M Ganesh Kumar
M Ganesh Kumar

M Ganesh Kumar

Research Scientist · Max Planck Institute for Biological Cybernetics, Tübingen
Learning and adaptation in biological and artificial systems

I study how humans and machines learn from sparse, delayed feedback and adapt under uncertainty. I build reinforcement learning agents, Bayesian models and theories of learning in deep networks, and test them against human behavior and neural recordings. My goal is a unified theory of learning and adaptation that explains when learning succeeds, why it fails in psychopathology, and how to design AI that learns efficiently and works well with people.

Reinforcement learning Bayesian inference Interpretable AI Computational psychiatry NeuroAI
CV Email Google Scholar GitHub ORCID LinkedIn Bluesky X

News

  • 11/2026Upcoming seminar, Ministry of Education, Singapore: Learning, Memory and Meta-cognition
  • 10/2026One-shot learning of paired association navigation with biologically plausible schemas published in the Journal of Neuroscience
  • 08/2026Keynote at the Tamil Teachers Conference, Ministry of Education, Singapore: Minds, Machines and Mother Tongue
  • 07/2026CompleteP for RL at ICML 2026, Seoul
  • 02/2026Invited talk, Department of Physiology, National University of Singapore
  • 01/2026Joined the Max Planck Institute for Biological Cybernetics as a Research Scientist with Peter Dayan
  • 01/2026Predictive coding of reward in the hippocampus published in Nature
  • 09/2025Invited talk, Institute of Cognitive Neuroscience, University College London

Selected publications

Predictive Coding of Reward in the Hippocampus
Yaghoubi, M., Kumar, M. G.+, Nieto-Posadas, A., … Pehlevan, C., Williams, S., Brandon, M. P.
Nature 651, 414–420 · 2026
PaperCode
CompleteP for RL: Maintaining Feature Learning When Scaling Deep Reinforcement Learning
Lee, A.*, Kumar, M. G.*, Bordelon, B., Pehlevan, C.
International Conference on Machine Learning (ICML), Seoul · 2026
Paper
A Model of Place Field Reorganization During Reward Maximization
Kumar, M. G., Bordelon, B., Zavatone-Veth, J., Pehlevan, C.
International Conference on Machine Learning (ICML), Vancouver · 2025
PaperCode
One-Shot Learning of Paired Association Navigation With Biologically Plausible Schemas
Kumar, M. G., Tan, C., Libedinsky, C., Yen, S.-C., Tan, A. Y.-Y.
Journal of Neuroscience · 2026
PaperCode

* equal contribution · + I developed the theory and ran all model simulations · All publications →

Research

Humans and artificial agents must infer the hidden causes of their outcomes from sparse, delayed and ambiguous feedback. A student who fails an examination may blame their skill, the difficulty of the test or bad luck, and that explanation shapes what they do next. I study how such inferences are learned, how they go wrong, and how they are implemented, at three levels of analysis.

Behavior

How do people form beliefs, assign credit and adapt?

Bayesian models of how people attribute success and failure to skill, difficulty and luck. Human experiments relate these biases to depression, anxiety, culture and trust in AI.

Kumar & Dayan, in prep · Cloos et al., 2024
Computation

What makes learning efficient, and what makes it fail?

Reinforcement learning, feature learning and scaling in deep networks, and interpretable meta-RL agents that can be probed and steered.

Nature 2026 · ICML 2025 · ICML 2026 · CCN 2025
Implementation

How do neural circuits carry out these computations?

Spiking networks with neuromodulated plasticity, and how perturbations such as stress or NMDA receptor hypofunction change learning.

J Neurosci 2026 · Cerebral Cortex 2022

Applications

Mental healthComputational phenotypes of attribution biases, and AI support for reattribution in cognitive behavioral therapy.
EducationMetacognition and feedback in learning, wearable measures of engagement, and language models for Tamil.
Human-AI systemsInterpretable, data-efficient agents that people can trust to the right degree.

Publications

* equal contribution · † equal senior contribution · + see contribution note · In machine learning, ICML, NeurIPS, ICLR and ICCV proceedings are selective (<30% acceptance) archival publications.

