Arna Ghosh
@arnaghosh.bsky.social
PhD student at Mila & McGill University, Vanier scholar • 🧠+🤖 grad student• Ex-RealityLabs, Meta AI • Believer in Bio-inspired AI • Comedy+Cricket enthusiast
Thanks Ken! ☺️
Here's the (more updated) NeurIPS version: proceedings.neurips.cc/paper_files/...
Also, more recently we extended the use of powerlaws for characterizing how representations change over (pre/post) training in LLMs. 🙂
🧵 here: bsky.app/profile/arna...
Here's the (more updated) NeurIPS version: proceedings.neurips.cc/paper_files/...
Also, more recently we extended the use of powerlaws for characterizing how representations change over (pre/post) training in LLMs. 🙂
🧵 here: bsky.app/profile/arna...
November 18, 2025 at 4:13 AM
Thanks Ken! ☺️
Here's the (more updated) NeurIPS version: proceedings.neurips.cc/paper_files/...
Also, more recently we extended the use of powerlaws for characterizing how representations change over (pre/post) training in LLMs. 🙂
🧵 here: bsky.app/profile/arna...
Here's the (more updated) NeurIPS version: proceedings.neurips.cc/paper_files/...
Also, more recently we extended the use of powerlaws for characterizing how representations change over (pre/post) training in LLMs. 🙂
🧵 here: bsky.app/profile/arna...
This is an excellent blueprint on a very fascinating use of AI scientist! And the results and super cool and interesting! 🤩
I have been asked this when talking about our work on using powerlaws to study representation quality in deep neural networks, glad to have a more concrete answer now! 😃
I have been asked this when talking about our work on using powerlaws to study representation quality in deep neural networks, glad to have a more concrete answer now! 😃
1. New preprint resolving a conundrum in systems neuroscience with an AI scientist, and humans Reilly Tilbury, Dabin Kwon, @haydari.bsky.social, @jacobmratliff.bsky.social, @bio-emergent.bsky.social, @carandinilab.net, @kevinjmiller.bsky.social, @neurokim.bsky.social
www.biorxiv.org/content/10.1...
www.biorxiv.org/content/10.1...
Characterizing neuronal population geometry with AI equation discovery
The visual cortex contains millions of neurons, whose combined activity forms a population code representing visual stimuli. There is, however, a discrepancy between our understanding of this code at ...
www.biorxiv.org
November 16, 2025 at 10:29 PM
This is an excellent blueprint on a very fascinating use of AI scientist! And the results and super cool and interesting! 🤩
I have been asked this when talking about our work on using powerlaws to study representation quality in deep neural networks, glad to have a more concrete answer now! 😃
I have been asked this when talking about our work on using powerlaws to study representation quality in deep neural networks, glad to have a more concrete answer now! 😃
Reposted by Arna Ghosh
Conrad Hal Waddington was born OTD in 1905.
His “epigenetic landscape” is a diagrammatic representation of the constraints influencing embryonic development.
On his 50th birthday, his colleagues gave him a pinball machine on the model of the epigenetic landscape.
🧪 🦫🦋 🌱🐋 #HistSTM #philsci #evobio
His “epigenetic landscape” is a diagrammatic representation of the constraints influencing embryonic development.
On his 50th birthday, his colleagues gave him a pinball machine on the model of the epigenetic landscape.
🧪 🦫🦋 🌱🐋 #HistSTM #philsci #evobio
November 8, 2025 at 4:03 PM
You mean the algorithms "generate" some auxilliary targets and then do supervised learning?
November 8, 2025 at 8:17 AM
You mean the algorithms "generate" some auxilliary targets and then do supervised learning?
Reposted by Arna Ghosh
I’m looking for interns to join our lab for a project on foundation models in neuroscience.
Funded by @ivado.bsky.social and in collaboration with the IVADO regroupement 1 (AI and Neuroscience: ivado.ca/en/regroupem...).
Interested? See the details in the comments. (1/3)
🧠🤖
Funded by @ivado.bsky.social and in collaboration with the IVADO regroupement 1 (AI and Neuroscience: ivado.ca/en/regroupem...).
Interested? See the details in the comments. (1/3)
🧠🤖
AI and Neuroscience | IVADO
ivado.ca
November 7, 2025 at 1:52 PM
I’m looking for interns to join our lab for a project on foundation models in neuroscience.
Funded by @ivado.bsky.social and in collaboration with the IVADO regroupement 1 (AI and Neuroscience: ivado.ca/en/regroupem...).
