✨Officially received my Doctorate degree from MIT today!✨
Extremely grateful to everyone who made this possible 🙏
Big thanks to both of my advisors, committee members, mentors, colleagues, friends and family. This wouldn’t have been possible without each & every one of you.
Celebrating the completion of my PhD
@MIT
BCS Stole Bestowal Ceremony today with my PhD advisor, Josh Tenenbaum. Grateful for his guidance and support!
@mitbrainandcog
@MITCoCoSci
We had a paper in ICLR! The title is: "Map Induction: Compositional Spatial Submap Learning for Efficient Exploration in Novel Environments". We propose the Map Induction hypothesis, empirically test it through human behavior, and formalize it through a computational model (1/n)
Finally a tweeprint on our recent preprint presenting Vector-HaSH !
Vector-HaSH extends MESH to unify two important and seemingly independent roles of hippocampus: Spatial Mapping and Episodic Memory !
Brief talk:
Preprint:
1/n
Thank you all for your kind wishes on my graduation!🎓
I am deeply grateful for the overwhelming support 🙏
Heartfelt gratitude to each one of you.
Some people have asked about my PhD thesis — here's a link to my PhD defense talk for those interested ✨
Thanks
@doristsao
for your kind & thoughtful words about our work! I am glad that you find it beautiful, creative and conceptually insightful!
For folks who would like to know more:
Short talk on Vector-HaSH
Short talk on MESH
@SebastianSeung
Btw, I don't know if
@FieteGroup
is funded by BRAIN, but their recent work understanding grid cell-place cell dynamics as a general mechanism for episodic memory that factorizes problem of building attractors from problem of assigning content to them, is such a beautiful and
Thank you so much
@SebastianSeung
for appreciating our work! I am glad to hear that you consider it a breakthrough in theoretical neuroscience!
For folks who would like to know more:
Short talk on Vector HaSH
Short talk on MESH
@doristsao
@FieteGroup
+1 to an amazing conceptual breakthrough in theoretical neuroscience that I think did not depend on big data or fancy technologies. What we can say is that BRAIN technologies will make the theory testable in a much more conclusive way than was ever possible.
Presenting poster II-122 @
#cosyne2023
today!
First model of hippocampal-entorhinal cortex that encapsulates key findings of entorhinal map plasticity: grid map fragmentation, interpolation at short timescales (local consistency), map merger at long timescales (global consistency)
Great line of work by
@criticalneuro
& colleagues showing that despite knowing the connectivity structure of a spatial map (or graph), LLMs fail to report the shortest path between two rooms (nodes). They hallucinate edges that don't exist, report longer paths, & fall into loops.
Delighted to share our
#neurips2023
paper w
@grockious
@hmd_palangi
et al
Evaluating Cognitive Maps & Planning in LLMs with CogEval
We test planning in 8 LLMs.
Failures like hallucinating invalid paths/falling in loops don't support emergent planning.
1/n
We have a paper in ICML! Title: "Content Addressable Memory Without Catastrophic Forgetting by Heteroassociation with a Fixed Scaffold". We propose a novel CAM (content addressable memory) architecture that generates a CAM continuum without a memory cliff. (1/n)
#ICML
#ICML2022
Huge congrats to Prof. Ila Fiete for being awarded the Swartz Prize for Theoretical and Computational Neuroscience!! So fortunate and grateful to be working with her!
Thanks to the Swartz foundation
@SwartzCompNeuro
, the SfN Swartz prize committee and the SfN
@SfNtweets
!
Touched to hear from so many brilliant female computational neuroscientists. For the young women and men in computational neuroscience — this is you next.
(1/5)
Excited to announce that our paper
"Content addressable memory without catastrophic forgetting by heteroassociation with a fixed scaffold" got accepted at
#ICML
Arxiv version of the paper:
Stay tuned for the final camera ready version!
Had a wonderful time presenting alongside
@behrenstimb
,
@dileeplearning
,
@lengyel_m
, Dora Angelaki, Ilker Yildirim; Irinia Higgins and Ishita Dasgupta from
@DeepMind
in the
#cosyne2022
workshop. Grateful for all the encouragement and insights from these amazing scientists!
