Patrick Mineault
@patrickmineault
Followers
19,491
Following
2,636
Media
753
Statuses
5,068
Neuro AI, vision, Python, open science. Previously engineer @ Google, Meta, Mila. Updates from ,
Montreal, QC, CA
Joined April 2011
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
España
• 431173 Tweets
Kroos
• 377861 Tweets
Project 2025
• 323058 Tweets
Portugal
• 272450 Tweets
Alemania
• 172076 Tweets
Nico
• 163582 Tweets
Lamine
• 150901 Tweets
Pedri
• 142909 Tweets
#ESPGER
• 98704 Tweets
المانيا
• 81035 Tweets
Almanya
• 72406 Tweets
Havertz
• 70704 Tweets
Morata
• 67061 Tweets
Wirtz
• 63871 Tweets
Carvajal
• 58985 Tweets
Musiala
• 49439 Tweets
Cucurella
• 49204 Tweets
スペイン
• 43096 Tweets
Merino
• 41833 Tweets
Dani Olmo
• 40083 Tweets
Rodri
• 38911 Tweets
Anthony Taylor
• 33915 Tweets
Ferran
• 31686 Tweets
De la Fuente
• 30523 Tweets
#Eurocopa2024
• 21851 Tweets
كروس
• 21832 Tweets
#GERESP
• 21747 Tweets
Spanien
• 18901 Tweets
Rudiger
• 18802 Tweets
クロース
• 16988 Tweets
Joselu
• 14824 Tweets
Fullkrug
• 14099 Tweets
Oyarzabal
• 13541 Tweets
Spain
• 10587 Tweets
ポルトガル
• 10547 Tweets
Pinned Tweet
Super excited to finally release this handbook on writing good research code. Based on my experiences going from research to industry and research again.
46
731
3K
Galactica is basically GPT-3 for science. It can write whitepapers, reviews, wikipedia pages and code. It knows how to cite and how to write equations. It's kind of big deal 1/ 🧵
🪐 Introducing Galactica. A large language model for science.
Can summarize academic literature, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more.
Explore and get weights:
285
2K
8K
15
274
2K
PaLM can answer questions that will make an average human's head explode
26
200
2K
It takes two minutes to integrate torch.nn.Transformer into a model. So why did I spend my Friday evening coding a visual transformer (ViT) from scratch? Sometimes you have to go slow to go fast 🧵1/
12
178
1K
There's a lot of bad vibes about academia, but if you're a big nerd, you enjoy thinking deeply about things for hours on end, and you're thinking about doing a PhD, by all means, do a PhD. Find a good lab with a supportive and respectful mentor, you'll cherish those 4+ years.
26
104
1K
I'm writing research code very differently than I used to in grad school - from anxiety-inducing brittle academic code to test-driven development. Would others be interested in learning how to write research code that won't bite back?
85
41
1K
A student writes: "I know python fairly well, but I want to improve my coding skills and learn to think in a mathematical and computational way". Some free resources to get started 📚🤔 1/
9
217
1K
I've written a booklet on creating good research code. Based on my experiences going from research to industry and research again. Releasing this semester. Looking for test readers. DM me!
63
118
967
Dall-E 2 has a bit of trouble understanding text. Here's an image unCLIP generated for the prompt, “A sign that says deep learning.”
30
157
955
There's been tremendous advancements in understanding the brain with deep unsupervised models. Here's my end-of-year write-up.
8
216
914
I searched "how to make an academic website" on a popular search engine, and boy, some of the advice in the top 10 results is outdated! Here's how you make an academic website👇 1/
17
138
904
Did you know you can use Python as a symbolic calculator, a la Wolfram or Maple? Sympy is pretty cool!
19
105
851
The field of neuroscience knows far too little about how nose boops affect the brain. To remedy this gaping hole in the literature, I scraped the
@neuralink
demo video to measure the response of neurons to nose boops. 1/
13
164
777
What are we going to do when we record 10k neurons at once? PCA?! I dive deep into the past and future of dimensionality reduction in neuroscience.
