The visual world is composed of objects, and those objects are composed of features. But do VLMs exploit this compositional structure when processing multi-object scenes? In our 🆒🆕 #ICLR2026 paper, we find they do – via emergent symbolic mechanisms for visual binding. 🧵👇
When we see something that's moving, our memories about it end up projected forward in time: We remember it further along than it was. In a new paper in 𝘗𝘴𝘺𝘤𝘩𝘰𝘭𝘰𝘨𝘪𝘤𝘢𝘭 𝘚𝘤𝘪𝘦𝘯𝘤𝘦, out today and led by @dillonplunkett.bsky.social, we demonstrate that this happens even when there is 𝙣𝙤 𝙢𝙤𝙩𝙞𝙤𝙣 𝙬𝙝𝙖𝙩𝙨𝙤𝙚𝙫𝙚𝙧.🧵
Another new paper from the lab: Predictive theories like the SR imply that navigators who navigate differently should have cognitive maps which differ in predictable ways. Here we show that this holds in mouse hippocampal CA1.
www.sciencedirect.com/science/arti...
It has been so so fun to think with some of my favorite scientists about what it means to understand!
Vase with Irises - 1890
https://botfrens.com/collections/46/contents/15543
New paper titled "Tracing the Representation Geometry of Language Models from Pretraining to Post-training" by Melody Z Li, Kumar K Agrawal, Arna Ghosh, Komal K Teru, Adam Santoro, Guillaume Lajoie, Blake A Richards.
LLMs are trained to compress data by mapping sequences to high-dim representations!
How does the complexity of this mapping change across LLM training? How does it relate to the model’s capabilities? 🤔
Announcing our #NeurIPS2025 📄 that dives into this.
🧵below
#AIResearch #MachineLearning #LLM
Excited to share a new preprint w/ @annaschapiro.bsky.social! Why are there gradients of plasticity and sparsity along the neocortex–hippocampus hierarchy? We show that brain-like organization of these properties emerges in ANNs that meta-learn layer-wise plasticity and sparsity. bit.ly/4kB1yg5
https://tinyurl.com/heavyrnn
Connectome suggests brain’s synaptic weights follow heavy-tailed distributions, yet most analyses of RNNs assume Gaussian connectivity.
🧵⬇️ Our @alleninstitute.org #NeurIPS2025 paper shows heavy-tailed weights can strongly affect dynamics, trade off robustness + attractor dimension.
**Discovering network dynamics**
One more on estimating dynamics of complex systems, this time with symbolic regression
doi.org/10.1038/s435...
This image combines 50 x 1" exposures, ISO 800, obtained with my CANON 5D Mark III attached at primary focus of my Skywatcher Black Diamond 80mm f600mm (F/7.5) during the Total Lunar Eclipse on Wednesday 26 May 2021, between 9:00pm and 9:04pm, Sydney local time. The frames were manually aligned with Photoshop, then extracted as new frames to Lynkeos, that did the fine alignment. Drizzling x2 was also used. Then the combined image was taken to Photoshop, where the colour, highlights/shadows, and contrast were tuned. This got only the image of the eclipsed moon. Additionally, I combined the full 266 x 1" images I took during totality using Siril to get the stellar background. They two combined frames (eclipsed moon + stars) were combined with Photoshop with a bit of extra tuning to get a nice background. The image taken at 9:03pm was used as reference for the alignment. All the data were taken during the 4 hours ONLINE event "Conversaciones astronómicas bajo la luna eclipsada" that I organised with the "Red Andaluza de Astronomía" (RAdA), "Agrupación Astronómica de Córdoba" (AAC), "Asociación de Lengua y Cultura Española" (ALCE) de Australia and the "Spanish Researchers in Australia-Pacific" (SRAP-IEAP), that was followed by many schools in Spain. I also want to thank Alicia Lozano (RAdA), Héctor Socas-Navarro (Instituto de Astrofísica de Canarias), María Ribes (Universidad de Alicante) and Alberto Aparici (In...
Total Lunar Eclipse - 26 May 2021 - From Ángel López-Sánchez - https://flic.kr/p/2m1PLyZ
A photo of the short story Everything and Nothing by Jorge Luis Borges, printed in a book.
A photo of the short story Everything and Nothing by Jorge Luis Borges, printed in a book.