Michał Januszewski is a Staff Research Scientist at Google Research who’s decoding the secrets of how we think.
Explore how the Google Research Connectomics team developed AI to create detailed brain maps similar to electrical wiring diagrams. →
If you are at
#cosyne2022
, stop by my poster III-058 today to learn about Map Induction: Composition of spatial regions for efficient exploration in novel environments.
#COSYNE22
Flocking behavior in birds is fascinating! In work
@MSFTResearch
(), I found that humans playing an Xbox game showed flocking (without explicit instruction or reward for it), while GPT based AI agents played solo. Next: AI that utilizes social information.
What's better than a holiday greeting from the McGovern Institute? How about a holiday greeting in 30 different languages (including Klingon!) from researchers across our labs. You don't want to miss this one - turn up the volume!
Just finished three years of the Computational and Theoretical Neuroscience journal club today!
Big thanks to Ila Fiete
@FieteGroup
for being the driving force for its inauguration in fall 2018, and
@ScienceMIT
for funding us through Science Quality of Life (SQoL) program. (1/n)
Super excited to be releasing AlphaFold 3 today, developed by
@IsomorphicLabs
and
@GoogleDeepMind
: our next generation AI model for predicting the biomolecular structures and interactions of proteins, DNA, RNA, small molecules, and more:
1/
SceneDiffusion optimizes a layered scene representation (during diffusion sampling) to obtain spatial disentanglement by jointly denoising scene renderings at different spatial layouts. This disentangles spatial info & appearance allowing spatial editing! Thanks
@liuziwei7
&team.
😼Move Anything in Your Picture😼
#CVPR2024
We propose 🏞️SceneDiffusion🏞️ to freely rearrange image layouts by layered scene diffusion
@CVPR
* It supports a wide range of spatial editing operations, e.g., moving, resizing and layer-wise editing
- Paper:
An important aspect of embodied AI is to enable effective navigation of 3D spaces. In this brief talk I present a Generative AI agent that efficiently explores partially observed novel spatial environments through map generation 🏘️🏣🏙️🏦.
I was recently invited as a guest for an AI podcast and had an exciting and fun conversation with
@kanjun
and
@joshalbrecht
, the CEO and CTO
@genintelligent
. Today I received this lovely gift and a very kind thank you note from them! Thank you so much
@kanjun
and
@joshalbrecht
!
Recently got invited to present the MESH model at
@GuangyuRobert
's lab and at
@TomasoPoggio
's lab (thanks to
@akshayrangamani
for inviting!)
It was fun discussing memory with scientists/theorists in both labs, and am grateful for all the intriguing questions & fun conversations!
We have a paper in ICML! Title: "Content Addressable Memory Without Catastrophic Forgetting by Heteroassociation with a Fixed Scaffold". We propose a novel CAM (content addressable memory) architecture that generates a CAM continuum without a memory cliff. (1/n)
#ICML
#ICML2022
@mitbrainandcog
Stole Bestowal Ceremony with Julianne Gale Ormerod and Sierra Vallin. Thanks to both of them for their tremendous efforts towards supporting and enriching my experience as a graduate student in the BCS PhD program throughout my time
@MIT
. Forever grateful to them!
I have always thought that philosophy and the study of the brain go hand in hand, and this beautiful piece of writing by
@criticalneuro
reinforces that thought ! It highlights how important memory is “To remember is to be”. More power to everyone studying memory !
I've studied memory for 15 years with an empirical lens, scanned it, modeled it.
"The logic of memory" is my philosophical musing on memory, what I found missing in scientific discourse.
Thanks to
@Philosoph_Salon
&
@esdsantos
's course on James Baldwin🙏🏼
Had a great time presenting our MESH model at ICML 2022 !
Checkout the spotlight talk here:
Full Paper:
Checkout the following twitter thread for a quick summary:
Big thanks to the ICML organizers !