Manuscripts in preparation

BDA Computational Model for Attribution Theory During Skill Inference
Kumar, M. G., Dayan, P.
BTempotron Learning in Deep and Recurrent Spiking Networks
Overwiening, J., Kumar, M. G., Sompolinsky, H.
BADScalable Yet Interpretable Meta-RL for Decision-Making With Uncertainty
Singh, A.*, Weichart, M.*, Rothe, K., Ansari, M. A., Zeraati, R., Dayan, P., Kumar, M. G.
BAStable Yet Volatile Representational Drift Facilitates Continual Learning
Huang, Y., Kumar, M. G., Dayan, P., Jacob, S. N.
BDTranscranial Direct Current Stimulation Rescues Metaplasticity in Hippocampal CA2 After Juvenile Stress
Chua, K., Kumar, M. G., Sreedharan, S.
BFeature Learning in the Hippocampus Encodes Prediction Errors
Han, Y., Yaghoubi, M., Brandon, M. P., Dayan, P., Kumar, M. G.
BAHeterogeneity Affords Stimulus-Association Learning
Wang, J., Kumar, M. G., Libedinsky, C., Yen, S.-C., Tan, A. Y.-Y.

Research articles

BPredictive Coding of Reward in the Hippocampus
Yaghoubi, M., Kumar, M. G.+, Nieto-Posadas, A., Mosser, C.-A., Gisiger, T., Wilson, E., Pehlevan, C., Williams, S., Brandon, M. P.
Nature 651, 414–420 · 2026
PaperCode
+ I developed the theory and ran all model simulations, as documented in the Author Contributions. The theory first appeared in Kumar et al. (ICML 2025) and predicted that reward-coding place fields would shift backward to fire before the reward. The experiments confirmed this prediction.
BAOne-Shot Learning of Paired Association Navigation With Biologically Plausible Schemas
Kumar, M. G., Tan, C., Libedinsky, C., Yen, S.-C., Tan, A. Y.-Y.
Journal of Neuroscience · 2026
PaperPreprintCode
BAHippocampal Neurogenesis With Cbln4 Deletion Impairs Fear ConditioningAccepted
Jagadeesh Shetru, V.*, Wang, J.*, Kumar, M. G., Libedinsky, C., Yen, S.-C., Tan, A. Y.-Y., Polepalli, J. S.
Communications Biology · 2026
ACompleteP for RL: Maintaining Feature Learning When Scaling Deep Reinforcement Learning
Lee, A.*, Kumar, M. G.*, Bordelon, B., Pehlevan, C.
International Conference on Machine Learning (ICML), Seoul · PMLR 306 · 2026
Paper
BAA Model of Place Field Reorganization During Reward Maximization
Kumar, M. G., Bordelon, B., Zavatone-Veth, J., Pehlevan, C.
International Conference on Machine Learning (ICML), Vancouver · 2025
PaperCodePoster
BDNeurocomputational Underpinnings of Suboptimal Beliefs in Reinforcement Learning Agents
Kumar, M. G., Manoogian, A., Qian, B., Pehlevan, C.†, Rhoads, S. A.†
Cognitive Computational Neuroscience (CCN), Amsterdam · 2025
PaperCode
BADetermiNet: A Large-Scale Diagnostic Dataset for Complex Visually Grounded Referencing Using Determiners
Lee, C.*, Kumar, M. G.*, Tan, C.
International Conference on Computer Vision (ICCV), Paris · 2023
PaperCode
BAA Nonlinear Hidden Layer Enables Actor–Critic Agents to Learn Multiple Paired Association Navigation
Kumar, M. G., Tan, C., Libedinsky, C., Yen, S.-C., Tan, A. Y.-Y.
Cerebral Cortex 32(18), 3917–3936 · 2022
PaperCode

Book chapter

BADTrends, Innovations and Challenges in Employing Interdisciplinary Approaches to Biomedical Sciences
Kumar, M. G., Ayyadhury, S., Murugan, E.
In Translational Research in Biomedical Sciences, Chapter 20. Springer Nature · 2024
Chapter

