Matt Perich
Neuroscience, engineering, AI, music. Asst. Professor / PI at University of Montréal and Mila.
- [Not loaded yet]
- Did a lot of outreach at Northwestern with other grad students there for a yearly "brain fair". Some fun things I remember: 1. Optical illusions are always fun 2. Put a sheep brain from the butcher in a blender to show brains are mostly fat
- 3. Backyard brains ephys stim kit hooked up to a cockroach leg. Play james brown from an iphone. Our body is electric! 4. Driving a lego mindstorms car w/ EMG signals from a surface electrode on the biceps. Pseudo-BCI demo!
- This year's Lab X-mas Family Photos are very zeitgeisty for those living in #Montreal: #Xmas #Transit
- [Not loaded yet]
- I mean, they're clearly on strike...
- I'm more and more convinced that low-dimensional manifolds in the brain are just an artifact of the experimental designs and analyses we use... 🧠📈 🧪
- Dimensionality reduction may be the wrong approach to understanding neural representations. Our new paper shows that across human visual cortex, dimensionality is unbounded and scales with dataset size—we show this across nearly four orders of magnitude. journals.plos.org/ploscompbiol...
-
View full threadSome personal speculation: orienting geometry w.r.t. readout buys flexibility and helps composition from lower-D representation, manifolds can be labile and change between tasks/contexts based on inputs etc, a low-D surface could store arbitrarily many representations (see SueYeon's capacity work)
- But ultimately I don't have much of "horse in the race" here about what is low-D, what is high-D. The brain probably uses a range of dimensionalities and strategies and some tasks are going to take a lot more dimensions than, say, limb movements. It's all fun to think about :)
- [Not loaded yet]
- Yes this I agree with 100%! I agree that the low dimensionalities we've seen to date are likely because we collect data in simple/limited behavioral regimes and short timescales.
- A simple bowl could be embedded in a 10^9 dimensional space. But naturally I don't think anyone argues "the brain is 10D" (or at they shouldn't).
- One of the reasons we wrote our recent perspective (www.nature.com/articles/s41...) was to try and argue that the talk around manifolds should be more about what it buys us conceptually as a framework, not about "is it low-D".
- [Not loaded yet]
- I could imagine lots of reasons they'd be seen. Robustness, predictability of future possible states, etc. Maybe you don't even want it, it's just a necessary consequence of wiring billions of neurons up and getting a somewhat-stable solution.
- For this paper, it’s worth considering that structure in brain activity need not be linear…
- I don’t have a strong opinion on what dimensionality brain activity should have. And dimensionality is just one aspect of manifolds. E.g., smoothness, which quite naturally fits with how we’d expect neural activity to transition between states
- Go work with Juan! I can vouch that this will be a very cool project.
- Super happy for having been awarded an @erc.europa.eu Consolidator Grant to continue our basic neuroscience work on the neural basis for motor control and motor learning. What a great way to set things up here at Champalimaud! 🎉 🚨 Job Alert for postdocs research technicians, and PhD students in 🧵
- There was never any point to having reference letters. That's why we've all started using AI to do this nonesense task. References should only be used for short-listed candidates for important positions/awards, and ideally, be done via a call to get the most honest opinion possible.
- Yes!
- [Not loaded yet]
-
View full threadYes but thankfully it's only a small cut in the current plan (which is kind of a win given how bad it could have gone). But certainly worse than a sustained increase, given the increasing applicant pool. www.science.org/content/arti...
- My CIHR committee was already at an 11% funding line last cycle (and I don't think it's the lowest)
- Fully agree. The unfortunate thing is that this is paired with cuts to federal research budgets. Importing highly competitive "big fish" while shrinking the pot for grants is not a great combo for the broader ecosystem... If Canada really wants to capitalize they need to think short and long term
- That being said, investment is investment, and it all could be a lot worse! Having a government acknowledging scientific excellence as a priority is a good signal.
