Lightnews — Scholar-powered news

Dmitry Krotov @krotov.bsky.social · May 29

The overlaid network illustrates a mathematical model, which represents the astrocyte as a network of processes that enables multi-synaptic communication and supports associative memory function.

Data credit: Leyka Nagendren, Cagla Eroglu.
Artist Credit: Annette Hui, IBM Research

Dmitry Krotov @krotov.bsky.social · May 29

The raw image was captured in October 2023 using an Olympus FV 3000 confocal microscope with a 60x objective lens and a z-step size of 0.5 μm across a 50–60 μm z-stack for high magnification and resolution.

1

Dmitry Krotov @krotov.bsky.social · May 29

A custom Imaris extension, 3D Sholl Analysis, was used to quantify the number of process intersections at 5 μm intervals from the cell soma, marked by the sphere, thereby visualizing the highly complex network of processes of the astrocyte.

1

Dmitry Krotov @krotov.bsky.social · May 29

The image summarizes the main message of our paper: astrocyte = network of processes that engage in information exchange and computation. The image shows a 3D reconstruction of an astrocyte from the visual cortex using Imaris Bitplane 9.9.

1

Dmitry Krotov @krotov.bsky.social · May 29

Key takeaways:
🔷 Astrocytes compute.
🔷 Dense Associative Memories can be built using astrocytes.
🔷 From the computational perspective astrocyte is a network of processes.
🔷 Memories can be stored, at least partially, in the biochemical machinery inside the astrocyte, as opposed to just synapses.

1 1

Dmitry Krotov @krotov.bsky.social · May 29

I am very excited about our new paper just published in PNAS.

www.pnas.org/doi/abs/10.1...

It was a pleasure to work on this idea together with @leokoz8.bsky.social and Jean-Jacques Slotine.

1 3 11

Reposted by Dmitry Krotov

Leo Kozachkov @leokoz8.bsky.social · May 28

Big week for astrocyte research: 3 new Science papers link astrocytes to behavior. We're excited to add to the momentum with our new PNAS paper: a theory, grounded in biology, proposing astrocytes as key players in memory storage and recall. w/ JJ Slotine and @krotov.bsky.social
(1/6)

1 13 34

Dmitry Krotov @krotov.bsky.social · Apr 25

John Hopfield on pursuing interesting things: “Doesn’t matter what they are - interesting is the important one”. www.youtube.com/watch?v=iYgs...

John J. Hopfield, Nobel Prize in Physics 2024: Official interview

YouTube video by Nobel Prize

www.youtube.com

4

Dmitry Krotov @krotov.bsky.social · Apr 23

More info here: nfam.vizhub.ai

Workshop on Associative Memory @ICLR 2025

New Frontiers in Associative Memories Workshop @ICLR 2025. Discuss the latest multidisciplinary developments in Associative Memory. Explore synergies between Associative Memories and LLMs.

nfam.vizhub.ai

2

Dmitry Krotov @krotov.bsky.social · Apr 23

I am heading to #ICLR2025 in Singapore. If you want to chat about associative memories, energy-based models, and related topics let’s connect! The highlight for me this year is the New Frontiers in Associative Memories workshop on Sunday April 27. Here is the schedule ⬇️

2 1 4

Reposted by Dmitry Krotov

Gaia Tavoni @gaia-tavoni.bsky.social · Mar 16

New ms! www.biorxiv.org/content/10.1...
I developed a normative coding theory for feedback-modulated canonical networks, providing a unified view of (1) efficient & predictive coding, (2) computational & algorithmic principles, (3) unimodal & multimodal sensory processing.

Convergence of efficient and predictive coding in multimodal sensory processing

The existence of pathways connecting different sensory modalities in the brain challenges the traditional view of sensory systems as operating independently. However, the reasons and mechanisms underl...

www.biorxiv.org

1 4 15

Dmitry Krotov @krotov.bsky.social · Jan 31

Website: nfam.vizhub.ai

Submissions on Open Review: openreview.net/group?id=ICL...

Workshop on Associative Memory @ICLR 2025

New Frontiers in Associative Memories Workshop @ICLR 2025. Discuss the latest multidisciplinary developments in Associative Memory. Explore synergies between Associative Memories and LLMs.

nfam.vizhub.ai

1 1

Dmitry Krotov @krotov.bsky.social · Jan 31

Associative memory is a prominent emerging paradigm in the landscape of mainstream AI and I expect exciting developments in this area in the next few years. Please take a look at the call for papers and submit by the deadline: February 14. It will be interesting! 😀

1

Dmitry Krotov @krotov.bsky.social · Jan 31

Now that ICML papers are submitted and we are in the midst of discussions on whether scaling is enough or new architectural/algorithmic ideas are needed, what can be a better time to submit your best work to our workshop on New Frontiers in Associative Memory @iclr-conf.bsky.social?