7
171
721
I often meet with grad students and postdocs interested in industry. I tell them how important it is to contact people. Most jobs aren't posted, and most people are nice. However, students often push back, thinking networking is transactional and a bit icky 1/
12
98
708
My guide on transitioning away from Matlab is out! 500 hours to Python proficiency
11
183
671
So it turns out bird's magnetic compass might be.. in their eyes? It boggles the mind what evolution can cook up
8
124
639
TIL Python classes can display custom HTML outputs in jupyter and colab via `_repr_html_`. That's how pandas table outputs work, for example. Here's Hello world and a map widget.
11
85
558
Life announcement 🎉: I joined
@Mila_Quebec
, the machine learning institute founded by Yoshua Bengio in Montréal. As senior ML scientist, I'll be pushing the boundaries of neuroAI. Stoked to be joining and working with
@g_lajoie_
,
@tyrell_turing
and the amazing Mila community
49
20
528
A solid list of neuroAI papers (and blog posts!) to get you up to speed on our exciting field
9
97
495
Many classical dimensionality reduction techniques, like PCA, can be expressed as maximum likelihood over latent factors of a generative model. That means you can do inference on these models via stochastic gradient descent in e.g. PyTorch. Notebook here:
7
69
463
We've uploaded ~100 talks from
@neuromatch
conference! Everything is tagged and linked and closed captioned via machine learning. That's 12% of our library of talks, the rest are coming over the next few days - don't forget to hit subscribe for the latest.
4
124
459
2022 saw a whirlwind of neuroAI research. Brain Dall-E. Neurons in a dish playing pong. GPT predicts how brains process language. I read through a bunch of papers so you don't have to. Read my review of 2022.
9
102
437
This is a cool paper: they train physical systems to perform neural-network-style computations. How does that work? 1/
6
97
426
Do you feel anxious that AI's emphasis on large-scale language models (LLMs) will crowd out academic labs? Few can afford 1000 GPUs drawing a small country's worth of electricity
56
32
403
When I get asked "what do you do exactly?", I usually ramble incoherently: I contain multitudes! No more – I now have a diagram!
13
18
395
I've experimented with a half-dozen interactive visualization frameworks over the last year. Here's my curriculum to make dynamic scientific visualizations in the browser from a Python user perspective.
2
80
393
A great series of lectures on linear algebra from Gilbert Strang, specialized for deep learning ; book here
4
98
359
A lot of folks are building semantic search engines on top of pre-trained word embeddings. Fine-tuning pre-existing embeddings can be finicky. A simple trick is to learn a matrix that multiplies the pretrained embeddings: inflate some dims, collapse others
2
50
340
If you don't have one already, make yourself an academic website. You, your interests, your CV, your publications, maybe a blog post or 10. Github pages, Hugo, 11ty, Wordpress, doesn't really matter how. Make it easy for people to find you in a space that you control
10
28
326
Some further tips for students thinking of a starting a PhD: 1. don't do a PhD with a bad mentor. Bad mentorship can run the gamut to being completely absent, sitting on drafts for years on end, being verbally abusive or a creep 1/
There's a lot of bad vibes about academia, but if you're a big nerd, you enjoy thinking deeply about things for hours on end, and you're thinking about doing a PhD, by all means, do a PhD. Find a good lab with a supportive and respectful mentor, you'll cherish those 4+ years.
26
104
1K
7
60
316
One of the wildest recent findings in matching AI vision to human vision is the surprising importance of adversarial robustness. 1/
7
41
310
Very cool interactive book on physics-based deep learning. So many jupyterbooks dropping - huge enabling technology for open education
0
56
299
We've prepared a great list of resources to get you up to speed for this summer's Neuromatch Academy! Free courses, lectures and tutorials
2
99
291
Every neuroscientist knows about the Jennifer Aniston neuron, but have you seen the Luke Skywalker cell? It's a single unit in the entorhinal cortex of an epilepsy patient which responds to Luke Skywalker... and Yoda. From Quiroga et al. (2009)
4
46
278
Neuro blogs thread 🎉 If there's a blog or mailing list about neuro, neuro-AI and related disciplines you like, reply here. I'll compile a list after. Self promotion is ok! 👇
15
58
276
I'm ready to call it: the ventral visual stream is not a convolutional neural net. 🧵1/
11
42
273
Old comp neuro papers can be a bit hard to read, but you find some real gems. Here's a proto-visual-transformer (ViT) from Ullman and Koch (1985).