A photo of the short story Borges and I by Jorge Luis Borges, printed in a book.
A photo of the short story On Exactitude in Science by Jorge Luis Borges, printed in a book.
Jorge Luis Borges
Happy birthday Kay Sage. The artist & poet, famous for her surrealist paintings, often referencing architecture & the built environment, was born today in 1898. #surrealism
What do we talk about when we talk about "readout"?
I argued that our overly specialized, modular approach to studying the brain has given us a simplistic view of readout.
🧠📈
A huge prominence erupted from the eastern limb of the Sun today, forming this towering wall of plasma around 8pm CST on October 26, 2010. The pale circle shows the size of Earth in comparison! Image has been rotated 90º ccw from the original. See a video of this region in action here www.flickr.com/photos/lightsinthedark/5149863234/ Courtesy SDO (NASA) and the AIA consortium. Edited by J. Major. Solar Dynamics Observatory main site: sdo.gsfc.nasa.gov/ www.lightsinthedark.com
Prominence 10-26-10 - From Jason Major (jpmajor.bsky.social) - https://flic.kr/p/8NkYgk
Credit: ESA/Rosetta/NAVCAM, CC BY-SA IGO 3.0 - Processing: Elisabetta Bonora & Marco Faccin / aliveuniverseimages.com
ESA ROSETTA 14 July 2015 - From 2di7 & titanio44 - https://flic.kr/p/vpqn2z
Mimas and Epimetheus captured along the main rings. This natural color narrow-angle image shows just how much darker the unlit side of the rings is compared to icy moons.
Mimas, Epimetheus and Rings - From Gordan Ugarković (ugordan.bsky.social) - https://flic.kr/p/5oPunj
🧠🚨 How does the hippocampus transform the visual similarity space to resolve memory interference?
In this new preprint, we found that the hippocampus sequentially inverts the behaviorally relevant dimensions of similarity 🧵
www.biorxiv.org/content/10.1...
Replace Variational Autoencoder (VAE) with pretrained representation encoders (e.g., DINO, SigLIP, MAE) paired with trained decoders, which they terms as Representation Autoencoders (RAE).
Preprint Alert 🚀
Can we simultaneously learn transformation-invariant and transformation-equivariant representations with self-supervised learning?
TL;DR Yes! This is possible via simple predictive learning & architectural inductive biases – without extra loss terms and predictors!
🧵 (1/10)
So excited to see this preprint released from the lab into the wild.
Charlotte has developed a theory for how learning curriculum influences learning generalization.
Our theory makes straightforward neural predictions that can be tested in future experiments. (1/4)
🧠🤖 🧠📈 #MLSky
Tomorrow the next meeting of MIT #Consciousness Club. This is a Zoom link 🔗⬇️
Correct me if I'm off base here, but the speaker's idea of an "intelligible theory" sounds a lot like an extended intuition pump a la Dennet. I'm wondering in particular what he means when he says understanding doesn't require true theories (what's a "true theory", anyway? all models are wrong etc.)
engage with art on a human level. It makes for better science. After all, the process of discovery is not like the deductive thing that ends up in the paper. And finally it goes without saying that it's preferable to have our top experts stay in touch with the human experience. 🙃
I don't think science will ever "explain" the qualia of artistic experience, but I think there are many layers before we hit that ceiling that we have as yet not explored, that science can meaningfully address. This to me is independent of its being desirable for scientists to
I think it may be possible to understand why playing with certain visual principles "tickles the brain" (maybe it's similar to the 'aha' moment when grasping a metaphor?) as well as why the artistic medium carries the signal better to another mind than does a very accurate, dry representation.
translates it into something that can induce some version of that experience in another person (the viewer). Understanding visual perception is only one small part -- for example, what is engaging about an impressionist work vs something like a journalistic photograph or a diagram of the same scene?
While I deeply agree with this last post, I think the bit about intuitive understanding of visual perception doesn't mean impressionists are reduced to amateur neuroscientists. Rather, art (not exclusively visual art) takes some experience of the world meaningful to the artist and
Max Ernst Showing a Young Girl the Head of his Father
Max Ernst Showing a Young Girl the Head of his Father www.wikiart.org/en/max-ernst/max-ernst-s...
Cielo di piombo, ispettore Callaghan...