Interested? See the details in the comments. (1/3)
🧠🤖
Funded by @ivado.bsky.social and in collaboration with the IVADO regroupement 1 (AI and Neuroscience: ivado.ca/en/regroupem...).
Interested? See the details in the comments. (1/3)
🧠🤖
Reposted by Arna Ghosh
A tad late (announcements coming) but very happy to share the latest developments in my previous preprint!
Previously, we show that neural representations for control of movement are largely distinct following supervised or reinforcement learning. The latter most closely matches NHP recordings.
Previously, we show that neural representations for control of movement are largely distinct following supervised or reinforcement learning. The latter most closely matches NHP recordings.
Here’s our latest work at @glajoie.bsky.social and @mattperich.bsky.social ‘s labs! Excited to see this out.
We used a combination of neural recordings & modelling to show that RL yields neural dynamics closer to biology, with useful continual learning properties.
www.biorxiv.org/content/10.1...
We used a combination of neural recordings & modelling to show that RL yields neural dynamics closer to biology, with useful continual learning properties.
www.biorxiv.org/content/10.1...
Brain-like neural dynamics for behavioral control develop through reinforcement learning
During development, neural circuits are shaped continuously as we learn to control our bodies. The ultimate goal of this process is to produce neural dynamics that enable the rich repertoire of behavi...
www.biorxiv.org
November 6, 2025 at 2:10 AM
A tad late (announcements coming) but very happy to share the latest developments in my previous preprint!
Previously, we show that neural representations for control of movement are largely distinct following supervised or reinforcement learning. The latter most closely matches NHP recordings.
Previously, we show that neural representations for control of movement are largely distinct following supervised or reinforcement learning. The latter most closely matches NHP recordings.
Indeed! We show in the paper that the DPO objective is analogous to contrastive learning objectives used for self-supervised vision pretraining, which is indeed entropy-seeking in nature (shown in prev works).
I feel spectral metrics can go a long way in unlocking LLM understanding+design. 🚀
I feel spectral metrics can go a long way in unlocking LLM understanding+design. 🚀
November 3, 2025 at 1:51 AM
Indeed! We show in the paper that the DPO objective is analogous to contrastive learning objectives used for self-supervised vision pretraining, which is indeed entropy-seeking in nature (shown in prev works).
I feel spectral metrics can go a long way in unlocking LLM understanding+design. 🚀
I feel spectral metrics can go a long way in unlocking LLM understanding+design. 🚀
A big shoutout to @koustuvsinha.com for insightful discussions that shaped this work, and
@natolambert.bsky.social + the OLMo team!
Paper 📝: arxiv.org/abs/2509.23024
👩💻 Code : Coming soon! 👨💻
@natolambert.bsky.social + the OLMo team!
Paper 📝: arxiv.org/abs/2509.23024
👩💻 Code : Coming soon! 👨💻
Tracing the Representation Geometry of Language Models from Pretraining to Post-training
Standard training metrics like loss fail to explain the emergence of complex capabilities in large language models. We take a spectral approach to investigate the geometry of learned representations a...
arxiv.org
October 31, 2025 at 4:19 PM
A big shoutout to @koustuvsinha.com for insightful discussions that shaped this work, and
@natolambert.bsky.social + the OLMo team!
Paper 📝: arxiv.org/abs/2509.23024
👩💻 Code : Coming soon! 👨💻
@natolambert.bsky.social + the OLMo team!
Paper 📝: arxiv.org/abs/2509.23024
👩💻 Code : Coming soon! 👨💻
This work was done with dream team 🤩
@melodylizx.bsky.social @kumarkagrawal.bsky.social Komal Teru @glajoie.bsky.social @adamsantoro.bsky.social @tyrellturing.bsky.social
at @mila-quebec.bsky.social @berkeleyair.bsky.social @cohere.com & @googleresearch.bsky.social!
🧵9/9
@melodylizx.bsky.social @kumarkagrawal.bsky.social Komal Teru @glajoie.bsky.social @adamsantoro.bsky.social @tyrellturing.bsky.social
at @mila-quebec.bsky.social @berkeleyair.bsky.social @cohere.com & @googleresearch.bsky.social!
🧵9/9
October 31, 2025 at 4:19 PM
This work was done with dream team 🤩
@melodylizx.bsky.social @kumarkagrawal.bsky.social Komal Teru @glajoie.bsky.social @adamsantoro.bsky.social @tyrellturing.bsky.social
at @mila-quebec.bsky.social @berkeleyair.bsky.social @cohere.com & @googleresearch.bsky.social!
