I really excited to say that the #PyTorch wrapper for #MachineLearning in #HighEnergyPhysics that I've been developing is now in beta and available on PIP!.Feedback, suggestions, and contributions are most welcome:

LLMs may be all the rage these days, but the key question on my mind is "can they help build Magic: The Gathering decks?". Let's dive in and take a look!. Code-base:

RT @MilesCranmer: SymbolicRegression.jl → 1.0 🎉 . After several years of work, I'm thrilled to announce some major new features! Let me sho….

Heavy rain in Tokyo last night, and my apartment got flooded. Water level outside reached 56cm!.Luckily most went down the drain before it could come in, but it still turned my sitting room into a puddle.

Very happy to finally have this work published!.Summary walkthrough here

RT @ATLASexperiment: Explore over 7 billion LHC collision events from home💻. The ATLAS Experiment at @CERN has made public two years' worth….

RT @dhpmrou: 10 years ago #OTD, @bingxu_ and @tqchenml quietly announced the birth of @XGBoostProject on the @kaggle #higgsml competition f….

Tagging @dorigo @pietrovischia @JanKieseler.

Delighted to say that the code for our paper on differential optimisation for muon tomography is now public!.Preprint (soon to be published) NeurIPS Poster @MODECollaborat1.

Been seeing a worrying number of DL libraries that gloss over param init; hardcoding some arbitrary scheme that doesn’t account for activation functions. If addressed in docs, correct init is treated as a special case, with little support. Do people just not bother these days?.

First session just finished for our poster for the Machine Learning and the Physical Sciences Workshop at NeurIPS 2023. Extended abstract: Full pre-print:

Despite the plan being shelved, it seems there’s still local support for a Japanese hosting of the International Linear Collider.

This was 2.5 years of work from a great bunch of people. Many thanks to all who helped with this!.

Key point in the optimisation process is the use of uncertainty-based weighting in the inference: all muons contribute to an average, but some are better measured and so contribute more. This involves the use of autograd in the forwards pass to propagate uncertainties.

By construction, the final detector costs the same as the initial one, but provides a much more precise measurement of the fill-height

Here it is in action: we begin the detector panels (coloured squares) from a sub-optimal position and overtime, they converge to a better position that carefully balances muon exposure and reconstruction precision, whilst learning a non-trivial assignment of available budget.

Using PyTorch as a backend, we can then compute derivatives of the loss wrt the learnable parameters (position and budget assignment) to update the detector in the direction that best improves our measurement.

On top of a differentiable model of the detector panels, and simulators for the muon generation and scattering, we use a task-specific inference and loss function to predict the fill height and quantify performance.

As a benchmark, we take the problem of estimating the fill-level of ladles containing liquid steel and slag at a metal refinery. Using cosmic muons, as a probe the fill-heights can be predicted from muon scattering.

In the context of muon-tomography, we develop a flexible system to learn detector layouts for specific tasks, by optimising directly for the measurements they are intended to be used to make: no proxy objectives or assumptions.

Very happy to finally have this out Demonstration of differentiable end-to-end, measurement-aware optimisation of particle detectors.