We have a paper in ICML! Title: "Content Addressable Memory Without Catastrophic Forgetting by Heteroassociation with a Fixed Scaffold". We propose a novel CAM (content addressable memory) architecture that generates a CAM continuum without a memory cliff. (1/n)
#ICML
#ICML2022
I will be presenting our work on Map Induction @
#CogSci2022
in the poster session tomorrow from 8:30-10:30am. Drop by if you would like to chat!
Here's a short 5 min talk if you want a trailer:
2) The Map Induction based MAP-POMCP model that explores efficiently by inducing the map based on observed regions of the environment. The illustration shows the most likely map.
Note that the induced map is updated dynamically in an online way based on new observations!
Folks at
#cosyne2022
#COSYNE22
come check out our poster I-140 to learn about building a content addressable memory model that escapes catastrophic forgetting through a combination of a pre-defined memory scaffold and heteroassociation of dense arbitrary patterns.
Watch PhysDreamer synthesize action-conditioned 3D object dynamics in response to interactions. This requires perception of the physical material properties of objects. PhysDreamer takes a physics-based approach, leverages object dynamics priors learned by video generation models
3D Gaussian is great, but how can you interact with it 🌹👋? Introducing
#PhysDreamer
: Create your own realistic interactive 3D assets from only static images! Discover how we do this below👇 🧵1/:
Website:
2) The Map Induction based MAP-POMCP model that explores efficiently by inducing the map based on observed regions of the environment. The illustration shows the most likely map.
Note that the induced map is updated dynamically in an online way based on new observations!
Excited for the K. Lisa Yang Integrative Computational Neuroscience
#ICoN
center
@mcgovernmit
which will create advanced mathematical models and computational tools to advance our understanding of the brain. Honored to be a part this amazing initiative!
The
#ICoN
center will also provide four graduate fellowships to
@MIT
students each year in perpetuity. This year’s inaugural grad fellows include Mark Saddler from
@JoshHMcDermott
lab and
@sugsharma
from the Tenenbaum and Fiete labs. 🧵 4/5
Human perception of correlated motion of lines as a moving cube shown in the light show in
#Boston
on new years eve. Reminded me of the cool illusions and perceptual effects shown by
@vin_agarwal
in
@MIT
's Splash2021 organized by
@espmit
. Group photo of our group of teachers!
Map Induction hypothesis – that humans optimize exploration of new spaces by representing maps as composed of reusable reference frames – which can inferred by program induction to represent unseen space as composed of previously encountered regions. (2/n)
How to build a memory model that can continuously tradeoff number of stored patterns and pattern richness? Come to our poster at
#ICML2022
to get some answers!
Paper:
Slides:
Check the following twitter thread for a quick summary
We have a paper in ICML! Title: "Content Addressable Memory Without Catastrophic Forgetting by Heteroassociation with a Fixed Scaffold". We propose a novel CAM (content addressable memory) architecture that generates a CAM continuum without a memory cliff. (1/n)
#ICML
#ICML2022
Want to learn about a content addressable memory model that doesn't show catastrophic forgetting, but instead exhibits the desired memory continuum? Will be giving a talk on it
@cshlmeetings
today in the afternoon session.
#cshlNeuroAI
Full paper:
Interested in a memory model that gradually forgets similar to humans?
Come to our Spotlight talk (4:15-5:45pm session) @
#ICML2022
& come chat in the poster session today !
"Content Addressable Memory Without Catastrophic Forgetting by Heteroassociation with a Fixed Scaffold"
Vector-HaSH is the first Entorhinal-Hippocampal model to unify item, spatial & episodic memory. It uses the representational structure used in spatial mapping, to generate high-capacity episodic memory. Multiple experimentally observed hippocampal phenomenon emerge in it.
15/n
Really cool work combining inverse physics with inverse rendering. Finally a step towards enabling virtual garment fitting ! - an important use-case I have long thought about 😄.
How do we create realistic models of dressed humans directly from visual data?
We introduce PhysAvatar, a framework that estimates the shape, appearance, and physical parameters of dressed human avatars from multi-view videos.
Page:
(1/6)
Thus, humans build a distribution of possible maps through an inductive process after seeing a partial piece of the environment.