- If you're interested in dynamical systems analysis for neuroscience, definitely check out @oliviercodol.bsky.social 's revised version of our RL paper! Very cool results in the new Fig 6, worth it regardless of if you saw our previous version or if it's all new. www.biorxiv.org/content/10.1...
- A big "get" for the Champalimaud! Excited to see what comes out of the Warehouse and the next phase of Juan's lab
- 🚨 I’m excited to say that my CIHR Project Grant was funded! My NHP lab is now full-speed-ahead, and I’m hiring experimentalists (postdoc, PhD student, and/or a tech/manager). We’ll do multi-region ephys during reaching/grasping in macaques, with behavioral and spinal perturbations.
- [Not loaded yet]
- Thanks!
- [Not loaded yet]
- Thanks!
- IMO Montreal is one of the best cities out there for neuro, and everyone in my lab will get to enjoy learning from and interacting with the NeuroAI folks at Mila. If modeling the computations behind multi-modal sensorimotor integration and adaptation is of interest to you, please reach out!
- More info on our funded project is here: webapps.cihr-irsc.gc.ca/decisions/p/.... In many ways, this directly follows our 2020 paper: www.biorxiv.org/content/10.1.... But we have much broader ambitions too and are building a flexible platform for NHP motor ephys. Should be a fun 5 years!
- Awesome work from @juangallego.bsky.social and lab. An interface from single motoneuron control in tetraplegia!
- 🚨 New preprint + thread 🧵 We've gone back to studying motoneuron control principles and their applications & here's paper #1: A proof-of-concept study showing that people with tetraplegic spinal cord injury can control up to 2DoF from a single intramuscular implant www.medrxiv.org/content/10.1...
- [Not loaded yet]
-
View full threadThough I will say that some evidence suggests it's not going to always be 1:1 with environment. E.g. the 2022 Gardner ERC place cell paper we mention in the article has place cells mapping even square environments into a toroid shape. Though the Guo 2024 CA1 paper has environment-hspaed manifolds
- Could be a cell type difference (Gardner 2022 focused on grid cells) or a region difference? Methods difference? Lots of interesting possibilities!
- [Not loaded yet]
- There's a long and fun conversation to be had here 🙂. But I agree, the "many-to-few" nature of neurons to manifolds allows considerable drift in single neuron activity without changing the manifold. Important in next steps to find out when, and how, neural drift changes manifold-level properties!
- [Not loaded yet]
- Indeed, IMO behavior (and environment, etc) are inextricably linked to manifold properties. For this reason, comparative (e.g. for same behavior, are manifolds different in different regions?) and causal (e.g., move activity on the manifold and predict behavioral changes) experiments are essential.
- Thanks for the kind words and really glad you enjoyed the article!
- 📰 I really enjoyed writing this article with @thetransmitter.bsky.social! In it, I summarize parts of our recent perspective article on neural manifolds (www.nature.com/articles/s41...), with a focus on highlighting just a few cool insights into the brain we've already seen at the population level.
- Neural manifold properties can help us understand how animal brains deal with competing and multifaceted information, execute flexible behaviors and reuse common computations, writes @mattperich.bsky.social. #neuroskyence www.thetransmitter.org/neural-dynam...
- Thanks to @emilysingerneuro.bsky.social for another opportunity to work with The Transmitter (which is an awesome publication), and of course the many, many long conversations on manifolds with @juangallego.bsky.social that shaped these articles 🙂
- Check out our new review/perspective (w/ @juangallego.bsky.social & Devika Narain) on neural manifolds in the brain! It was a lot of fun to think through these ideas over the past couple of years, and I'm excited it's finally out in the world! 🔗: www.nature.com/articles/s41... 📄: rdcu.be/ex8hW
- A lot has changed since we wrote our last perspective piece in 2017 (www.cell.com/neuron/fullt..., both in how we think about neural manifolds and in the prevalence in the field. We hope this paper provides a good primer for the ideas, and points towards some big open questions in this space.
