Exploring End-to-end Differentiable Neural Charged Particle Tracking – A Loss Landscape Perspective
Authors: Tobias Kortus, Ralf Keidel, Nicolas R. Gauger
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate empirically that including differentiable variations of discrete assignment operations allows for efficient network optimization, working better or on par with approaches that lack E2E differentiability. In additional studies, we dive deeper into the optimization process and provide further insights from a loss landscape perspective, providing a robust foundation for future work. We demonstrate that while both methods converge into similar performing, globally well-connected regions, they suffer under substantial predictive instability across initialization and optimization methods... 5 Experimental Results and Analysis: Dataset: For the studies reported in this work, we rely on MC simulation of detector readout data... |
| Researcher Affiliation | Academia | Tobias Kortus EMAIL Chair for Scientific Computing University of Kaiserslautern-Landau (RPTU); Ralf Keidel EMAIL Chair for Scientific Computing University of Kaiserslautern-Landau (RPTU); Nicolas R. Gauger EMAIL Chair for Scientific Computing University of Kaiserslautern-Landau (RPTU) |
| Pseudocode | No | The paper describes the methodology using mathematical formulations (e.g., Eq. 1-15) and schematic diagrams (Figure 1-3) but does not contain a dedicated 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | No | All source code that supports the findings of this study will be openly available open source after paper acceptance under https: //github.com/SIVERT-p CT/e2e-tracking. |
| Open Datasets | Yes | For comparability purposes, the test set, containing 10,000 particles in total, is taken from Kortus et al. (2022). A detailed listing of all data sources is provided in Appendix F. ... Tobias Kortus, Alexander Schilling, Ralf Keidel, Nicolas R Gauger, and on behalf of the Bergen p CT collaboration. Particle Tracking Data: Bergen DTC Prototype, dec 2022. URL https://doi.org/10. 5281/zenodo.7426388. |
| Dataset Splits | Yes | We generate different datasets for training (100,000 particles) and validation (5,000 particles)... For comparability purposes, the test set, containing 10,000 particles in total... Further, we generate hit graphs for the test set with 50, 100 and 150 p+/F to assess the generalization ability to unseen densities. |
| Hardware Specification | Yes | All results are generated on an NVIDIA A100 GPU for all readout frames in the test set, with 10 evaluations for each readout frame. |
| Software Dependencies | Yes | We rely on MC simulation of detector readout data, which we generate using GATE (Jan et al., 2004; 2011) version 9.2 based on Geant4 (version 11.0.0) (Agostinelli et al., 2003; Allison et al., 2006; 2016). |
| Experiment Setup | Yes | Hyperparameter settings: We share model hyperparameters for both PAT and PTT framework, documented in detail in Appendix B. As we consider the interpolation parameter λ as an important tunable parameter... we select values of 25, 50 and 75... Table 4: Selected hyperparameters used in the studies in Section 5. ... dhidden 16 Hidden size of the network layer scaling 0.001 ... batch size 32 ... lr 1e-3 Learning rate ... λ {25, 50, 75} Interpolation factor... |