🧵9/9
@melodylizx.bsky.social @kumarkagrawal.bsky.social Komal Teru @glajoie.bsky.social @adamsantoro.bsky.social @tyrellturing.bsky.social
at @mila-quebec.bsky.social @berkeleyair.bsky.social @cohere.com & @googleresearch.bsky.social!
🧵9/9
Takeaway: LLM training exhibits multi-phasic information geometry changes! ✨
- Pretraining: Compress → Expand (Memorize) → Compress (Generalize).
- Post-training: SFT/DPO → Expand; RLVR → Consolidate.
Representation geometry offers insights into when models memorize vs. generalize! 🤓
🧵8/9
- Pretraining: Compress → Expand (Memorize) → Compress (Generalize).
- Post-training: SFT/DPO → Expand; RLVR → Consolidate.
Representation geometry offers insights into when models memorize vs. generalize! 🤓
🧵8/9
October 31, 2025 at 4:19 PM
Takeaway: LLM training exhibits multi-phasic information geometry changes! ✨
- Pretraining: Compress → Expand (Memorize) → Compress (Generalize).
- Post-training: SFT/DPO → Expand; RLVR → Consolidate.
Representation geometry offers insights into when models memorize vs. generalize! 🤓
🧵8/9
- Pretraining: Compress → Expand (Memorize) → Compress (Generalize).
- Post-training: SFT/DPO → Expand; RLVR → Consolidate.
Representation geometry offers insights into when models memorize vs. generalize! 🤓
🧵8/9
BONUS: Is task-relevant info contained in the top eigendirections?
On SciQ:
- Removing top 10/50 directions barely hurts accuracy.✅
- Retaining only top 10/50 directions CRUSHES accuracy.📉
As supported by our theoretical results, eigenspectrum tail encodes critical task information! 🤯
🧵7/9
On SciQ:
- Removing top 10/50 directions barely hurts accuracy.✅
- Retaining only top 10/50 directions CRUSHES accuracy.📉
As supported by our theoretical results, eigenspectrum tail encodes critical task information! 🤯
🧵7/9
October 31, 2025 at 4:19 PM
BONUS: Is task-relevant info contained in the top eigendirections?
On SciQ:
- Removing top 10/50 directions barely hurts accuracy.✅
- Retaining only top 10/50 directions CRUSHES accuracy.📉
As supported by our theoretical results, eigenspectrum tail encodes critical task information! 🤯
🧵7/9
On SciQ:
- Removing top 10/50 directions barely hurts accuracy.✅
- Retaining only top 10/50 directions CRUSHES accuracy.📉
As supported by our theoretical results, eigenspectrum tail encodes critical task information! 🤯
🧵7/9
Why do these geometric phases arise?🤔
We show, both through theory and with simulations in a toy model, that these non-monotonic spectral changes occur due to gradient descent dynamics with cross-entropy loss under 2 conditions:
1. skewed token frequencies
2. representation bottlenecks
🧵6/9
We show, both through theory and with simulations in a toy model, that these non-monotonic spectral changes occur due to gradient descent dynamics with cross-entropy loss under 2 conditions:
1. skewed token frequencies
2. representation bottlenecks
🧵6/9
October 31, 2025 at 4:19 PM
Why do these geometric phases arise?🤔
We show, both through theory and with simulations in a toy model, that these non-monotonic spectral changes occur due to gradient descent dynamics with cross-entropy loss under 2 conditions:
1. skewed token frequencies
2. representation bottlenecks
🧵6/9
We show, both through theory and with simulations in a toy model, that these non-monotonic spectral changes occur due to gradient descent dynamics with cross-entropy loss under 2 conditions:
1. skewed token frequencies
2. representation bottlenecks
🧵6/9
Post-training also yields distinct geometric signatures:
- SFT & DPO exhibit entropy-seeking expansion, favoring instruction memorization but reducing OOD robustness.📈
- RLVR exhibits compression-seeking consolidation, learning reward-aligned behaviors at the cost of reduced exploration.📉
🧵5/9
- SFT & DPO exhibit entropy-seeking expansion, favoring instruction memorization but reducing OOD robustness.📈
- RLVR exhibits compression-seeking consolidation, learning reward-aligned behaviors at the cost of reduced exploration.📉
🧵5/9
October 31, 2025 at 4:19 PM
Post-training also yields distinct geometric signatures:
- SFT & DPO exhibit entropy-seeking expansion, favoring instruction memorization but reducing OOD robustness.📈
- RLVR exhibits compression-seeking consolidation, learning reward-aligned behaviors at the cost of reduced exploration.📉
🧵5/9
- SFT & DPO exhibit entropy-seeking expansion, favoring instruction memorization but reducing OOD robustness.📈
- RLVR exhibits compression-seeking consolidation, learning reward-aligned behaviors at the cost of reduced exploration.📉
🧵5/9
How do these phases relate to LLM behavior?