2
51
267
We're at a golden age of neuro AI. I chatted about it with Yannic Kilcher - out now
At one point, Neuroscience informed Deep Learning. Today, it might be the other way around! I spoke to Patrick Mineault (
@patrickmineault
) about how to understand the brain by inspecting deep models. How are they similar? How are they different? Watch here
1
48
256
6
33
258
Favorite podcasts on machine learning, neuroscience or general nerdy matters?
53
15
258
NeuroAI is a big tent. Working on a new map of neuroAI approaches. 2 axes: 1) are you driven by curiosity or by applications? 2) what's your home turf, neuro or AI? Here's what I have so far. Comments appreciated
17
41
255
Excited to launch the new
@neuromatch
NeuroAI course this summer! This is our first advanced course on the shared principles of intelligence in artificial and natural systems. Applications are open until March 17th (for TAs) and March 24th (for students).
2
81
244
It's always been a dream of mine to be able to do neuroscience work from anywhere with internet access. I set up a deep learning rig at home that I could remotely access... then went to Mexico for a wedding and stress tested it. It worked pretty well! 1/ 🧵
7
22
238
Deep learning scientists hate him: one weird scikit-learn class to beat Kaggle competitions.
⚡️Preprint: Why do tree-based models still outperform deep learning on tabular data?
We give solid evidence that, on tabular data, achieving good prediction is easier with tree methods than deep learning (even modern architectures) and explore why
1/9
36
489
2K
4
40
239
Did you know there's an interactive debugger built into jupyter notebook/colab? Use %debug to start a debugger after an error. Use u and d commands to step up and down the call stack.
5
51
235
I made this scrollytelling simulation of balanced excitation and inhibition in biological neural networks last year. JS+observable+React+Idyll. Everything runs live in the browser. It's the most in the weeds I've ever gone into JS
6
34
233
I'm not vegetarian, but as a neuroscientist, I don't eat squid - any other neuroscientists feel the same?
"Octopus farming is immoral, given everything we know about this highly intelligent animal."
A thread 🐙
@philip_ciwf
336
3K
6K
37
19
226
What does it mean for a neural network to "be like the brain"? I revisited this paper from Alex Williams and co today, which I think asks the right questions and offers some tantalizing answers. 1/
A very interesting trend in neuroscience is the emergence of datasets with large-scale recordings repeated across many animals (see e.g.
@IntlBrainLab
)
How to make sense of variability across animals in these data?
Our NeurIPS '21 paper has a few ideas
6
162
649
6
50
220
A free textbook on control systems and accompanying Python code and toolbox. It's amazing the free educational resources that people are putting out there.
2
64
216
What's something every prospective PhD student should know or read before deciding on joining a lab?
47
22
208
My latest roundup of neuroAI papers! Equivariance, retinal computations, emotions, faces, wormholes and topography. Featuring work by
@niru_m
,
@JamesJDiCarlo
,
@TimKietzmann
,
@naturecomputes
,
@talia_konkle
,
@tylerraye
2
40
204
Calling all grad students - I'm giving a couple of workshops tomorrow Wednesday 28th on writing good research code, 10AM and 7PM Eastern. Learn the secrets of a former Google software engineer - me! Live and informal &
Structuring Code & Data workshops. Wed, July 28, 2pm – 4pm UTC & Wed, July 28, 11pm – Thu, July 29, 1am UTC
A workshop by
@patrickmineault
(computational neuroscientist, former software engineer @ Google & Facebook) on how to write research code that won't bite back.
