Anna Sergeevna Bosman

We have reviewed over 100 papers on regularisation, and conducted experiments to show: regularisation doesn’t always help! 💯 human-made review of regularisation in NNs available here: ieeexplore.ieee.org/abstract/docum… #neuralnetworks #regularisation #ieeetransactions

@yule_gan Isn’t this just another manifestation of mode connectivity, i.e. multiple minima on the same manifold?

Simply adding Gaussian noise to LLMs (one step—no iterations, no learning rate, no gradients) and ensembling them can achieve performance comparable to or even better than standard GRPO/PPO on math reasoning, coding, writing, and chemistry tasks. We call this algorithm RandOpt. To verify that this is not limited to specific models, we tested it on Qwen, Llama, OLMo3, and VLMs. What's behind this? We find that in the Gaussian search neighborhood around pretrained LLMs, diverse task experts are densely distributed — a regime we term Neural Thickets. Paper: arxiv.org/pdf/2603.12228 Code: github.com/sunrainyg/Rand… Website: thickets.mit.edu

Drumroll for the IJCNN decision letters, due this week 🥁 My research group submitted three papers on three very distinct topics. I have also acted as both a reviewer and an AC, and I confirm that reviews were stern and demanding! Hoping for the signal to outweigh the noise 🤞🤞

Thrilled to announce the 3rd African Computer Vision Summer School (ACVSS) 👉acvss.ai 🎓Learn computer vision from the best scientists 🌍Google, Accra, Ghana 📆July 19-29, 2026 🎓Apply before March 1st (full grants available) ❤️ Follow us #ACVSS #computervision #ai

Discrete Diffusion just got a huge upgrade! with "Categorical Flow Maps", it is now much faster + capable of test time inference. This paper shows you can make diffusion/flow models for discrete stuff (text, graphs, binary images) generate in 1–2 steps by flowing continuously on the probability simplex and self-distilling an endpoint-consistent “flow map”. This keeps predictions valid (on-simplex) and unlocks the same fast guidance tricks we use in continuous diffusion. So few-step speedups to categorical generation without the usual discrete-step pain!

I knew it!

If you want to do diffusion on discrete data, you have three choices: discrete, Gaussian, or simplicial. How are they related? Which should you use? We theoretically unify all three and train one model to do them all! @AlinaChandra @yucenlily @alex4ali @andrewgwils 1/7

Recursion, I have always believed in you 🤟

New paper on covariant #neuromorphic networks! We're connecting decades of work in computer vision with decades of work in spiking networks to present spatio-temporal spiking processing. nature.com/articles/s4146…

This blog page is the best place to learn everything about Diffusion. Like the very essence of it!!

UP is ranked No. 1 in South Africa for artificial intelligence research, No. 2 in Africa, and within the top 600 globally. Learn how UP is shaping ethical, African-led AI for real-world impact. Read more: ow.ly/xLjC50WFhpT #UniversityOfPretoria #ProudlyUP #AI

Does Adam oscillate? The answer might surprise you! Check out our new paper "Visualising Adam Oscillations in Neural Network Loss Landscapes": dl.acm.org/doi/10.1145/37… Main takeaway: if you're pruning weights early, look at their magnitude - not their gradients. Gradients deceive!

“Can Physics-Informed Neural Networks (PINNs) simulate 3D turbulence?” A question we've been asked repeatedly since developing the framework in 2017. After nearly a decade of progress, we now have a conclusive answer. Thread below 🧵

It is extremely fulfilling to be a part of such beautiful, meticulous, relevant applied projects — ML used for good rather than evil. 🌿☀️🪴 sciencedirect.com/science/articl…

1/10 ML can solve PDEs – but precision🔬is still a challenge. Towards high-precision methods for scientific problems, we introduce BWLer 🎳, a new architecture for physics-informed learning achieving (near-)machine-precision (up to 10⁻¹² RMSE) on benchmark PDEs. 🧵How it works:

📢 Counting down to the Deep Learning Indaba𝕏🇿🇦2025! 🚀Workshop: Understanding AI Safety – Why It Matters. 💡Participate in dynamic discussions and interactive group activities. 🗓️Date: Wednesday, 9 July 2025 ⏰Time: 13:30 - 16:00 📍 Venue: Stellenbosch University #Indaba𝕏SA

1/ AI is accelerating. But can we ensure that AIs truly share our values and follow our goals? We argue that aligning advanced AI systems requires cracking a core scientific challenge: how data shapes AI's internal structure, and how that structure determines behavior.

Heya science twitter! I am looking for an excellent CS student to do fully-funded MSc research starting 2026 on foundation models for radio astronomy. Does it even get any geekier🤓?! AI +✨=😍 sarao.ac.za/researchandstu… Interested? Email me. Please share far and wide!🪐👩‍🚀💫

Spread the Word! Free student applications for #IndabaXSA2025 are OPEN—spaces are limited. Encourage friends to apply and join us at Stellenbosch University, July 7-11, for 70 years of AI innovation. Repost and DM to grow the community! indabax.co.za/register

Student applications are OPEN for the Indaba𝕏🇿🇦'25 (it's FREE!). Academic and industry professionals, registration will open on April 30th Join us at Stellenbosch University from July 7-11 as we celebrate 70 years of AI. Apply now: indabax.co.za/register/ #IndabaX2025