Dan Goodman

Yeah it seems weird to cast it as being about the language when it should be about how good the new algorithms they've developed are (assuming that's the main difference).

My guess is that it won't scale beyond a few hundred or thousand neurons because you're basically setting it up as a very high dimensional nonlinear DE (dimension proportional to number of neurons).

Trees with leaves of every shade from dark to light green, yellow, orange, brown, red and purple. Background is flat white overcast sky.

The skies in your new home are not so blue, but the colours can be beautiful at this time of year, even in the rain.

Simulation of networks of spiking neurons: A review of tools and strategies - Journal of Computational Neuroscience We review different aspects of the simulation of spiking neural networks. We start by reviewing the different types of simulation strategies and algorithms that are currently implemented. We next revi...

It's probably all in here:

link.springer.com/article/10.1...

Continuous synapses would solve that problem but raise others, namely you'd lose the efficiency you get from temporal sparsity (every neuron would have to continuously contact every other connected neuron).

People do this but it doesn't scale well to larger networks.

The problem you'd run into is that if spikes can occur off the grid then you have to numerically solve for the threshold crossing time (expensive) and you have to propagate the effect of the spike across the network at that time which means interrupting solvers for other neurons.

If spikes have to be a multiple of dt then the error is order dt even if the DE solver has better order, so it's kind of pointless using a higher order solver. This isn't entirely correct but not far off.

I don't really use differential equation solvers myself. You generally need to simulate SNNs with spikes constrained to a fixed time grid which introduces error of order dt, so not much point going beyond Euler method.

Yeah it might well be true in this case, although just from a quick look at those graphs it looks like they're using different algorithms so it's probably more about that than the code. Generally you don't see 1000x differences from the way something is coded.

Generally not a good idea to believe benchmarks run by one org/library that show that their solution is the fastest. The only performance benchmarking I'd believe is where different groups of developers agree methodology in advance. That said, Julia does have a good rep for DEs.

This can't be real can it? An AGI paper from many big names in the field that invents a metric of AGI and claims we're 58% of the way there, and then the references are fake, generated by an LLM? That's got to be an elaborate hoax doesn't it? It's too perfect.

Thanks! Saw him outside my place when I was taking the bins out, so I rushed back in to grab my camera and he obliged. 🙂

It's better than it sounds. 😂 Well I say that but it's probably been 30 years since I last had one.

It can't possibly be true either. It's not possible to have grown up in Britain and never had curry or kebab.

I agree with her that scones are great though.

Squirrel on a concrete wall with patches of moss, a blurry street in the background.

Let the nut collecting begin! #photography

Yeah getting old sucks. 😉

Excellent funding opportunity: $100k/year for two years for a postdoc position in Montreal. Apply if you're interested in joining Montreal's thriving #NeuroAI ecosystem.

Neuromatch Academy 2026 is coming!

✅ Hands-on projects
✅ Global community
✅ Affordable fees

Applications open Feb 2026.

Learn more: neuromatch.io/courses/
Sign up for updates: neuromatch.io/mailing-list/

#ComputationalNeuroscience #DeepLearning #ClimateScience #NeuroAI #Neuroscience #AI

Yes maybe it's worth lecturing a course the first couple of times to iron out those bugs before switching out to flipped.

I guess I can see it could work as a sort of first pass for answering questions not covered in the course, giving you ideas of where to look. A sort of search engine.

Does it make self education efficient? I haven't looked into this but I would have assumed that the very high rate at which LLMs produce false information would impair their use in this way. Given the free availability of well designed and tested courses, why would anyone do this?

It's tricky to get right I think, but so are lectures in a different way.

Flipped classroom works if it's well designed. I still use it and my teaching evaluations have generally been much higher since I started doing so. It's so much better use of my time to spend the 2h per week with my class walking around interacting with them while they solve problems than lecturing.

AI is no threat to teaching, but the free availability of high quality courses online is, to some extent. It's not the only reason students go to university though. AI is a threat to assessment, and we still don't have a good solution. Neither does this piece.

Dear @elsevierconnect.bsky.social, having a default opt-in to marketing emails is in breach of the GDPR. Please correct this.

You're assuming they didn't know...

📣 BIG NEWS EVERYONE. I am so excited to announce…

🎉 I’m moving to University College London @ucl.ac.uk to join the Experimental Psychology department in @uclpals.bsky.social! 🎉

The big move happens in spring/summer. So I’m already exploring recruiting staff & students at UCL for fall 2026!

🚨Registration for Mechanistic Basis of Foraging 2025 closes October 20th 🚨

Speakers
@mkflugge.bsky.social @aaronbornstein.bsky.social @benhayden.bsky.social @ellileadbeater.bsky.social
@davidrobbe.bsky.social @dlbarack.bsky.social @becketebs.bsky.social

uobevents.eventsair.com/the-mechanis...

Deloitte to pay money back to Albanese government after using AI in $440,000 report Partial refund to be issued after several errors were found in a report into a department’s compliance framework

Deloitte repaying the Australian government for a $440k report written by ChatGPT. Wasn't this always the model of consultancy firms? They write a plausible sounding case for doing what you already decided to do using recent grads with no real expertise?

www.theguardian.com/australia-ne...

Schematic of awake-dreaming learning. (a) Awake phase: over 100 real frames, the agent and model networks interact with the Pong environment. (b) Dreaming phase: over 50 imaginary frames, the agent is disconnected from the real environment and instead interacts only with the world model. This alternation between real and simulated experience boosts sample efficiency.

#Spiking #neural #networks ( #SNN ) running on continuous-time, noisy, and highly variable computing substrates can learn reliably with #ReinforcementLearning ... Not only in real brains, but also in mixed-signal #neuromorphic hardware! 😇