Short papers

BDUnsupervised Hebbian Learning Drives Biologically Interpretable Pattern Separation in a Hippocampal–Striatal NetworkSpotlight
Wang, J., Jagadeesh Shetru, V., Kumar, M. G., Libedinsky, C., Yen, S.-C., Tan, A. Y.-Y., Polepalli, J. S.
AAAI Workshops, Singapore · 2026
ATiny Personal Critic: A Lightweight Critic for Low-Compute Language Model Personalization
Singh, A., Kumar, M. G.
AAAI Workshops, Singapore · 2026
BAHuman-Like Compositional Learning of Visually Grounded Concepts Using Synthetic Environments
Lin, Z., Kumar, M. G., Tan, C.
ICLR Workshops, Singapore · 2025
Paper
ADGenerating and Validating Agent and Environment Code for Simulating Realistic Personality Profiles With Large Language Models
Cloos, N., Kumar, M. G., Manoogian, A., Cueva, C., Rhoads, S. A.
NeurIPS Workshop on Foundation Models for Science, Vancouver · 2024
Paper
BACompositional Visual Grounding of Word Concepts Through Embodied Reinforcement Learning
Lin, Z.*, Azaman, H.*, Kumar, M. G., Tan, C.
CVPR Workshops, Seattle · 2024
Paper
ADReject Option to Reduce False Detection Rates for EEG Motor-Imagery-Based BCI
Kumar, M. G., Ang, K. K., So, R. Q.
IEEE Engineering in Medicine and Biology Conference (EMBC), Korea · 2017
Paper

Students I supervise

Aditya Singh
Research Assistant · funded by my NUS Development Grant
Interpretable yet scalable neuro-cognitive models
First-author paper, AAAI 2026 workshop
Maximilian Weichart
Master's thesis, University of Tübingen · funded by MPI
Interpretable yet scalable neuro-cognitive models
Yuzhe Han
Master's thesis, University of Tübingen · funded by MPI
Feature learning in the hippocampus

Teaching

  • 2025Teaching Fellow, APMTH 50: Introduction to Applied Mathematics, Harvard University60 students. Led weekly tutorials of 30 students and supervised group paper presentations and final project proposals.
  • 2019Guest Lecturer and Practical Facilitator, LSM4213: Systems Neurobiology, NUSLecture on brain-computer interfaces. Facilitated the electromyography practical for 60 students.
  • 2019 – 2020Graduate Research Mentor, Special Programme in Science, NUSSupervised a group project on brain activity in schizophrenia.
  • 2019 – 2022Outreach and continuing education, NUSNUS Braincamp (2019), Beginning AI Through Neuroscience (2021), Neuroscience, AI and Medicine Workshop (2021), STEP NUS Braincamp (2022)

Teaching interests: computational models of cognition, brain-inspired AI, reinforcement learning, and computational psychiatry.

Service and leadership

Singapore Armed Forces

2011 – present · Major, volunteer reservist

I plan operations for a battalion of about 700 National Servicemen. I previously served as a combat engineer platoon and company commander. I hope to apply my research on learning to the resilience and well-being of National Servicemen.

Tamil community

Centre for Singapore Tamil Culture · Neuro-AI Expert

I am leading an initiative to develop a language model with strong Tamil capabilities. I organized the Tamil + AI symposium in 2019, presented AI Singapore's AI for Everyone series in Tamil, and speak to the Ministry of Education on learning, metacognition and AI for Tamil teachers. I have also produced, directed and acted in Tamil theatre.

Tamil Open Data Repository →

Industry and innovation

Nugen.ai

2019 – 2024 · Co-founder

Wearables that measure students' physiological states during learning, giving educators and parents feedback on engagement and stress. We funded and mentored an SUTD Capstone team that built the prototypes, and ran a preschool pilot.

Noova Tech

2026 – present · Consultant

Locally deployed agentic AI systems for data security.

Certified Scrum Product Owner and Certified ScrumMaster.

Outside work

✈️ Pilot

FAA-certified private pilot. Navigating landmarks at 5,000 ft requires 3D place fields.

🏍️ Motorcycle

It forces you to be present and situationally aware, or end up in the mud like I have.

🚗 Road trips

Long drives are good for thinking through hard problems, or for letting system one take over.