1
6
29
6
72
203
Is early vision like a convolutional neural net? My long overdue blog post
@neurograce
@tyrell_turing
@tiagogmarques
@ArtDeza
@skornblith
@dileeplearning
@ylecun
2
42
201
Landmark paper in Nature demonstrating walking in a paraplegic patient via direct stimulation of the spine, decoding intent to walk from a brain implant (ECoG). 1/
This is a landmark achievement. 25 yrs ago, I watched Superman in a wheelchair after falling from a horse tell me that one day neuroscientists would fix spinal cord damage. Today, a digital bridge over the broken spinal cord enabled this guy to walk again
5
136
605
2
34
194
Did you know that matplotlib accepts 954 named colors from the xkcd color survey? Spruce up your plots with 'blue with a hint of purple', 'browny orange', and 'icky green'. Also includes actually useful colors like light gray, orange, deep green
6
41
194
I am really digging this mashup of network neuroscience and control theory
6
22
194
The phylogeny of CNNs is well known, from Hubel & Wiesel to Fukushima to
@ylecun
to AlexNet. By contrast, the deep lore of Transformers is far muddier. Why is it even called attention, and what does it have to do with visual attention? I need answers! Help me out 1/
11
22
192
I'm thinking of offering my coding handbook as a series of emails. Sometimes you don't have quite enough energy to read it all in one go. By getting an email everyday, it's a little nudge to keep going. Do you like this idea?
11
0
190
My wife asked me how I generate ideas and publish blog posts on . I think writing blog posts is hugely beneficial as a budding scientist to find your voice, so I figured I would share my process more broadly 🧵1/
6
17
191
Gflownets are slick methods of sampling from sets and graphs, but they're notoriously hard to grok. I went to a 3-day workshop and wrote all about it 🎉 Think MCMC + RL rolled up into one
2
32
190
When you teach, you sometimes realize you don't quite understand something. It forces you to reevaluate what you thought you understood, and breeds deeper knowledge. I find this oddly rewarding.
4
10
186
Diffusion models generate photorealistic images from noise, but don't have latent variables–or do they? This paper shows that the intermediate upsampling blocks extract a latent semantic segmentation at intermediate denoising steps
6
30
190
Send me your NeurIPS stuff! I'm gathering a list of neuroAI papers and workshops to share on the NeuroAI archive.
65
27
189
How do you balance a pole on a moving cart? You don't need RL! I wrote an interactive demo of classical control theory policies (LQR). Bonus: you can try to balance the cart manually, QWOP style. It's hard! Inspired by
@eigensteve
's control bootcamp
11
23
185
New tweeprint time! The brain is exquisitely sensitive to complex motion... how can it learn how to do that?
TL;DR by learning to estimate the organism's own motion from the visual input.
Paper here:
1/
8
49
183
I made a little power calculator in
@observablehq
. The power of a two-sample statistical test is the area of the test distribution (in orange) that is to the right of the critical value (in green) of the reference distribution. A little hard to explain, easier to visualize
1
20
181
It brought a tear to my eye to see the map of countries and territories (108!) served by
@neuromatch
academy 2021:
4
10
182
I'm using Github copilot like a private tutor to learn Haskell. Here it is telling me how to translate Python dicts to Haskell's maps. And of course it knows how to build a Markdown table. Pretty close to magic 🪄✨
3
23
176
I just finished Reading in the Brain by
@StanDehaene
. It's a damn good book – here's my review 1/
7
28
178
Before bilateral symmetry evolved, the sea floor was dominated by strange animals: some had 3-lobed symmetry, glide symmetry or no symmetry at all. They could look like mats, tubes or toroids, and some resembled modern plants. Welcome to the Ediacaran period.
3
23
177
The TL;DR paper abstract summarizer from
@ai2_allennlp
is seriously impressive. Best of all: it takes three lines of Python to run this on your computer thanks to
@huggingface
2
14
175
It's really impressive how similar CLIP is to the brain, R2-wise. Seems like language supervision forces the visual encoder to match the representational geometry of language, which is similar to that of the brain. From
@ariairaw
in Leila Wehbe's group
4
37
173
Kind of wild, but not too surprising: ChatGPT has the political leanings of a median Bay-Area-software-engineer
9
35
168
With the arbitrary milestone of 10k followers looming, it's time for a meta thread: how to be good at science twitter 🧵 1/
2
27
172
This study is wild - putting the experimenters' arm in casts for two weeks to measure the brain effect of limb disuse. Now that's commitment to science!