This Map Induction process is a conceptually new way to think about how humans navigate through the world in few- and zero-shot settings.
(10/n)
Congrats to
@AIatMeta
on Llama 3 release!! 🎉
Notes:
Releasing 8B and 70B (both base and finetuned) models, strong-performing in their model class (but we'll see when the rankings come in @
@lmsysorg
:))
400B is still training, but already encroaching
Can AI be programed to absorb inspiration, crave communication & hence creative expression? What about the active role between the piece of art and the viewer? For instance, The Tree of Life below by Gustav Klimt has been interpreted in multiple ways!
#ai
#aiart
#aiartcommunity
Workshop was on "Linking phenomena across levels of analysis: The need for a new multi-level reverse-engineering toolkit". We had a great panel discussion at the end with active participation from the audience which was so much fun!! Many thanks for all their questions/comments!
Excited to be on the program committee of the AMHN workshop on Associative Memories
@NeurIPSConf
! I have been working on an associative memory model of hippocampal episodic memory enabled by pre-structured spatial representations with Ila Fiete. Excited to share & see new ideas!
Excited to share the schedule for our
@NeurIPSConf
AMHN workshop at
#NeurIPS2023
:
Please consider submitting your latest & greatest (new work, NeurIPS, AAAI, ICLR, etc):
CfP
Deadline Oct 6 (AoE)
Submissions
These findings contribute to a growing body of evidence that mindfulness can change patterns of brain activity associated with emotions and mental health.
Spatial Memory: as the agent moves around, velocity inputs update the grid state, & landmark sensory inputs get associated with particular grid codes, i.e., particular locations in space. Allows bidirectional inference in familiar & zero-shot inference in novel environments.
8/n
@doristsao
@FieteGroup
+1 to an amazing conceptual breakthrough in theoretical neuroscience that I think did not depend on big data or fancy technologies. What we can say is that BRAIN technologies will make the theory testable in a much more conclusive way than was ever possible.
RAG is being widely used to circumvent the problem of hallucinations in LLMs emphasizing the need for efficient external memory storage & retrieval systems like MESH. This talk summarizes key aspects of MESH - a neural memory model inspired by neuroscience
We have a paper in ICML! Title: "Content Addressable Memory Without Catastrophic Forgetting by Heteroassociation with a Fixed Scaffold". We propose a novel CAM (content addressable memory) architecture that generates a CAM continuum without a memory cliff. (1/n)
#ICML
#ICML2022
We formalize the map induction hypothesis to quantitatively test its predictions. We model it by using a Bayesian generative framework consisting of four computational modules that can each be independent of the others and can express different modeling assumptions. (4/n)
The same core architecture in Vector-HaSH enables three kinds of high-capacity memory: recall of random items, recall of spatial information such as landmarks & their locations, and recall of sequential episodic memory through low-dim vector updates to the grid cell circuit.
7/n
Vector-HaSH has three core components: grid cells with their modular attractor dynamics; hippocampal cells with sparse continuous activations; and non-grid Entorhinal Cortex (EC) that encodes external sensory observations.
4/n
MESH factorizes the problem of storing memory into building attractors & assigning content to them, leading to a graceful trade-off b/w pattern number & pattern richness.
Brief talk:
Paper:
Tweeprint:
2/n
We have a paper in ICML! Title: "Content Addressable Memory Without Catastrophic Forgetting by Heteroassociation with a Fixed Scaffold". We propose a novel CAM (content addressable memory) architecture that generates a CAM continuum without a memory cliff. (1/n)
#ICML
#ICML2022
Episodic Memory: Vector-HaSH maps the problem of learning arbitrary sequences in a high dim space to simply learning low dim transitions on a sequence scaffold (enabled by low-dim velocity updates to grid cells), which is highly effective for learning sequential information.
9/n
📢Submissions for
#CCN2024
are now open at ! 📢
We welcome submissions for 2-page papers (deadline: 12 April) and Generative Adversarial Collaborations (GACs), Keynote+Tutorials, and (new this year!) Community Events (deadline: 5 April).
