Tech Founder & Data Scientist. Interested in scientific innovation, computer science, machine learning, AI, statistics, history, astronomy, 🇮🇳 and green tech.
Gurgaon, India
Joined April 2009
- Tweets227K
- Following2.1K
- Followers1.5K
- Likes32K
Pinned Tweet
Thoughts on all the #AI doomsday scenario being discussed on the internet #existential #Crisis :
[TL,DR : AI is very far from being an existential threat, they just want regulatory capture ]
quora.com/What-are-some-pote…
Muktabh Mayank retweeted
📝Fresh from the press:
"Geometric Interpolation and Geometric Jensen-Shannon Divergence Between Gaussian Measures On Hilbert Space", joint work with Hà Quang Minh
IEEE Transactions on Information Theory.
Publisher:ieeexplore.ieee.org/document…
Arxiv:arxiv.org/abs/2506.10494
Muktabh Mayank retweeted
A zero gradient doesn’t always mean there’s nothing left to learn.
I’m happy to share PolyStep, now published in #TMLR. It explores a question I find fascinating: how do we train networks when hard spikes, quantization, or discrete routing leave us with little useful gradient information?
These operations can produce losses with flat regions and sudden jumps. A tiny parameter change may tell us almost nothing, while a larger one reveals a meaningful difference.
PolyStep works at that finite scale. It evaluates structured parameter probes along the vertices of a rotated polytope, then uses the observed losses to weight the probe directions and take a step. The weights come from an optimal transport formulation, with a simple softmax solution in the main method. Training uses only forward passes.
The intuition I find most useful is that the scale at which we look determines which learning signals become visible. Probe radius controls what the optimizer can see, temperature controls how strongly it responds, and step scale controls how far it moves.
The PolyStep repo includes a small synthetic classification example with hard LIF spikes and no surrogate gradients, alongside CartPole policy search that learns from episode returns without gradients through the actions or simulator.
There is a practical tradeoff: those forward evaluations cost compute, and backpropagation is usually faster when useful gradients are available. I’m especially interested in the cases where hard operations are part of the behavior we want to preserve during training.
Many thanks to Viet T. Nguyen for the lovely interactive visualization that makes the geometry much easier to see.
If you’re working with spiking models, quantization, discrete routing, or simulators, I’d love to hear what you try, what works, and where it gets stuck. Questions, feedback, and contributions are very welcome.
Code: github.com/anindex/polystep
Paper: openreview.net/forum?id=8mlc…
Interactive visualization: vietngth.github.io/polystep-…
Muktabh Mayank retweeted
Expository article: "Exploring Neural Dynamics with Time Delays: A Numerical and Bifurcation Analysis Approach" (by Martin R Nelson, Jonathan J Crofts): arxiv.org/abs/2610.36920
Muktabh Mayank retweeted
UAE President Mohamed bin Zayed Al Nahyan hugs brave Indian pilot Smit Machchhar, crew who thwarted the Fly Dubai terror attack.
Vdo csty: UAE's News Agency WAM
Muktabh Mayank retweeted
(1/n) We’re sharing Nemotron-Labs-Diffusion-Uni: a unified diffusion language model for multimodal understanding and generation. It achieves SoTA understanding and generation performance by learning from autoregressive VLMs and continuous flow models via On-Policy Distillation
Muktabh Mayank retweeted
Tutorial/Perspective article: "Multi-time-scale Control of Power Systems" (by Emiliano Dall'Anese, John W. Simpson-Porco, Florian Dörfler, Saverio Bolognani, Sairaj Dhople): arxiv.org/abs/2610.36357
Muktabh Mayank retweeted
It's insane that you can now run your own coding agent (Pi) w/ LFM3-1.7B *in the browser* at over 300 tok/s on an M4 Max, using custom WebGPU kernels & 4-bit GGUF weights.
No data sent to a server. All your chats stay private.
Try it yourself on Hugging Face 👇
Muktabh Mayank retweeted
Review article: "The geometry of cognition: from neural populations to neurotechnology" (by Lucas Bayones, Devika Narain, Roman Rossi-Pool): arxiv.org/abs/2610.36541
Muktabh Mayank retweeted
New Diffusion LLMs 👀
huggingface.co/Diffuspace/Cl…
- 35B/A3B
- 128k context
- 680 tok/s on GB300
- Runs on 16GB VRAM (INT4)
- 73.6% SWE bench Verified
- Apache 2.0
Meet Clover-35B-A3B-Preview 🍀 the first open-weight production diffusion language model (dLLM) with 35B total parameters capable of 680 tps on a single card at BF16.
