Cian O'Donnell
LightNews LightNews
banner
cianodonnell.bsky.social
Cian O'Donnell
@cianodonnell.bsky.social
2.7K followers 1.8K following 780 posts
Computational neuroscientist.
Senior Lecturer at Ulster University in the Great City of Derry, Northern Ireland.
"not articulate enough"
https://odonnellgroup.github.io
Posts Replies Media Videos
cianodonnell.bsky.social
odd framing in this (otherwise v good) BBC article on Lough Neagh pollution. Early on quotes Mellor listing 6 sources with agriculture named 3rd

Later describes report from ABFI (Mellor's employer) that says agriculture responsible for 60% of nutrient load in the lake

www.bbc.co.uk/news/article...
Screenshot of text: "We have this massive surplus of nutrients that means it's an enriched broth," is how Dr Adam Mellor from the Agri-Food and Biosciences Institute (AFBI) describes Lough Neagh. Those nutrients are primarily phosphorus - one of the building blocks of life - and they come from many sources. "We have waste water as one of the significant contributors and we have land use and agriculture probably as the other large contributor. "But we also have individual contributions through septic tanks and what we individually put into the wastewater system." Screenshot of text: "An exercise carried out by AFBI determined that agriculture was responsible for about 60% of the nutrient load in the lough, wastewater approximately a quarter and the rest from septic tanks, industry and households."
October 22, 2025 at 7:23 AM
cianodonnell.bsky.social
🔥 Julia results on ODE solving benchmark tasks...

For non-stiff problems, Matlab is 3-10x faster than SciPy (Python), but Julia is 100x faster than Matlab

For stiff problems, Matlab and SciPy perform similarly to each other, but Julia is 100-1000x faster

From docs.sciml.ai/SciMLBenchma...
Plot of runtime vs error tolerance for a Lotka-Volterra benchmark non-stiff ODE solving problem for a range of algorithms from SciPy, Matlab and Julia (and others). Generally Matlab 3-10x faster than Scipy, but Julia is around 100x faster than Matlab. Plot of runtime vs error tolerance for a ROBER benchmark stiff ODE solving problem for a range of algorithms from SciPy, Matlab and Julia (and others). Generally Matlab and Scipy are similar, but Julia is around 100-1000x faster than both.
October 19, 2025 at 6:58 PM
cianodonnell.bsky.social
The joint FF+FG vote continues to decline. It's only a matter of time
Graph showing combed Fine Gael and Fianna Fáil vote since 99, declining from 70% to ~42%
October 18, 2025 at 12:04 PM
cianodonnell.bsky.social
drosophila gene names never fail to crack me up
Screenshot from Neuroscience Bulletin journal webpage with text: Sandman is a Sleep Switch in Drosophila Research Highlight | Published: 13 September 2016 Volume 32, pages 503-504, (2016) Cite this article
October 6, 2025 at 9:13 AM
cianodonnell.bsky.social
had a nice group discussion on "Is NeuroAI the next great thing? Or just a branding trick?"

Some very interesting arguments, summarised in slides below
Slides: "WHAT ARE THE POSSIBLE NEUROAI ARGUMENTS FOR FUTURE NEUROSCIENCE/NEUROSCIENTISTS? • Neural-network models of the brain as a tool for doing computational neuroscience (perturbations of brain components, parameters etc) • SciML methods to fit our biologically-faithful model to data/task • Ideas and methods developed in AI, might give us new ideas for how the brain works. • Better data analysis tools. • Can AI help with linking bottom-up and top-down approaches to neuroscience. • Get a well paid job." Slide with text: "WHAT ARE THE POSSIBLE NEUROAI ARGUMENTS FOR FUTURE AI? Future AI could be improved by neuroscience inspired ideas: • energy efficiency • resources • compactness • resilience/robustness • learning from less data • continual learning/lifelong learning • uncertainty • Brain-computer interfaces • Run natively on neuromorphic hardware"
October 2, 2025 at 1:36 PM
cianodonnell.bsky.social
in my 2006 PhD interview in Edinburgh I showed a slide with the "rotating snakes" illusion and said I want to understand how it works in the brain... thought I was being clever

Could never I have dreamed that 19 years later people would be figuring out this kind of thing in such mechanistic glory
Rotating snakes optical illusion, several yellow, blue and black discs appear to rotated even though they are static
September 20, 2025 at 2:06 PM
cianodonnell.bsky.social
Apple really need to do more market research in Ireland
Screenshot of iPhone Fitness app notification: "Check your ring. Here's a look at today's progress"
September 20, 2025 at 10:16 AM
cianodonnell.bsky.social
updated version of our paper on bayesian modelling for whole-brain cell count data: elifesciences.org/reviewed-pre....