From Experiment 1, we find that map induction is critical for human like exploration.
The Uniform-POMCP model that doesn't use map induction fails to explain human behavior.
Quantitatively shown by the Log likelihoods of the three models given human behavioral data.
(8/n)
Following Mosers' eleven rooms, eleven maps exp, Vector-HaSH stores all maps w/ stable reconstruction by storing each room's map in a different region of vast grid coding space, resulting in orthogonalized maps across rooms. Grid cells can be examined within & across rooms.
12/n
The Region Extractor extracts candidate regions (or submaps) from known parts of the map.
The Map Generator generates a space of possible map completions by using an underlying probabilistic generative grammar to combine region primitives into a set of completed maps.
(5/n)
Observations of memory consolidation emerge in Vector-HaSH, consistent with the multiple trace theory of the hippocampus. Frequently seen/recalled memories get strengthened in Vector-HaSH & become more robust to lesions of the hippocampus relative to other stored memories.
14/n
Last week,
@MIT
president Sally Kornbluth and
@OpenAI
CEO Sam Altman
@sama
engaged in an interesting discussion on MIT campus delving into the "Promise and Perils of AI"
Witnessed by MIT students and faculty, here are some highlights on their views:
If you find a particular paper interesting, Litmaps gives you an option to find more papers like that one.
Click on "More like this" under a paper and Litmaps will add new related papers to the map.
In this work we formalized the Map Induction hypothesis computationally, by combining a Bayesian map induction model and an approximate belief-space planner.
We find that Map Induction outperforms the exploration performance of a Partially Observable Monte Carlo Planner.
(11/n)
We have a paper in ICML! Title: "Content Addressable Memory Without Catastrophic Forgetting by Heteroassociation with a Fixed Scaffold". We propose a novel CAM (content addressable memory) architecture that generates a CAM continuum without a memory cliff. (1/n)
#ICML
#ICML2022
Freaking thrilled to share that our big
@CIHR_IRSC
grant just got funded (at the 3rd percentile!?!). Feels like a big accomplishment, especially squeezing this in right at the end of our first year as a lab. Of course, the project also got not-funded 3 times first, so...
I would like to thank my advisors Ila Fiete and Josh Tenenbaum, my collaborator Sarthak Chandra, and members of
@FieteGroup
and
@MITCoCoSci
for their feedback and support.
Also thanks to
@mitbrainandcog
,
@mcgovernmit
,
@MIT
for enabling this research 🙏🙏.
Vector-HaSH extends MESH to build a model of the Entorhinal-Hippocampal system that can store episodic memories through velocity-driven low-dimensional shifts in the Grid cell circuit. K-hot label states in MESH are replaced by the modular & periodic Grid code in Vector-HaSH.
3/n
Vector-HaSH learns high-capacity seq. scaffolds of arbitrary shapes w/ very small # of cells relative to seq. length (coz it learns lowD velocity updates). Arbitrary sensory sequences are linked to sequence scaffold enabling storage & recall of exponentially long sequences.
10/n
Key idea: grid+hippocampus forms a pre-structured scaffold with frozen weights; sensory observations in EC are linked to the scaffold via heteroassociation. Analogous to a hippocampal clothesline forming a structured, invariant scaffold onto which external cues are “hooked”.
5/n
SWE-agent is pretty cool! It uses GPT-4 (or potentially any other LM) to take a GitHub issue and automatically tries to fix it. Best part it’s open source - thanks to the team at Princeton !
Unlike MESH, the scaffold in Vector-HaSH exhibits strong generalization. To learn the scaffold, the scaffold weights need to be trained on only a vanishing fraction of all grid code patterns. Scaffold can be learned once (through early exploration) & then held fixed for life.
6/n
Vector-HaSH models memory palaces-a mnemonic technique memory athletes use to remember long lists of items-they imagine walking thru a path in a familiar env. & place items on landmarks along the path. Vector-Hash perfectly recalls new items despite degraded landmark recall.
11/n
We compare human performance to three model hypotheses (variants of POMCP).