After 10 years in
#neuroscience
as an undergrad and
#PhD
student, I'm thrilled to announce my first ever scientific paper: Plasticity and Spontaneous Activity Pulses in Disused Human Brain Circuits
(thread below)
29
215
921
6
26
168
The Good Research Code Handbook just got its 200th star on Github. Feeling grateful 🥰. Read it here
0
30
165
Here's a rough timeline for setting up research lab code management, based on this thread. WDYT?
Give me your best tips/resources for code management in a research lab (with the typical challenges: people w/o proper CS training working largely individually and frequently coming and going)
23
31
172
5
15
166
Some people see a shiny new package that requires a CUDA upgrade and think "heck no"!, but not me! Adversity can only make me stronger! In unrelated news, my deep learning rig has been down for 36 hours.
8
9
159
Tim Roughgarden's class on algos and data structures is a great intro to the classic CS curriculum, and will make you a better coder over the long term. 2/
2
7
152
How do you do AI with low resources as an academic? You can't compete directly with Google, Facebook and OpenAI, but you can position yourself differently. A lot of important lessons also for neuroAI
0
27
157
While we are all laughing at slightly conscious matrix multiplications, the theoretical computer scientists were busy cranking out Conscious Turing Machines. This paper I tell ya...
2
16
156
What can AI get from Neuroscience? This short 2007 paper from
@stevempotter
is a time capsule of the pre-DL days: it's prescient and dated all the same and it's darn interesting. It starts off with a heck of a metaphor: 1/
5
24
159
I wrote this blog post a year ago about how the visual system is not convolutional. I didn't think much of it until somebody told me last week they had found it useful. I re-read it today: it *is* really interesting! You should totally read it
5
19
158
I just put the finishing touches on the matching algo for
@neuromatch
academy. NMA is going to be 1780 interactive students and 194 TAs spanning 66 countries and 9 time slots. 9:1 student: TA ratio.
3
19
157
It's very cheap to append to a Python list. Numpy arrays, however, are expensive to append to. What's up with that? Let's look under the hood! 1/
3
10
151
I'm getting questions from early-stage PhD students who eventually want to move to industry and want to prepare. Seems like the most useful stepping stone would be to go for an internship in years 3 & 4. But where does one find out about neurotech internships? 🤔
16
9
149
I've been meaning to write this post for a long time: what does it mean for a neural network to be like the brain? I get into the nitty gritty of scores that compare a neural net vs. the brain.
4
30
148
It's a strange world we live in that we gather ML training data from button presses rather than directly from brains. If you train a deep neural net directly to match fMRI, it can become better at image recognition. Too bad scanning is 1000$ an hour! 1/
6
20
148
I'm preparing for a 3-day workshop about neuro & AI. I wanted to make sure that I was well-prepared. I asked chatGPT to write a Python script to grab all the abstracts from all the workshop participants from OpenAlex. 1/
4
15
143
GPTs are all neocortex, no hippocampus. They store information in their weights, as opposed to having access to a semantically indexed db of memories. With chunk-based retrieval, RETRO achieves similar perplexity with 3x smaller models.
A comprehensive study of LLMs with and without retrieval. Retrieval helps with knowledge-intensive tasks significantly.
@wbx_life
@_weiping
@PengXu51108979
@MohammadShoeybi
@ChaoweiX
@ctnzr
2
23
91
3
24
141
Did you know that gray hair acts as an optic fiber? A very short nature article from 1989
6
29
143
Networking isn't about asking randos for a connection to a job. The best kind of conversation you can have is just informational. Learn about a person's work, how they came to be in their job, how you would fit into the ecosystem 2/
1
3
143
It took me a stint as a software engineer at Google to figure it out - but I think I could have shaved a few months off of my PhD if I had written code that didn't spontaneously combust.
6
4
141
Friend of mine asked for a good intro to comp neuro.
@neuromatch
academy's CN course has a comprehensive overview . For paper books, Neuronal Dynamics (
@compneuro_epfl
@NeuroNaud
) coherently present modelling from HH to cognition
4
22
141
Quick experiment: Every once in a while, I get approached by companies doing cool neuro/AI things and looking for staff. Would love to connect them to graduating students, postdocs, or industry people. If you're interested, submit your CV here:
3
31
135