1 2 14

Reposted by Dmitry Krotov

Kyle Cranmer @kylecranmer.bsky.social · Jan 14

I am very honored to return to my alma mater to give a distinguished lecture on AI for Science at Rice University's Ken Kennedy Institute.
events.rice.edu/event/spring...

1 3 17

Dmitry Krotov @krotov.bsky.social · Jan 14

I am excited to announce the call for papers for the New Frontiers in Associative Memories workshop at ICLR 2025. New architectures and algorithms, memory-augmented LLMs, energy-based models, Hopfield nets, AM and diffusion, and many other topics.

Website: nfam.vizhub.ai

@iclr-conf.bsky.social

6 16

Dmitry Krotov @krotov.bsky.social · Dec 11

Wow, what a night! A heartfelt congratulations to my dear friend and collaborator John Hopfield and all the 2024 Laureates. #NobelPrize

1 13

Dmitry Krotov @krotov.bsky.social · Dec 9

What a wonderful day of Nobel lectures and inspiring conversations at the beautiful Aula Magna, Stockholm University! #NobelPrize

4

Dmitry Krotov @krotov.bsky.social · Dec 8

Physics is a point of view. Wise words from John Hopfield’s Nobel lecture. #NobelPrize

1 8

Dmitry Krotov @krotov.bsky.social · Dec 5

Please check out the post in the first quote and join us at the Scientific Methods for Understanding Deep Learning workshop at #NeurIPS2024 on December 15 to learn more.

2

Dmitry Krotov @krotov.bsky.social · Dec 5

We demonstrate empirically that spurious states exist in conventional DMs. Their existence is a distinct prediction of DenseAM theory, yet these states have been overlooked in the DM literature. Spurious states mark the onset of the memorization-generalization transition in DMs.

1 3

Dmitry Krotov @krotov.bsky.social · Dec 5

In our latest work, we leverage the correspondence between DenseAMs and Diffusion Models (DMs) to make a theoretical prediction: a similar phenomenon must occur during the memorization-to-generalization transition in DMs.

1 1

Dmitry Krotov @krotov.bsky.social · Dec 5

DenseAM theory predicts a distinct phenomenon when the amount of data surpasses the critical memory capacity: spurious states emerge. Spurious states are local minima of the energy function that fail to store each pattern in a unique basin of attraction (two or more patterns sharing the same basin).

1 1

Dmitry Krotov @krotov.bsky.social · Dec 5

Most of the work on Dense Associative Memory (DenseAM) thus far has focused on the regime when the amount of data (number of memories) is below the critical memory storage capacity. We are beginning to explore the opposite limit, when the data is large.

Bao Pham @baopham.bsky.social · Dec 5

Diffusion models create beautiful novel images, but they can also memorize samples from the training set. How does this blending of features allow creating novel patterns? Our new work in Sci4DL workshop #neurips2024 shows that diffusion models behave like Dense Associative Memory networks.

On the left figure, it showcases the behavior of Hopfield models. Given a query (the initial point of energy descent), a Hopfield model will retrieve the closest memory (local minimum) to that query such that it minimizes the energy function. A perfect Hopfield model is able to store patterns in distinct minima (or buckets). In contrast, the right figure illustrates a bad Associative Memory system, where stored patterns share a distinctive bucket. This enables the creation of spurious patterns, which appear like mixture of stored patterns. Spurious patterns will have lower energy than the memories due to this overlapping.

1 3 16

Dmitry Krotov @krotov.bsky.social · Dec 3

Check out our new #NeurIPS2024 paper, where we develop a distributed representation for a broad class of Dense Associative Memories. Similarly to the traditional Hopfield Network, the number of weights can be kept fixed and independent of the number of memories when new patterns are introduced.

Ben Hoover @bhoov.bsky.social · Dec 3

Excited to share "Dense Associative Memory through the Lens of Random Features" accepted to #neurips2024🎉

DenseAMs need new weights for each stored pattern–hurting scalability. Kernel methods let us add memories without adding weights!

Distributed memory for DenseAMs, unlocked🔓

1 4