🏋️ CrossFit

My wife convinced me it was fun and useful, so that I can run after my son. I believe she is right.

Download full CV (PDF)

Appointments

  • 2026 – presentResearch Scientist, Department of Computational NeuroscienceMax Planck Institute for Biological Cybernetics, Germany · with Peter Dayan
  • 2023 – 2025Postdoctoral Fellow, School of Engineering and Applied SciencesHarvard University, United States · with Cengiz Pehlevan
  • 2022 – 2023Research Scientist I, Centre for Frontier AI ResearchA*STAR, Singapore
  • 2017 – 2018Research Engineer, A*STAR Artificial Intelligence InitiativeA*STAR, Singapore

Education

  • 2018 – 2023Ph.D. in Computational Neuroscience, National University of SingaporeAdvisors: Andrew Tan, Shih-Cheng Yen · Thesis committee: Thomas Yeo
  • 2013 – 2017B.Sc. (Honours with Distinction) in Life Sciences, National University of SingaporeMinors: University Scholars Programme and Special Programme in Science

Grants, fellowships and awards

  • 2025, 2026Young NUS Fellow, NUS Development Grant (SGD 40,000)
  • 2025Spotlight talk and Travel Award (EUR 1,000), Computational Psychiatry Conference
  • 2025Spotlight poster and Travel Award (USD 500), COSYNE
  • 2023Postdoctoral Fellowship in Computer Science, Harvard University
  • 2019Fujitsu Laboratories Fellow, MIT Center for Brains, Minds and Machines
  • 2018NUS Graduate School Scholarship
  • 2017NUSS Gold Medal for Outstanding Achievement
  • 2013A*STAR Undergraduate Scholarship

Selected invited talks

  • 2026Ministry of Education, Singapore (keynote, Tamil Teachers Conference; workplan seminar)
  • 2026Department of Physiology, National University of Singapore
  • 2025Institute of Cognitive Neuroscience, University College London
  • 2025Kempner Institute, Harvard · LMU–Harvard Young Scientists' Forum, Munich
  • 2022McGovern Institute and Department of Brain and Cognitive Sciences, MIT

Service

  • ReviewerNature Communications, Neural Networks, PLOS Computational Biology, Frontiers in Systems Neuroscience, IEEE TCDSNeurIPS, ICLR, ICML, CCN, COSYNE
  • OrganizingOrganizing Committee, Society for Neuroscience Singapore Chapter
← Back to Beyond Research
Pilot · Proof of concept

Tamil Open Data Repository

தமிழ்த் திறந்த தரவுக் களஞ்சியம்

A pilot collection of high-quality, teacher-created Tamil teaching material — built so that future AI systems can actually learn Tamil properly. Every contribution is credited, anonymised, and openly licensed.

Please read before contributing. This is a preliminary pilot, not a government platform. Its purpose is to demonstrate that Tamil teachers can collectively produce a high-quality corpus — evidence we intend to take to MOE, IMDA and AI Singapore to argue for a properly funded national repository. Contribute only material you own or are entitled to share, and only after removing every student name.

Why this matters

First — what is a "token"?

A token is the unit an AI language model actually reads and writes. It is not a word. Before a model sees any text, a tokenizer chops it into chunks drawn from a fixed dictionary of about 100,000 to 200,000 entries. Everything a model knows about a language, it learned by reading tokens of that language. Fewer Tamil tokens in, worse Tamil out. That is the whole problem in one sentence.

Counting tokens: tokens are occurrences, not distinct words

English"me, me, me, me"
me, me, me, me
7 tokens — not 1. Saying a word four times costs four times as much. One distinct word, four occurrences; corpus sizes always count occurrences.
English"The teacher marked the essay."
The teacher marked the essay.
6 tokens for 5 words. Common English words each get their own dictionary entry, so English runs at roughly one token per word.