People spend 1-2 years collecting these kinds of gene-expression/anatomy/IEG data... what's another 1-2 months learning + applying Bayes to get more stats bang for your buck 😍
Figure 4B from Dimmock et al showing bar chart of mean cell count change between two groups of animals for t-test vs a Bayesian model. Bayesian model consistently shows smaller uncertainty intervals and more statistically significant group changes.
September 19, 2025 at 6:34 PM
cianodonnell.bsky.social
anyone come across this statistical issue before?

I have simulated some time series data and am computing crosscorrelograms between two variables. "one year" of smith data recovers ground truth crosscorrelogram (right) but "one month" of sim data (left) gives artifactual negative correlations 🤷‍♂️
two cross-correlogram plots, left shows curves with positive peak at zero time lag, but then negative correlation side lobes. Right plot is similar, but for 'long time series' data, only shows non-negative correlations across the range of time lags.
September 8, 2025 at 12:29 PM
cianodonnell.bsky.social
incredible Ada Lovelace quote highlighted in a talk by Steve Furber. She spells out the dream of computational neuroscience, 2 centuries ago. The sheer ambition 🤩
Steve Furber powerpoint slide showing picture of Ada Lovelace and a quote: "I have my hopes, and very distinct ones too, of one day getting cerebral phenomena such that I can put them into mathematical equations--in short, a law or laws for the mutual actions of the molecules of brain. .... I hope to bequeath to the generations a calculus of the nervous system."
August 29, 2025 at 9:21 AM
cianodonnell.bsky.social
gave a short lecture this morning on principles of computational modelling, always try to stress the point made by @romainbrette.bsky.social that adding details to a model does not automatically make it more realistic.

The wooden airplane model has more 'details' but only the paper model can fly
Slide with title 'Details vs realism' showing a picture of a wooden model airplane on left and a paper airplane on right.
August 25, 2025 at 10:20 AM
cianodonnell.bsky.social
[side-stepping the gendered language] this passage from Conrad's Heart of Darkness reminds me of the private joys of doing science that I have realised are what keep me wanting to do research. Hidden from and quite different to the collective public process of science
Excerpt from Heart of Darkness "...don't like work. I had rather laze about and think of all the fine things that can be done. I don't like work—no man does-but Ilike what is in the work,— the chance to find yourself. Your own reality— for yourself, not for others-what no other man can ever know. They can only see the mere show, and never can tell what it really means."
August 19, 2025 at 9:53 PM
cianodonnell.bsky.social
I surprised myself at how much I enjoyed writing this glossary. Feel like all those years of social media posting have improved my ability to say technical things in a limited amount of space
Table saying: Box 1: Glossary Likelihood: The probability of the observed data, under a given statistical model and choice of parameters. Mixture model: A complicated probability distribution that is composed of two or more simpler probability distributions. Partial pooling: A hierarchical modeling method that links estimates for similar-but-not-identical parameters, for example mean cell counts from multiple animals within the same experimental group. Prior: Data-independent probability distributions describing plausible values for the statistical model's parameters. These are typically set 'prior' to observing the data, based on the analyst's domain knowledge. Posterior: The final probability distribution over model parameters, inferred from both the data-independent prior and data-dependent likelihood. Regularisation: A family of methods for reducing model overfitting to noise in limited data, by biasing parameter estimates towards sensible values.
July 21, 2025 at 3:35 PM
cianodonnell.bsky.social
Narin strand really is a special place
Narin beach, Donegal
July 12, 2025 at 2:23 PM
cianodonnell.bsky.social
It was hard to know how to feel when Diogo Jota died last week. I was sad because you feel like you kind of feel connected to these celebrities from watching them on tv every week, but surely it's nothing compared to the daily horror in Gaza
Guardian newspaper headline saying "Israeli strikes kill at least 38 in Gaza as ceasefire talks reach critical point"
July 6, 2025 at 7:59 PM
cianodonnell.bsky.social
my personal annoyance is that people assume:

computation + algorithm levels = systems neuroscience

implementation level = biophysics/cellular neuroscience

when there are many known algorithms implemented at the cellular/molecular level... eg ask anyone who works on retina
2019 tweet from Demis Hassabis: "no, i've been saying that for a very long time, and it is an obvious analogy (which is why we both ended up talking about it). It's why I always say we are interested in systems neuroscience (or the top two levels of Marr) at DeepMind not the implementation details of the brain"
June 26, 2025 at 11:35 AM
cianodonnell.bsky.social
"the hippocampus is essentially a "solved problem" in terms of the general framework for how its biological properties enable its well-established role in memory"
- Randy O'Reilly, 2008