- Entropy-seeking: Correlates with short-sequence memorization (♾️-gram alignment).
- Compression-seeking: Correlates with dramatic gains in long-context factual reasoning, e.g. TriviaQA.
Curious about ♾️-grams?
See: bsky.app/profile/liuj...
🧵4/9
- Entropy-seeking: Correlates with short-sequence memorization (♾️-gram alignment).
- Compression-seeking: Correlates with dramatic gains in long-context factual reasoning, e.g. TriviaQA.
Curious about ♾️-grams?
See: bsky.app/profile/liuj...
🧵4/9
October 31, 2025 at 4:19 PM
How do these phases relate to LLM behavior?
- Entropy-seeking: Correlates with short-sequence memorization (♾️-gram alignment).
- Compression-seeking: Correlates with dramatic gains in long-context factual reasoning, e.g. TriviaQA.
Curious about ♾️-grams?
See: bsky.app/profile/liuj...
🧵4/9
- Entropy-seeking: Correlates with short-sequence memorization (♾️-gram alignment).
- Compression-seeking: Correlates with dramatic gains in long-context factual reasoning, e.g. TriviaQA.
Curious about ♾️-grams?
See: bsky.app/profile/liuj...
🧵4/9
LLMs have 3 pretraining phases:
Warmup: Rapid compression, collapsing representation to dominant directions.
Entropy-seeking: Manifold expansion, adding info in non-dominant directions.📈
Compression-seeking: Anisotropic consolidation, selectively packing more info in dominant directions.📉
🧵3/9
Warmup: Rapid compression, collapsing representation to dominant directions.
Entropy-seeking: Manifold expansion, adding info in non-dominant directions.📈
Compression-seeking: Anisotropic consolidation, selectively packing more info in dominant directions.📉
🧵3/9
October 31, 2025 at 4:19 PM
LLMs have 3 pretraining phases:
Warmup: Rapid compression, collapsing representation to dominant directions.
Entropy-seeking: Manifold expansion, adding info in non-dominant directions.📈
Compression-seeking: Anisotropic consolidation, selectively packing more info in dominant directions.📉
🧵3/9
Warmup: Rapid compression, collapsing representation to dominant directions.
Entropy-seeking: Manifold expansion, adding info in non-dominant directions.📈
Compression-seeking: Anisotropic consolidation, selectively packing more info in dominant directions.📉
🧵3/9
When investigating OLMo (@ai2.bsky.social) & Pythia (@eleutherai.bsky.social) model checkpoints, as expected, pretraining loss ⬇️monotonically.
BUT
🎢The spectral metrics (RankMe, αReQ) change non-monotonically (with more pretraining)!
Takeaway: We discover geometric phases of LLM learning!
🧵2/9
BUT
🎢The spectral metrics (RankMe, αReQ) change non-monotonically (with more pretraining)!
Takeaway: We discover geometric phases of LLM learning!
🧵2/9
October 31, 2025 at 4:19 PM
When investigating OLMo (@ai2.bsky.social) & Pythia (@eleutherai.bsky.social) model checkpoints, as expected, pretraining loss ⬇️monotonically.
BUT
🎢The spectral metrics (RankMe, αReQ) change non-monotonically (with more pretraining)!
Takeaway: We discover geometric phases of LLM learning!
🧵2/9
BUT
🎢The spectral metrics (RankMe, αReQ) change non-monotonically (with more pretraining)!
Takeaway: We discover geometric phases of LLM learning!
🧵2/9
📐We measured representation complexity using the #eigenspectrum of the final layer representations. We used 2 spectral metrics:
- Spectral Decay Rate, αReQ: Fraction of variance in non-dominant directions.
- RankMe: Effective Rank; #dims truly active.
⬇️αReQ ⇒ ⬆️RankMe ⇒ More complex!
🧵1/9
- Spectral Decay Rate, αReQ: Fraction of variance in non-dominant directions.
- RankMe: Effective Rank; #dims truly active.