The same word in Tamil costs about twelve times more

Englishteacher
teacher
1 token. The whole word is a single entry in the dictionary.
Tamilஆசிரியர்
▪▪▪▪▪▪▪▪▪▪▪▪
About 12 tokens under the GPT-3.5 and GPT-4 tokenizer — for the same single word. Each red chip is a fragment too small to be a letter, let alone a word. Measured across matched sentences, Tamil costs 9.9 times what English costs for identical content.
Don't take our word for it — count them yourself. OpenAI publishes a free tokenizer tool that shows you exactly how any text is chopped up. No account needed. It is worth five minutes, and it makes an excellent classroom demonstration.
  1. Type an English sentence. Watch the token count sit close to the word count.
  2. Now type the same sentence in Tamil. The count jumps several times higher.
  3. Try a single word — teacher, then ஆசிரியர். Same meaning, wildly different cost.
  4. Click "Show token IDs" to see the fragments. For Tamil, many will be meaningless slivers rather than letters.
  5. Switch the tokenizer between GPT-4o and GPT-3.5 & GPT-4 using the selector, and watch the Tamil count drop by roughly two-thirds. Nothing about the Tamil changed — only the dictionary did. That is the whole argument for this repository in one click.
Open the OpenAI tokenizer →
This is OpenAI's own tool and shows OpenAI's tokenizers. Other companies use different ones, so the exact numbers will vary — but the pattern holds across all of them.
A common misunderstanding: nothing is translated into English. Tamil text is not converted to English at any point. What happens is narrower and more mundane. The tokenizer's dictionary of chunks was built by counting what appeared most often in its training text, and that text was overwhelmingly English — so the dictionary is full of English words and pieces, and contains almost nothing from Tamil. When it meets Tamil it has no matching entry, so it falls back to ever-smaller pieces until it is emitting raw bytes. Under the GPT-4 tokenizer, 34% of Tamil tokens are stranded single bytes, against 9% for English: one token carries 4.86 bytes of English but only 2.03 bytes of Tamil. It is a gap in the dictionary, not a translation step — which is exactly why it is fixable. GPT-4o's newer tokenizer cut the penalty by 73% without changing anything about the Tamil language.
So 20 billion tokens is not 20 billion Tamil words. Because each Tamil word is shattered into several tokens, 20 billion tokens works out to roughly 1.6 billion Tamil words under the GPT-4 tokenizer, or about 6 billion under GPT-4o's newer one. The honest, tokenizer-independent figure is the one in pages: around 20 million pages of Tamil text.

There is a sting in this. Token counts flatter Tamil. The same idea written in Tamil produces more tokens than in English while carrying no more meaning — identical content, chopped finer. Measured in tokens, Tamil looks less far behind than it is. Measured in ideas, the gap is worse than it appears.

The consequence teachers feel without knowing why: at the same context limit, a Tamil user fits about 15% of what an English user fits. Paste a long composition into a chatbot in Tamil and it runs out of memory roughly six times sooner.

Large language models are trained largely on Common Crawl, an open snapshot of the public web. The July 2026 crawl contains about 2.14 billion web pages. Here is how the languages divide it up.

LanguageShare of pagesEst. pagesEst. tokens

Percentage figures are Common Crawl's own published statistics for crawl CC-MAIN-2026-30 (July 2026), which identifies the primary language of each page. Page and token counts are estimates derived from the 2.14 billion page total, assuming roughly 1,000 tokens of extracted text per page. Source: Common Crawl language statistics.

The goal: get Tamil to 1%

Around 80 million people speak Tamil, and it is one of the world's oldest living classical languages. Yet it makes up 0.0488% of Common Crawl — while English makes up 40.6%. English has roughly 830 times more presence on the web that AI systems learn from.

0.0488%
Tamil today
1.00%
Our target
~20.4B
Tokens still needed

To reach 1% of Common Crawl, Tamil needs roughly 20 million more pages of text — about 20 billion additional tokens.

An honest note on scale. Teacher uploads alone will not produce 20 billion tokens, and we should not pretend otherwise. What this pilot proves is something different and, for the pitch, more useful: that corrected, curriculum-aligned, expert-annotated Tamil can be collected systematically. Data of that quality is worth far more per token than scraped web text — a single marked composition teaches a model something no amount of scraped forum posts can. We are demonstrating the method so that it can be funded at national scale.

