The work was led by @magnusaross and was co-authored by an amazing team of clinical and machine learning experts (Nel Swanepoel, Akish Luintel, Emma McGuire, Ingemar Cox, Steve Harris).
6/
6/
1
0
0
20
Within this process, we show that forecasting patient physiology offers a principled foundation for decision support in antibiotic stewardship.
5/
5/
1
0
0
4
The latter required a substantial effort from @magnusaross & the team at UCLH and is a milestone for further collaborative research exploration. The proposed machine learning solution greatly optimises the process of selecting patients for a switching review by a clinician.
4/
4/
1
0
0
8
The design yields interpretable outputs, adapts to updated guidelines without retraining, and preserves clinical judgement. We validate models across two data sets, a public one (MIMIC-IV), but also a new one (non-public) that we formed from patients admitted at @uclh.
3/
3/
1
0
0
40
We use neural processes to forecast vital sign (e.g. heart rate, oxygen saturation) trajectories probabilistically, predicting switch-readiness for patients by comparing their projected rates against clinical guidelines, and ranking patients to prioritise clinical review.
2/
2/
1
0
0
42
In our paper, the final version of which was published this week in @NatureComms, we study the clinical task of switching patients from intravenous to oral antibiotic treatment.
Paper: nature.com/articles/s4146…
Code: github.com/SAFEHR-data/iv…
1/
Paper: nature.com/articles/s4146…
Code: github.com/SAFEHR-data/iv…
1/
1
2
5
320
(3) The excellence and research readiness of our industrial partners. @locai_labs, and George Drayson in particular, provided a rigorous, research-led supervision to this project.
cc: @ai_ucl
cc: @ai_ucl
0
0
0
32
(2) The excellence of our Machine Learning MSc students at @uclcs. The work was led by Minoo Kim and was based on his MSc Project. Congratulations Minoo!
/3
/3
1
0
0
48
Reasons to be excited about:
(1) NEEDLE is a method for removing LLM backdoors while retaining model performance & safety. It's not a complete & final solution, but points to a methodological path for more reliable removal algorithms once backdoors have been established.
/2
(1) NEEDLE is a method for removing LLM backdoors while retaining model performance & safety. It's not a complete & final solution, but points to a methodological path for more reliable removal algorithms once backdoors have been established.
/2
1
0
0
37
Excited to share our (w/ Minoo Kim and @g_drayson) new preprint "Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation".
Paper: arxiv.org/abs/2610.00348
Code: github.com/LocaiLabs/NEED…
Models: huggingface.co/collections/lo…
/1
Paper: arxiv.org/abs/2610.00348
Code: github.com/LocaiLabs/NEED…
Models: huggingface.co/collections/lo…
/1
1
1
5
165
Vasileios Lampos retweeted
People are increasingly worried about backdoors in LLMs…
so we developed a method to remove them.
New paper: Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation
Congratulations to Minoo Kim for leading this work and @lampos
so we developed a method to remove them.
New paper: Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation
Congratulations to Minoo Kim for leading this work and @lampos
1
3
5
97
Huge congratulations to my amazing PhD student, Niels D. Cariou Kotlarek, for leading this!
In this work, we address a fundamental gap in generative TPPs: the lack of global sequence-level loss functions and mathematically justified distributional discrepancy measures.
2/
In this work, we address a fundamental gap in generative TPPs: the lack of global sequence-level loss functions and mathematically justified distributional discrepancy measures.
2/
1
2
3
308
Delighted to share that our paper "From Jumps to Signatures: a Generative Method for Temporal Point Processes" has been accepted at #NeurIPS2026.
Preprint: arxiv.org/abs/2607.06652
Code: github.com/Code-Cornelius…
1/
Preprint: arxiv.org/abs/2607.06652
Code: github.com/Code-Cornelius…
1/
1
2
4
172
Vasileios Lampos retweeted
Artificial Intelligence is a critical national infrastructure. Government and industry increasingly depend on foundation models whose security, transparency, and reliability are essential.
/1
/1
3
6
15
2,972