⬇️αReQ ⇒ ⬆️RankMe ⇒ More complex!
🧵1/9
October 31, 2025 at 4:19 PM
📐We measured representation complexity using the #eigenspectrum of the final layer representations. We used 2 spectral metrics:
- Spectral Decay Rate, αReQ: Fraction of variance in non-dominant directions.
- RankMe: Effective Rank; #dims truly active.
⬇️αReQ ⇒ ⬆️RankMe ⇒ More complex!
🧵1/9
- Spectral Decay Rate, αReQ: Fraction of variance in non-dominant directions.
- RankMe: Effective Rank; #dims truly active.
⬇️αReQ ⇒ ⬆️RankMe ⇒ More complex!
🧵1/9
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
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
October 31, 2025 at 4:19 PM
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
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
Very cool study, with interesting insights about theta sequences and learning!
1/
🚨 New preprint! 🚨
Excited and proud (& a little nervous 😅) to share our latest work on the importance of #theta-timescale spiking during #locomotion in #learning. If you care about how organisms learn, buckle up. 🧵👇
📄 www.biorxiv.org/content/10.1...
💻 code + data 🔗 below 🤩
#neuroskyence
🚨 New preprint! 🚨
Excited and proud (& a little nervous 😅) to share our latest work on the importance of #theta-timescale spiking during #locomotion in #learning. If you care about how organisms learn, buckle up. 🧵👇
📄 www.biorxiv.org/content/10.1...
💻 code + data 🔗 below 🤩
#neuroskyence
September 19, 2025 at 7:07 AM
Very cool study, with interesting insights about theta sequences and learning!
Reposted by Arna Ghosh
Together with @repromancer.bsky.social, I have been musing for a while that the exponentiated gradient algorithm we've advocated for comp neuro would work well with low-precision ANNs.
This group got it working!
arxiv.org/abs/2506.17768
May be a great way to reduce AI energy use!!!
#MLSky 🧪
This group got it working!
arxiv.org/abs/2506.17768
May be a great way to reduce AI energy use!!!
#MLSky 🧪
Log-Normal Multiplicative Dynamics for Stable Low-Precision Training of Large Networks
Studies in neuroscience have shown that biological synapses follow a log-normal distribution whose transitioning can be explained by noisy multiplicative dynamics. Biological networks can function sta...
arxiv.org
July 9, 2025 at 2:34 PM
Together with @repromancer.bsky.social, I have been musing for a while that the exponentiated gradient algorithm we've advocated for comp neuro would work well with low-precision ANNs.
This group got it working!
arxiv.org/abs/2506.17768
May be a great way to reduce AI energy use!!!
#MLSky 🧪
This group got it working!
arxiv.org/abs/2506.17768
May be a great way to reduce AI energy use!!!
#MLSky 🧪
Congratulations, Dan!! 😁
June 24, 2025 at 9:44 PM
Congratulations, Dan!! 😁
This looks like a very cool result! 😀
Can't wait to read in detail.
Can't wait to read in detail.
Does the brain learn by gradient descent?
It's a pleasure to share our paper at @cp-cell.bsky.social, showing how mice learning over long timescales display key hallmarks of gradient descent (GD).
The culmination of my PhD supervised by @laklab.bsky.social, @saxelab.bsky.social and Rafal Bogacz!
It's a pleasure to share our paper at @cp-cell.bsky.social, showing how mice learning over long timescales display key hallmarks of gradient descent (GD).
The culmination of my PhD supervised by @laklab.bsky.social, @saxelab.bsky.social and Rafal Bogacz!
Dopamine encodes deep network teaching signals for individual learning trajectories
Longitudinal tracking of long-term learning behavior and striatal dopamine reveals
that dopamine teaching signals shape individually diverse yet systematic learning
trajectories, captured mathematical...
www.cell.com
June 15, 2025 at 11:03 PM
This looks like a very cool result! 😀
Can't wait to read in detail.
Can't wait to read in detail.
Preprint Alert 🚀
Multi-agent reinforcement learning (MARL) often assumes that agents know when other agents cooperate with them. But for humans, this isn’t always the case. For example, plains indigenous groups used to leave resources for others to use at effigies called Manitokan.
1/8
Multi-agent reinforcement learning (MARL) often assumes that agents know when other agents cooperate with them. But for humans, this isn’t always the case. For example, plains indigenous groups used to leave resources for others to use at effigies called Manitokan.
1/8
June 9, 2025 at 6:16 AM