🚨 New paper alert!
Have you ever suspected that spikes, Dale's law, and E/I balance might be more than just biological constraints, but rather fundamental to how brains compute? Check out my latest work with Christian Machens @Neuro_CF: https://t.co/5lZ8TlDmIE
🧵 (1/5)
We are excited to share our work on dynamical constraints on neural population activity, published as a cover article in @NatureNeuro. It was led by Emily Oby, @AlanDegenhart, @ErinnGrigsby, with @aaronbatista and team.
https://t.co/Oi2wXl0MIz
(1/n)
Registration & Abstract Submission is open for the 2025 "Engram and Ensembles in Learning and Memory" meeting, at @tcddublin in May.
Abstract deadline February 14.
#EngramsandEnsembles2025
Preliminary Agenda at event website:
https://t.co/i9RDN3u0Rp
Missed #SNUFA24 spiking neural network workshop last week? No worries, we uploaded all 11 talks and the flash talks to Youtube. Check it out, along with talks from every year since we started in 2020 📺 71 videos, 36k views and 3.4k hours watched so far.
https://t.co/Y5FLOJ0Q2M
“Where was I again?”
Our study published today https://t.co/SssPQUDv7l reveals brain cells can form a coordinate system for our behaviours. Instead of locating where we are in the world, this coordinate system tells us “where we are” in a sequence of behaviours: 🧵below:
#SNUFA24 starting at 14:00 CET (in just over 2 hours). Say no to doomscrolling, and yes to a free online workshop on spiking neural networks.
https://t.co/9aOMHNCLkU
The #SNUFA24 final program is out and the event is next Tue-Wed! If you love spiking neural networks, click on the link below to check it out and register (free).
https://t.co/U7TGfWN815
🧵on Japan's underrated contributions to neural nets. Shun-ichi Amari @UTokyo_News_en@riken_en is another one of my heroes. His 1972 paper on associative memory models modeled Hebbian plasticity using an outer product weight matrix.
I am excited to announce our workshop at the upcoming Bernstein Conference 2024 @BernsteinNeuro
How much biological inspiration makes a system "neuromorphic"?
What can we learn from Neuroscience breakthroughs and Machine Learning success?
link: https://t.co/2xU3tr5nMo
1/n
Back in 2022, @roxana_zeraati & I organized a Cosyne workshop on neural timescales, and after working on it for the last 2 years together, it's now a review paper!
https://t.co/LgVdULAQ2U
w/ @SelfOrgAnna & @jakhmack
(2nd blogpost to turn into a real review paper this year lol)
#Neuromorphic#computing just got more accessible! Our work on a Neuromorphic Intermediate Representation (NIR) is out in @SpringerNature Communications. We demonstrate interoperability with 11 platforms. And more to come!
https://t.co/dEnqKMdL2C
A thread 🧵 1/5
SPIKING NEURAL NETWORKS!
If you love them, join us at SNUFA24. Free, online workshop, Nov 5-6 (2-6pm CET). Usually ~700 participants.
Invited speakers: Chiara Bartolozzi, David Kappel, Anna Levina, Christian Machens
Posters + 8 contributed talks selected by participant vote.
How can we train biophysical neuron models on data or tasks? We built Jaxley, a differentiable, GPU-based biophysics simulator, which makes this possible even when models have thousands of parameters! Led by @deismic_, collab with @CellTypist @ppjgoncalves https://t.co/iMVKnw9YHz
Proud to share our paper in @Nature that Andrei and I co-first authored with @TFlogel@SWC_Neuro
We recorded ~15000 neurons across the brain while mice made perceptual decisions to reveal how sensory evidence controls actions through global neural dynamics https://t.co/LABeAGSh8N
We're hiring! Come build models of how the brain learns and simulates a world model. We have several openings at PhD and postdoc levels, including a collab with @georg98keller lab on designing regulatory elements to target distinct neuronal cell types.
https://t.co/DakerWs5Sz
it's fellowship szn!
I spent *a lot* of time applying for postdocs + independent positions last year, and I want to share my notes.
In this doc, I list deadlines, pay, and other 🔑 info on bio/theory positions I considered.
DM me for advice anytime.
https://t.co/CbUpOHXyQR
My #AI4Neuro magnum opus:
Discovery of spiking network model parameters constrained by neural recordings, using simulation-based inference & generative “AI”.
(aka the answer to “how the f did you end up in Tübingen?”)
Here's what we have in store:
https://t.co/UpnjVaLp8G
Sex or survival—what’s more important? Excited to share our @Nature paper on how flies resolve this conflict.
We found a dopamine-based filter that reduces threat perception, helping flies focus on courtship when close to mating.
https://t.co/ZZ0UoKTqgn
AI hardware systems such as @nvidia's GPUs are a true marvel of technology. Extremely complex structure that also harnesses enormous amounts of energy. That is because modern AI algorithms involve multiple functions, and GPUs implement those over multiple physical components. This requires careful orchestration & heavy communication among the components, through wires spanning entire chips and even multiple chips.
In new work with @ETH_en published in @NatureComms, we demonstrate that six key functions of modern AI algorithms can in fact be included in a single nanometre-scale device.
Namely, Chris Weilenmann and team, under the guidance of Dr Alexandros Emboras & Prof. Mathieu Luisier, shows an individual memristor that comprises not only the ability to store a memory of the network structure (i.e. a synaptic connection weight) and to transmit information between neurons (i.e. a weighting operation), but also working memory (i.e. recurrency), learning (i.e. synaptic plasticity), selective context retention/forgetting (i.e. short-term plasticity), and meta-learning (i.e. learning how to learn and forget). These are arguably the key functions that have brought modern AI models such as LLMs to their impressive performance.
To manage this result, we previously took inspiration from the brain to devise a new algorithm (the STPN - Short-Term Plasticity Neuron) that is suitable for such #neuromorphic hardware implementations, and reaches high performance in complex tasks. We started this line of work on STP at @IBMResearch, with @abuseb Abu Sebastian & Evangelos Eleftheriou (IJCNN, IEEE Nano, arXiv etc 2017-2020), it continued at @Huawei with @hector_grhv (ICML 2022), and culminated in this latest paper where we finally demonstrate a device that realizes the model physically. Moreover, we use measurements from such devices, in a neural network that plays an Atari video game not only better than a human, but also with a 100x reduction in energy consumption compared to a GPU.
This approach does not only improve AI in comparison to GPUs, but also compared with other in-memory computing (#IMC) hardware, and also with other neural network algorithms. Often, IMC for AI implies memory devices with "just" the two functions of weight storage and input-weight-multiplication, which is already a significant improvement compared to the (von Neumann) architecture of GPUs, which separates memory from computation. Our work further pushes the envelope of IMC, by including more of the key computational operations within the memory. As a result, our model (STPN) improves the efficiency of memristive hardware (see ICML 2022). Moreover, our algorithm is not only more efficient but also performs better than other models such as LSTM, even on GPU, as we showed in our previous work.
Sincerely thank you, Chris, Alex, Mathieu, and team for realizing this vision and for involving me.