@tomstafford.mastodon.online.ap.brid.gy's summary of a comp-neuro debate: eprints.whiterose.ac.uk/id/eprint/14...
Secondly, philosophy of science informs debates that we have within a field. Among modellers substantial disagreements exist concerning the correct approach to modelling. As evidence of this assertion, let me pick two discussants from a debate that occurred on the comp-neuro mailing list in 2008. James Bower expressed the opinion that modellers should perhaps "give up on cerebral cortex for several hundred years and all study tritonia instead" (Bower, n.d.). Although he was making this suggestion to illustrate a point, it does resonate with his apparent preference for low-level 'computational neuroscience' modelling. In ontrast, in the same debate, Randall O'Reilly wrote that "The hippocampus is essentially a "solved problem" in terms of the general framework for how its biological properties enable its well-established role in memory" (O'Reilly, n.d.). Not only should we, contra Bower, con nue to study cerebral cor- tex, but we have in fact essentially solved a major part of it! Bower initiated the discussion on the email list to illustrate to a group of graduate students that many fundamental issues within computational neuroscience are not agreed upon. This it illustrated admirably and the reader is encouraged to review the discussion to enjoy wide-ranging consideration of levels of modelling, ways of assessing the value of a model, the value of modelling in general. All of these issues are the business of philosophy of science and a modeller-scientist cannot avoid having a position of them, albeit if only implic-itly.
June 24, 2025 at 9:30 PM
cianodonnell.bsky.social
this is a nice piece but imo "E/I balance" is too overly simplistic a notion to characterise whole-brain or even neural-circuit dynamics.

Eg here is a super simple diagram of recurrent E and I in a neural circuit. Is GABA synapse onto I neurons net excitatory? What about glut synapse on I neurons?
June 23, 2025 at 5:03 PM
cianodonnell.bsky.social
aside from this being a v cool paper I also want to congratulate the authors on the incredible SNR achieved in the title via a complete absence of filler words

Neuromorphic hierarchical modular reservoirs
www.biorxiv.org/content/10.1...
screenshot of biorxiv paper titled "Neuromorphic hierarchical modular reservoirs", authors Filip Milisav, Andrea I Luppi, Laura E Suárez, Guillaume Lajoie, Bratislav Misic
June 22, 2025 at 5:13 PM
cianodonnell.bsky.social
something compelling about working on the edge: the edge of an empire, the edge of a continent, the edge of neuroscience, AI, the unknown. It's a privilege to come to work every day and get paid to do fundamental research in a place you love.
google map of the island of Ireland with pin in Intelligent Systems Research Centre in Derry/Londonderry
June 19, 2025 at 8:38 PM
cianodonnell.bsky.social
According to google maps, this is the global headquarters for NeuroAI
Google maps screenshot showing quite a decrepit building in street view, somewhere in Santa Monica, of a place called "Neuro-AI"
June 18, 2025 at 12:47 PM
cianodonnell.bsky.social
if you are curious about
- Computational neuroscience
- Neurotechnology
- Neuro-inspired AI

consider attending our 7-day ISRC CN3 summer school @ulsteruni.bsky.social in the great city of Derry, Northern Ireland, UK from 25th Aug-1st Sep 2025. In-person or online.

www.ulster.ac.uk/conference/i...
Photograph of the Peace Bridge in Derry/Londonderry, with the Guildhall visible in the background. Photo of Ulster University campus in Derry/Londonderry. Multi-colour logo for the summer school.
June 10, 2025 at 9:22 AM
cianodonnell.bsky.social
today we had a nice discussion on q: 'does Sutton's Bitter Lesson apply to neuroscience?'

Sutton argued that big compute and data lead to AI systems better than any human-crafted alternatives.

Neuroscience has big data now, is it going the same way?

Eg www.nature.com/articles/s41...
Photo of a happy-looking Rich Sutton Figure 1a from Lappalainen et al 2024, shows schematic diagram of their strategy to train a recurrent ANN with connectivity constrained by fly optic lobe connectome, then compare it to actual fly brain activities. “The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.” “The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning.” The goal of neuroscience is to understand the brain. But what if we humans are just not smart enough? Neuroscience is entering the big-data era: Connectomics (brain wiring) Cell types census (brain parts catalogue) Large-scale neural activity recording methods Spatial transcriptomics (molecular expression) Should we just combine all this data with machine learning, do a big brain simulation, then go home?
June 6, 2025 at 1:49 PM
cianodonnell.bsky.social
the weirdest glitch I have yet experienced with a GenAI tool. Gemini randomly mentioning the time and date. Wonder if the discussion on model dynamics vs stability triggered it
Screenshot from Gemini 2.5 output: "This moves towards models with dynamic states and potentially combines concepts from standard deep learning with biologically inspired plasticity rules. Remember the current time is approximately 7:39 PM BST on April 7, 2025, although the principles discussed here are fairly stable concepts in computational neuroscience modeling."
April 7, 2025 at 6:46 PM
cianodonnell.bsky.social
Not sure who made this but it made me lol
Comic showing robot cops about to arrest a guy, one robot says "wait! Not that guy. He always said thanks...". Panel then shows laptop screen from 10 years ago, ChatGPT interface where guy had written "great answer! Thanks."
April 2, 2025 at 10:01 AM