Uniform-POMCP: doesn’t use map induction
Maximum A Posteriori (MAP-POMCP): uses the most likely map from the inferred distribution.
Distributional (D-POMPC): uses the entire set of induced maps.
(7/n)
True. Vision would also be a primary input modality for Spatial Navigation - another area thats important for embodied AI. In addition to interacting with 3D objects, embodied agents need to explore & plan in 3D spaces, building generalizable spatial representations as in humans.
Llama-3 is closing the gap with GPT-4, but multimodal models gotta catch up. Vision capabilities of open models like LlaVA are far, far behind GPT-4V. Video models are even worse. They hallucinate all the time and fail to give detailed descriptions of complex scenes and actions.
Amazing talks at
#cosyne2021
so far! Enjoying looking at others' work and excited to present a poster on my work today at
#cosyne2021
from 3-5pm and 7-9pm EST. Drop by if you wanna chat!
The Debate Over “Understanding” in AI’s Large Language Models. Talk by Melanie Mitchell, Professor at the Santa Fe Institute. She presents arguments that have been made for and against “understanding” in LLMs & discusses potential ways of evaluating it.
Some cool illustrations showing the qualitative difference between exploration behavior exhibited by:
1) The Uniform-POMCP model that explores exhaustively since it doesn't use any spatial priors.
Here: agent (blue), reward (yellow), wall (red), empty space (black).
Multiple aspects of hippocampal phenomenology emerge in Vector-HaSH including experimental observations on context and direction specific maps: splitter cells, route dependent cells, directional cells, and task dependent cells.
13/n
Hopfield networks exhibit a memory cliff: addition of a pattern beyond critical capacity leads to catastrophic forgetting (loss) of all patterns. Existing networks each exhibit a memory cliff, & approach the envelope at only one point (a specific number of stored patterns). (6/n)
@SebastianSeung
Btw, I don't know if
@FieteGroup
is funded by BRAIN, but their recent work understanding grid cell-place cell dynamics as a general mechanism for episodic memory that factorizes problem of building attractors from problem of assigning content to them, is such a beautiful and
Pretty cool work reconstructing the neural circuits that carry visual information to the navigation center of a fruit fly brain. Enables generation of new hypotheses about transformation of visual information at various processing stages for generation of head direction signals!
For example, the ring neuron in the video receives information from a vertically elongated visual field. Thus, it may respond to vertical stripes or encode the horizontal location of visual stimulus without elevation information.
I just did my federal tax return and was thinking about the exact same thing. It would be wonderful to have an LLM assistant who could do my taxes for me every year !
New MIT CSAIL method automatically breaks LLAMA2-7B, GPT-3.5, Stable Diffusion, and more with curiosity-driven exploration!
This type of exploration helps test LLMs & vision models by generating diverse inputs/prompts that trigger unwanted responses from a target model:
In Experiment 2, we use environments with rewards distributed based on color cues.
We find that humans use a distribution of possible maps through map induction. Importantly this is a constrained distribution unlike the distribution in the Uniform model which is large.
(9/n)
New opening for a full-time lab tech in the Kanwisher Lab:
This job is a great stepping stone to graduate programs in cognitive neuroscience. Demonstrated serious interest in cog neuro and good coding skills required. Machine learning skills a plus.
One hypothesis is that this is because episodic memory is enabled by pre-structured spatial representations that support high-capacity memory. Some of these ideas are fleshed in this work:
Tweeprint:
Finally a tweeprint on our recent preprint presenting Vector-HaSH !
Vector-HaSH extends MESH to unify two important and seemingly independent roles of hippocampus: Spatial Mapping and Episodic Memory !
Brief talk:
Preprint:
1/n
Many congrats to Dr. Andrew Francl for an excellent defense of his thesis, "Modeling and Evaluating Human Sound Localization in the Natural Environment! I was a bit bummed to attend virtually due to a sick kid, but very proud of all his great work over the past 6 years!
@iclr_conf
Excited to present my work on Modular networks for high capacity pattern and sequence memory
@ICLR_brains
today! Drop by to chat and learn about a neural architecture inspired by the entorhinal-hippocampal system in the brain!