much steeper than for GeNN. C) A bar graph comparing the training time for different elements of the network as a function of timesteps. The most time is spent on the synapses, followed by neuron simulation and other bits. The compile time is negligible. All times increase roughly linearly with number of time steps and somewhat sub-linearly with the number of hidden neurons.

We also classified the Spiking Speech Commands (SSC) with good success. Finally, the GeNN implementation of Eventprop has very beneficial computational scaling properties compared to BPTT in Spyx (github.com/kmheckel/spyx). All details at doi.org/10.1088/2634.... @sussexai.bsky.social 5/5

Bar graph showing the performance of our algorithm in classifying the Spiking Heidelberg Digits data set with four different variations: plain, with delay line input, with delay and shift, and with delay, shift and blend. The performance increases in this order. Bars are shown for all four conditions for three hidden layer sizes. The performance is best for the largest hidden layer but the differences are not very large.

We extended Eventprop to a wider class of loss functions and found that, with a ‘loss-shaping’ term, we could achieve fast and reliable learning. Combining this with 3 data augmentations, we obtained a SOTA SHD classification accuracy of 93.5±0.7% (n=8) on the test set after rigorous validation. 4/5

This initially failed due to average cross-entropy loss creating unhelpful gradients in the hidden layer and the fact that spike creation and deletion is not “visible” in the exact gradients calculated by Eventprop. 3/5

We implemented Eventprop (Wunderlich & Pehle, 2021) in our GeNN (genn-team.github.io) simulator and attempted to classify the Spiking Heidelberg Digits (SHD, zenkelab.org/resources/sp...) in a 3-layer network. 2/5

Just out in J Neuromorph Comput & Eng: Loss shaping enhances exact gradient learning with Eventprop in Spiking Neural Networks doi.org/10.1088/2634.... With @neworderofjamie.bsky.social.
We report how Eventprop scales to harder learning tasks.
TL;DR: It works great but not without extra effort. 🧪🧵

"A study of federally funded research projects in the United States estimated that principal investigators spend on average about 45% of their time on administrative activities related to applying for and managing projects rather than conducting active research"

www.pnas.org/doi/10.1073/...

We are proposing PhD project ideas along these lines:

More than 60 German-speaking universities and research institutes have just jointly announced they will cease activities on X because “the current orientation of the platform is not compatible with their core values” incl. scientific integrity, transparency and democratic discourse” 🧪

Hi Manu,
I am a Professor of Informatics and work in Computational Neuroscience and neuromorphic computing. I would like to contribute to the Science feed. My staff page is profiles.sussex.ac.uk/p206151-thom... and Google Scholar scholar.google.com/citations?us...

New paper on spike sorting www.sciencedirect.com/science/arti...
with Lydia Ellison, Georg Raiser, Alicia Garrido Peña and George Kemenes. TL;DR: SSSort software now handles overlapping spikes in addition to the extreme spike shape changes it was already good at. github.com/grg2rsr/SSSort