- [Not loaded yet]
- [Not loaded yet]
- I guess I wouldn't think of babies as "learning" a foundation model. IMO a better analogy is that evolution learned the foundation model and babies during development are fine-tuning it (albeit in a "multi-task" way) for their bodies/experiences .
- Our new approach for scalable, generalizable, and efficient neural population decoding is now online! Here we focus on real-time BCI but I'm excited about all of our next steps building on this. Awesome work led by @averyryoo.bsky.social @nandahkrishna.bsky.social @ximengmao.bsky.social
- Explainer threads 🧵 are generous. They help science be a community, not a contest. Keep 'em coming!
- [Not loaded yet]
- BluePrints, naturally
- [Not loaded yet]
- [Not loaded yet]
- Subjectivity is hard to investigate, indeed. But if this is true then all "non-aphantasics" are equally misdescribing their subjective experience too... (said as someone with what I believe to be quite extreme aphantasia, and who cannot square my experiences with those described by others)
- [Not loaded yet]
- [Not loaded yet]
- Yeah and let’s not forget that in the US that college degree could easily put you $200,000+ in debt… it’s hard to justify knowledge for knowledge’s sake at that extreme personal cost, which IMO justifies the voiced regrets among American grads
- [Not loaded yet]
- [Not loaded yet]
- That's the earliest one I know of, at least on neural trajectories. There's a paper from ~ the 70s (I think?) that I've seen that used dim reduction to look at coactivation of motoneurons which starts to look slightly trajectory-like but not fully the same. Can't recall the authors though...
- [Not loaded yet]
- [Not loaded yet]
- 💯 I'm not convinced the most fun and innovative science is actually all that predictable from existing literature/things for model training. I could see it someday being useful to ask "here's my study, tell me all of the necessary controls and alternative explanations you can think of"
- [Not loaded yet]
- Amazing, congrats!
- [Not loaded yet]
- [Not loaded yet]
- This one from Fig 2 of POYO is a bit clearer; still a relatively small effect but this is in a simple task where single session models are pushing saturation, with 75%+ performance. I would hypothesize that if we subsampled single sessions to even smaller sets we'd get more gains from pre-training.
- I stumbled upon this website recently, and feel the need to make sure others know about it. Nice to be reminded that sometimes good things do come from the internet. applerankings.com
- [Not loaded yet]
- Yeah, there's a huge variance with that one. A honeycrisp at peak is a beautiful thing; almost as beautiful as an average pink lady! A honeycrisp off peak...
- [Not loaded yet]
- Yeah, I don't know nearly all of the apple varieties/hybrids here but something is off when Empire is lower than Gala and Pink Ladies are lower than Honeycrisp...
- Great initiative! The Transmitter is awesome.
- I'm excited to share another project we've been working on: The Transmitter's first book, which includes some of our favorite essays from our first ~9 months. Download a copy here: www.thetransmitter.org/transmitter-...
- This is an excellent thread summarizing many of my own objections. I’d like to add two thoughts: 🧵 1/n
- [Not loaded yet]
- 100% in agreement. Though I'd take a small bolder step and say there's a non-zero value for understanding too. If a big model shows that a thing is possible, we now have places to look (and potentially phenomena to explain) that helps progress towards insight/understanding long-term.
- [Not loaded yet]
- [Not loaded yet]
- I think about it somewhere in between, perhaps. Consider that the point of, say, practice is increase the bit rate of our outputs even in the face of the low bit rate constraints of our teleological bottleneck. True for me playing guitar, not really true for me on the couch...
- [Not loaded yet]
- Oooh excellent! Thanks for sharing!
- [Not loaded yet]
- Yup. If you only measure/quantify experimenter-defined task output you're ignoring all of the things the brain+body is doing in parallel, often subconsciously. Though caveat that I'm not on a university network and the article is paywalled so I can't actually read what they say right now....
