I am a Research Scientist at Google Research, where I work on biologically-inspired artificial intelligence.

Previously, I completed a PhD in Computer Science at Mila-Quebec AI Institute and McGill University, under the supervision of Blake Richards, where I focused on neuro-inspired unsupervised representation learning and biologically-plausible credit assignment algorithms. My PhD interships at Meta involved developing deep learning models for neuromotor interfaces.

During my MSc, also at McGill University, under the supervision of Marie-Hélène Boudrias and Georgios Mitsis, I developed deep learning techniques to advance neuroimaging analysis. I obtained my undergraduate degree in Electrical Engineering from Indian Institute of Technology Kharagpur.

---

📚 Selected Publications

Sorted By: Most recent
* denotes equal contribution

Tracing the Representation Geometry of Language Models from Pretraining to Post-training
M.Z. Li*, K.K. Agrawal*, A. Ghosh*, K.K. Teru, A. Santoro, G. Lajoie, B.A. Richards
NeurIPS 2025 | Paper → | Project website →

Harnessing small projectors and multiple views for efficient vision pretraining
A. Ghosh*, K.K. Agrawal*, S. Sodhani, A. Oberman, B.A. Richards
NeurIPS 2024 | NeurIPS Paper→ | arXiv Paper → | Project website →

Synaptic Weight Distributions Depend on the Geometry of Plasticity
R. Pogodin*, J. Cornford*, A. Ghosh, G. Gidel, G. Lajoie, B.A. Richards
ICLR 2024 Spotlight | Paper → | Code →

How gradient estimator variance and bias impact learning in neural networks
A. Ghosh, Y.H. Liu, G. Lajoie, K.P. Körding, B.A. Richards
ICLR 2023 | Paper → | Code →

α-ReQ: Assessing representation quality by measuring eigenspectrum decay
K.K. Agrawal*, A.K. Mondal*, A. Ghosh*, B.A. Richards
NeurIPS 2022 | Paper → | Blogpost → |

Estimating brain age from structural MRI and MEG data
A. Xifra-Porxas*, A. Ghosh*, G.D. Mitsis, M.H. Boudrias
NeuroImage (2021) | Paper → | Code → |

🚀 Current Research Directions

📐 Representation Geometry in Artificial and Biological Networks

LLMML / AINeuroscience

We are investigating the geometric structure and spectral properties of neural network representations to understand how artificial and biological systems encode information. This can be further used to develop task-agnostic metrics that predict model behavior (e.g., generalization) and reveal how learning occurs in large-scale vision and language models.

1. Li*, Agrawal*, Ghosh*, et al., NeurIPS 2025.
2. Ghosh*, Chorghay*, Bhaktiari, and Richards, NeurReps 2025.
3. Ghosh*, Agrawal*, et al., NeurIPS 2024.
4. Agrawal*, Mondal*, Ghosh*, and Richards, NeurIPS 2022.

🧠 Biologically-plausible Credit Assignment Algorithms

LLMML / AINeuroscience

We are investigating how the brain solves the "credit assignment" problem by characterizing how biological constraints—such as non-Euclidean plasticity and noisy gradient estimation—impact learning. This work provides a foundation for understanding how biological and neuromorphic systems can achieve efficient learning and superior generalization despite imperfect gradient signals.

1. Pogodin*, Cornford*, Ghosh, et al., ICLR 2024.
2. Ghosh, et al., ICLR 2023.

🔬 AI for Biological Applications

LLMML / AINeuroscience

Here are some of the projects where I developed AI pipelines for biological applications:

1. Xifra-Porxas*, Ghosh*, Mitsis, and Boudrias, NeuroImage 2021.
We developed and applied a machine learning framework to predict brain age of healthy subjects from MRI and MEG recordings.
2. Ghosh*, Sivakumar*, and Tom*, Intel Innovate FPGA Project (Silver & Iron Award Winners).
We built a deep learning framework on an Intel FPGA board for decoding imagined motor movements from EEG recordings to enable real-time typing.
3. Ghosh et al., Frontiers in Neuroscience.
We built a deep learning framework to identify exercise-induced EEG signatures, and improve population-level generalization of deep learning methods in a limited sample setting.
4. Ghosh, Singh, and Sheet, IEEE ICIIS.
We implemented deep learning algorithms for detection of mitotic nuclei from histopathology images, enabling detection of breast cancers and leukemia.

﹏𓂁﹏