My school CSE @UNSW is hiring 12 positions, including 6 Lecturers/Senior Lecturers (Assistant Professors), and 6 Associate Professors, including in ML and NLP, neurosymbolic AI, systems (for AI), cybersecurity, software engineering, cryptography. Check out UNSW Careers website.
#facultyjobs #hiring
🚀 Introducing our #ICLR2026 paper DRIFT-Net: A Spectral-Coupled Neural Operator for PDE Learning!
Simulating Partial Differential Equations (PDEs) is computationally expensive. While neural operators help, many PDE foundation models struggle with error drift. Current attention-based models often struggle with error accumulation and drift during long rollouts.
DRIFT-Net fixes this using:
Dual-Branch Design: It explicitly combines global spectral structure with local spatial detail, with controlled low-frequency mixing.
🌊 Spectral branch for global, low-frequency dynamics. 🔍 Image branch for local, high-frequency details.
🛡️ Stable training with no width inflation via bandwise fusion: The two branches are effectively coupled at every layer. By fusing these branches via bandwise weighting, it avoids the training instability and width inflation caused by naive concatenation.
The results?
📉 7%-54% lower relative L1 error on Navier-Stokes tasks.
⚡ ~15% fewer parameters & higher throughput than scOT!
Code is available here 👇 🔗 https://t.co/62x7f39TLt
Model & checkpoints on @huggingface : https://t.co/OmCduyS9Ht
Paper: https://t.co/ggLhpcdCGx
Whether you are working with fluid dynamics, wave propagation, or phase separation, this modular design can be swapped into existing multi-scale operator backbones for an immediate upgrade.
This work was done with my Master student Jiayi Li during his 2-term thesis project at UNSW Computer Science and Engineering. Indeed, it was his first paper submission!
We've been exploring ODEs, PDEs, SDEs for a while. If you're going to be in Rio for ICLR, would be great to chat.
#AI #MachineLearning #DeepLearning #PDE #physics #neuraloperators #PINNs #ICLR2026
p.s. posting after a while... Giving this platform another go.
We REALLY REALLY need a "Findings" for NeurIPS, ICLR, and ICML. 25,000 submissions at this year's NeurIPS represents extreme excess pressure. It takes valuable time away from legitimate new research.
One question is how to administer it. I suggest that Findings go through a lightweight round focused on improving clarity, reeling in overselling, etc. An AC can then "sign off". Authors can always decline the opportunity, if they want to try for the next conference.
The NeurIPS's proceedings are called "Advances in Neural Information Processing Systems". We could have "Findings in Neural Information Processing Systems", or to not trample on their brand, perhaps "Contributions to Neural Information Processing Systems".
ICLR could have "Letters on Learning Representations."
ICML could have "Machinations on Machine Learning."
Today witnessed an electrifying talk by the legendary Professor Ben Shneiderman (@benbendc) at UNSW AI Institute! "Generative AI: With Great Power Comes Great Responsibility". Mind officially blown. https://t.co/lNv4ElsmWd #GenerativeAI#AI#UNSWAI - My key takeaways - 1/4
A new stark reality to hit us and the entire planet ecosystem. And it’s only going to get worse unless we do something about it
https://t.co/mYO2c2k8rv
Join this challenge. To find out more you can find us at NeurIPS if you’re attending.
BTS: Building Timeseries Dataset
🎥 https://t.co/cMo0XtxdGW
⚒️ https://t.co/CAlMjgceCW
📍 Fri 13 Dec 11-2, West Ballroom A-D #5106
1/ 💡🏢 New Challenge Alert: Join the Brick by Brick: Automating Building Data Classification Challenge and tackle one of the biggest hurdles in sustainable building tech - automated building data classification. 📷 Join now: https://t.co/KoNiLCzskh
1/ 💡🏢 New Challenge Alert: Join the Brick by Brick: Automating Building Data Classification Challenge and tackle one of the biggest hurdles in sustainable building tech - automated building data classification. 📷 Join now: https://t.co/KoNiLCzskh
Come and find us. I have upcoming openings in my group (postdocs and PhDs). Keen to reconnect, discuss potential new collaborations, and meet new friends. Our school @UNSWCOMPUTING is advertising for 7 Associate Professor positions. Feel free to reach out.
@unsw_ai#NeurIPS2024
Resolution-Agnostic Transformer-based Climate Downscaling
🎥 https://t.co/Pj9CrLXUQz
📍Tackling Climate Change workshop
🔗https://t.co/xOYwSM4IHv
We introduced a cost-efficient downscaling method using a pretrained Earth-ViT model, to generalize across different resolutions
Threads or BlueSky or something else?
I’m a newbie on both:
@ https://t.co/TNiQ3Jr0TZ on BlueSky.
@ fsalim on Threads.
BlueSky seems more civilized with more paper discussions, but mostly NLPs.
Threads seems more casual.
I need help finding folks on both platforms. Find me there?