TL;DR
This paper introduces sMask-RCNN, a modified Mask R-CNN network optimized for speed on CPUs by using sparse convolutions. It effectively clusters cosmic ray muons in the MicroBooNE liquid argon time projection chamber, significantly improving the ability to reject background events and enhance the purity of electron neutrino signals.
What Problem Does This Paper Address?
The MicroBooNE experiment, designed to study neutrinos, operates a large liquid argon time projection chamber (LArTPC) without significant overhead shielding. This leads to a high rate of cosmic ray muons interacting within the detector, which can be mistaken for neutrino interactions. Effectively distinguishing between these signals is crucial for accurate neutrino physics measurements. Traditional methods and even standard deep learning approaches can be computationally intensive, posing a challenge for processing the vast amount of data generated by MicroBooNE, especially when relying on CPU farms rather than GPUs.
Methods
The core of this work is the development of sMask-RCNN, a sparse implementation of the Mask R-CNN object detection and segmentation network. The key innovation lies in replacing the standard convolutional layers in the ResNet backbone with sparse submanifold convolutions. This technique significantly speeds up processing on sparse data (like the MicroBooNE wire readout images, where only a small fraction of pixels have signal) by avoiding computations on zero-valued pixels. This optimization is crucial for enabling the network to run efficiently on CPUs, which are more readily available in large-scale computing facilities like FermiGrid. The sMask-RCNN network takes 2D wire readout images as input and outputs individually labeled particle interactions, classifying them as either cosmic ray muons or electron neutrinos.
Datasets
The study utilizes:
- Simulated MicroBooNE data: This includes simulated electron neutrino interactions and simulated cosmic ray muons (generated using CORSIKA).
- Real MicroBooNE data: This includes off-beam data (cosmic ray-only background) and on-beam data (containing both neutrinos and cosmic rays).
The training data for the dense Mask-RCNN included various interaction classes, while the sparse Mask-RCNN was trained specifically on cosmic ray muons and electron neutrino interactions. Performance evaluations were conducted on a validation set of 9400 full-sized images.
Key Insights and Ideas
- Sparse convolutions are key for CPU-based deep learning: The paper strongly demonstrates that sparse submanifold convolutions are essential for making complex deep learning models like Mask R-CNN computationally feasible on CPUs for large-scale physics experiments. This opens up possibilities for deploying advanced AI techniques where GPU resources are limited.
- Cosmic ray identification is a critical first step: The effectiveness of sMask-RCNN in identifying and vetoing cosmic ray-only events highlights the importance of robust background rejection early in the reconstruction chain. This significantly improves the purity of the desired neutrino signal.
- "Identification by negative" is more effective for background veto: The study found that using sMask-RCNN to identify and remove cosmic ray events (an "identification by negative" approach) is more effective than trying to directly identify neutrino events ("identification by positive"). This is attributed to the more consistent topology of cosmic ray muons compared to the diverse final states of electron neutrino interactions.
- Combining sMask-RCNN with existing tools yields state-of-the-art results: Integrating sMask-RCNN's cosmic tagging capabilities with the existing Wire-Cell cosmic tagger significantly enhances background rejection, demonstrating the synergistic benefits of combining different techniques.
Key Results
- Speedup: sMask-RCNN achieves a 20x speedup in ResNet processing time on a CPU compared to the dense Mask-RCNN.
- Pixel Clustering Efficiency: sMask-RCNN has an average pixel clustering efficiency of 85.9%, slightly lower than the dense network's 89.1%.
- Cosmic-Only Event Veto: When combined with the Wire-Cell cosmic tagger, sMask-RCNN can remove 70% of the remaining cosmic ray muon background events at the same electron neutrino signal efficiency.
- Improved Signal-to-Background Ratio: This event veto strategy can provide 99.7% rejection of cosmic ray-only background events while maintaining an electron neutrino event-level signal efficiency of 80.1%. This improves the signal-to-background ratio from 1.26 to 4.14 for general electron neutrino events.
Limitations and Future Work
- Dense vs. Sparse Performance: While sMask-RCNN offers significant speed advantages, the dense implementation generally shows slightly better performance in terms of efficiency and purity metrics. The paper notes that it's difficult to definitively attribute this difference solely to the sparse convolutions versus other training differences (e.g., training on crops vs. full images).
- Unresponsive Wire Regions: The network struggles with interactions that fall in regions of unresponsive wires in the LArTPC, leading to zero efficiency. This is an inherent limitation of the detector geometry and data acquisition.
- Small/Low-Charge Interactions: Visually small interactions or those with less deposited charge can also lead to lower efficiency, as they are harder for the network to detect and cluster.
- Future Work: The authors suggest that sMask-RCNN could be extended to identify other particle interaction types in 3D-tracking detectors. Further improvements could involve increasing the required confidence score for neutrino classification or retraining the network with adjusted penalties. The study also highlights the potential for further optimization by applying sparse convolutions to later stages of the network.
Summary updated