What this repository has collected so far

0
Estimated tokens
0
Pages
0
Contributions
0
Teachers
Progress toward the 20.4 billion tokens needed for 1% 0.0000%

Token counts are estimates: approximately 1,200 tokens per page of digital text and 600 tokens per page of handwritten or scanned material. They are indicative of scale, not exact measurements.

What we are looking for

The teacher's markings are the point

A clean, unmarked worksheet teaches an AI model very little. What is genuinely rare — and what no web scraper can ever collect — is your judgement made visible: the error identified, the correction supplied, and the reasoning behind it. Please contribute material that carries your input, whether as red ink, margin comments, tracked changes, rubric scores or annotations.

📄 Digital files

  • Word documents, PDFs, PowerPoint, plain text, spreadsheets
  • Worksheets and comprehension passages with model answers
  • Exam or assessment questions paired with correct answers and marking schemes
  • Student compositions with tracked changes or inserted comments
  • Lesson plans, grammar explanations, vocabulary lists, rubrics
  • Anything reflecting Singapore Tamil usage, register, or local context

📷 Handwritten and scanned material

  • Photographs or scans of marked student work — red ink corrections are ideal
  • Compositions with your annotations in the margins
  • Marked answer scripts showing what was wrong and what the correct answer is
  • Handwritten model answers and worked examples
  • Whiteboard or blackboard explanations you have photographed

Especially valuable

  • Question → student answer → teacher's correct answer triples. This is the single most useful format for training and evaluating AI.
  • Composition drafts showing the before and after of your feedback.
  • Anything involving spoken or colloquial Singapore Tamil, which current AI handles poorly.
  • Classical and literary Tamil with commentary — where AI is least reliable today.

Student privacy is mandatory

Remove every student name before you upload

This is not optional. Student names, class registers, index numbers, NRIC or FIN numbers, addresses, photographs of faces, and parent or guardian details must all be removed or blacked out before a file is submitted. Please also remove your school's identifying marks if the material is sensitive. If in doubt, leave it out.

A redaction tool is built into the uploader below. For photographs and scans, you can drag black boxes over any name directly in your browser. The redaction is permanently burned into the image, and the original file never leaves your computer — only the redacted version is uploaded. Camera and location metadata are stripped automatically in the process.

Why we do not erase names automatically. You may reasonably ask whether this could be automated. It can be attempted — handwriting recognition followed by name detection — but for handwritten Tamil it is unreliable, and an automated tool that misses one name in fifty is worse than no tool at all, because it invites people to stop checking. A teacher taking ten seconds to black out a name is both more accurate and more defensible. We would rather be slow and correct.

Where a name is part of the pedagogical content — a composition about a person, for instance — replace it with a generic name such as மாறன் or கமலா rather than deleting it, so the sentence still reads naturally.

Teacher access

Sign in or register

Access is limited to Tamil language teachers and educators. Enter your details and we will email you a secure sign-in link — clicking it verifies your address. There is no password to remember.

We store your name, email, role and institution solely to verify contributors and to credit contributions. Your email address is never published.

Questions

Who can see what I upload?

During the pilot, only the project maintainer. Nothing is published automatically. Before any material is released or shared with a partner organisation, contributors will be told what is being released and given the opportunity to withdraw their material.

Will I be credited?

Yes. Your name and institution are recorded with every contribution, and any published dataset will carry a contributor list. If you would prefer to contribute anonymously, say so in the notes field and we will honour it.

Can I withdraw material later?

Yes. Email m.ganeshkumar138@gmail.com and it will be removed. Once a dataset has been publicly released, prior copies cannot be recalled — which is why nothing is released without notifying contributors first.

What licence will the data carry?

The intention is an open licence permitting research and AI development with attribution required — most likely CC BY 4.0. This will be settled in consultation with contributors and any institutional partner before release, not decided unilaterally.

Is this an official MOE project?

No. This is an independent pilot run by Dr M Ganesh Kumar to demonstrate feasibility. If it succeeds, the case for a properly resourced national repository becomes much easier to make.

I have material but I am not a teacher — writers, publishers, historians?

Please get in touch directly. High-quality Tamil prose of any kind is valuable; this pilot is simply scoped to teachers first because classroom material with corrections is the rarest and most useful kind.