Clover exhibits strong performance against both dLLM and AR models and is available today at @huggingface @ModelScope2022 and at our website
Muktabh Mayank retweeted
Vision foundation models like DINO and SigLip now power much of robot learning.
But they are built for RGB. What about depth?
Introducing DeFM: a foundation model for depth images, built for robotics.
🌐 de-fm.github.io
Weights, code, and demo are out today 🧵
Muktabh Mayank retweeted
"Mathematical Theory of Deep Learning" is a free book that introduces the mathematical foundations of deep learning.
The book focuses on three central areas of neural network theory: 1. approximation theory, which studies how neural networks represent functions; 2. optimization theory, which examines how their parameters are learned; 3. statistical learning theory, which investigates how models generalize beyond their training data.
I think it's a very useful resource for anyone interested in studying deep learning from a mathematical perspective.
arxiv.org/pdf/2407.18384
Muktabh Mayank retweeted
Solving Problems In #Algebra and #Trigonometry – V. Litvinenko, A. Mordkovich
intended for students of physical and mathematical faculties of pedagogical institutes; contains about 2000 examples, #problems, and #exercises
#soviet #mathematics #mir
mirtitles.org/2018/10/12/sol…
Muktabh Mayank retweeted
Lecture notes: "The Disordered Universe 2025: Statistical Systems Out of Equilibrium" (by Weam Abou Hamdan, Martí Berenguer Mimó, Rafael Carrasco Carmona, Silvia Pla, Omkar Shetye, Tarek Anous, Paolo Benincasa, Guilherme Leite Pimentel): arxiv.org/abs/2610.36997
Muktabh Mayank retweeted
After the evidence of fully functional hospitals in 5th century India (as noted by Faxian) now we have a much older evidence of public healthcare in 3rd century India.
Around 1,700 years ago, a woman in India put more than 3,500 karshapanas (an ancient Indian silver punch-marked coin) into craft guilds to fund medicines for sick monks.
The obligation was meant to continue even as the guilds changed.
Her name was Vishnudatta. This old impression of her inscription at Nashik preserves the arrangement.
1,000 karshapanas to one guild. 2,000 to another. 500 to a 3rd, whose name is now missing. The oil-millers held another deposit, but the amount is unreadable. That makes 3,500+ karshapanas in all.
Why give the money to guilds if the purpose was medicines?
She was setting up a perpetual endowment. The guilds held the deposits and owed a continuing return to provide medicines for the sick monks of the monastery on Mount Trirasmi. This was funding for supplies over time, not money for a single purchase.
E. Senart, who edited the record, explains that the obligation continued even if the guilds changed their constitution. It bound the guild as it existed then and as it might exist in the future.
The inscription gives her family too: daughter of Agnivarman, wife of Rebhila, mother of Visvavarma. It calls her a lay devotee. The medicines were for monks "of whatever sect and origin" living in that monastery.
The date is the ninth year of the Abhira king Isvarasena, generally placed in the 3rd century CE. The inscription was in print by 1865. Senart's later edition includes this impression, the reading and a translation.
I have attached that translation with the relevant lines marked. You can see the medicine provision, the deposits and his explanation of the continuing obligation on the same page.
Source: E. Senart, "The Inscriptions in the Caves at Nasik", Epigraphia Indica VIII, inscription 15, pp. 88-89, Plate VII. Publication history: H. Luders, Epigraphia Indica X, Appendix, entry 1137.
Photo 1: Vishnudatta's medicine endowment at Nashik, recorded in the ninth year of King Isvarasena. Historical impression of inscription 15, published in Epigraphia Indica, Vol. VIII, Plate VII.
Photo 2: The published translation records Vishnudatta's guild deposits and the continuing obligation to fund medicines. Epigraphia Indica, Vol. VIII, p. 89. Yellow highlights added.
Muktabh Mayank retweeted
🟦 Vision transformers emergently learn to segment objects, but how? 100 years ago, Gestalt psychologists had the same question about people. In new work with @Mikey_Lepori, we take inspiration from Gestalt psychology to study how ViTs bind image patches into objects.
Muktabh Mayank retweeted
NCP-ArchPreview: toward latent-space language models
An 8.9B model trained on 5.73T tokens that adds Next Concept Prediction, predicting discrete concepts spanning multiple tokens. It matches OLMo-3-7B pretraining loss using only 51.3% of the tokens.
