Refine Neutrino Events Reconstruction with BEiT-3

TL;DR

This paper introduces ISeeCube, a Transformer-based model leveraging TorchScale for neutrino event reconstruction, outperforming previous solutions with simpler implementation and new evaluation metrics, and demonstrating potential for broader applications.

Contribution

The paper presents a novel Transformer model, ISeeCube, based on TorchScale, achieving superior performance in neutrino event reconstruction and simplifying model development and testing.

Findings

Outperforms second-place solutions in Kaggle competition.

Reduces code complexity by 80% using TorchScale.

Introduces overlap ratio as a new performance metric.

Abstract

Neutrino Events Reconstruction has always been crucial for IceCube Neutrino Observatory. In the Kaggle competition "IceCube -- Neutrinos in Deep Ice", many solutions use Transformer. We present ISeeCube, a pure Transformer model based on TorchScale (the backbone of BEiT-3). When having relatively same amount of total trainable parameters, our model outperforms the 2nd place solution. By using TorchScale, the lines of code drop sharply by about 80% and a lot of new methods can be tested by simply adjusting configs. We compared two fundamental models for predictions on a continuous space, regression and classification, trained with MSE Loss and CE Loss respectively. We also propose a new metric, overlap ratio, to evaluate the performance of the model. Since the model is simple enough, it has the potential to be used for more purposes such as energy reconstruction, and many new methods such as combining it with GraphNeT can be tested more easily. The code and pretrained models are available at https://github.com/chenlinear/ISeeCube