Unpaired Image Translation to Mitigate Domain Shift in Liquid Argon Time Projection Chamber Detector Responses

TL;DR

The feasibility of using an alternative way to solve the domain shift problem that is not specific to any downstream algorithm is explored and several popular UI2I translation algorithms are adapted to work on scientific data and demonstrated the viability of these techniques for solving the domain shift problem with LArTPC detector data.

Abstract

Deep learning algorithms often are trained and deployed on different datasets. Any systematic difference between the training and a test dataset may degrade the algorithm performance--what is known as the domain shift problem. This issue is prevalent in many scientific domains where algorithms are trained on simulated data but applied to real-world datasets. Typically, the domain shift problem is solved through various domain adaptation methods. However, these methods are often tailored for a specific downstream task and may not easily generalize to different tasks. This work explores the feasibility of using an alternative way to solve the domain shift problem that is not specific to any downstream algorithm. The proposed approach relies on modern Unpaired Image-to-Image translation techniques, designed to find translations between different image domains in a fully unsupervised fashion. In this study, the approach is applied to a domain shift problem commonly encountered in Liquid Argon Time Projection Chamber (LArTPC) detector research when seeking a way to translate samples between two differently distributed detector datasets deterministically. This translation allows for mapping real-world data into the simulated data domain where the downstream algorithms can be run with much less domain-shift-related degradation. Conversely, using the translation from the simulated data in a real-world domain can increase the realism of the simulated dataset and reduce the magnitude of any systematic uncertainties. We adapted several UI2I translation algorithms to work on scientific data and demonstrated the viability of these techniques for solving the domain shift problem with LArTPC detector data. To facilitate further development of domain adaptation techniques for scientific datasets, the "Simple Liquid-Argon Track Samples" dataset used in this study also is published.

Figures (9)

• Figure 1: Learning to translate without pairing. An unpaired translation problem features two domains with samples that are not paired, e.g., cats and dogs. For an input image from the source domain, a neural translation algorithm needs to produce translations resembling samples in the target domain. In the meantime, the translations must retain certain consistency with their input. The first row demonstrates that a deep neural network model can be trained to translate cats into lifelike yet nonexistent dogs while maintaining features such as fur color patterns and facial orientations. Our work investigates if UI2I translation can be adapted to translate between two domains of LArTPC images.
• Figure 2: Signal formation in a three-wire plane LArTPC. An illustration from MicroBooNE:2016pwy. LArTPC detectors enclose a volume of liquid argon. Energetic charged particles ionize electrons from nearby argon atoms as they pass through the volume. An external electric field causes the electrons to drift toward the detector's readout. The readout consists of three parallel planes of sense wires. Each wire plane generates one tomographic view of the tracks. The 3D particle tracks can then be reconstructed by combining the three tomographic views.
• Figure 3: Response functions and ADC waveforms. The ionization electron distribution (Panel A) is convolved with two types of response functions to produce the SLATS dataset's two domains. The 2D response (Panel B top) is used to produce domain $B$ samples, while the quasi-1D response (Panel B bottom) is used to create domain $A$ samples. The quasi-1D data are constructed by masking the 2D response so all contributions from neighboring wires are removed. Panel C shows examples of the ADC waveforms used as input to the translation algorithm.
• Figure 4: Preprocessing for the SLATS dataset. Panel A features an example of a full ADC waveform of the U plane from domain $B$ (generated with a 2D response). The full image has dimension $(\text{channel}, \text{time}) = (800, 6000)$. The portion bounded by the red box is the center crop of dimension $(768, 5888)$. The center crop is divided into $3\times23$ tiles of size $(256, 256)$ and shown as the gray grid. Panels B1 and B2 show a pair of tiles in the test dataset from the domain $A$ (generated with a quasi-1D response) and the domain $B$ (generated with a 2D response), respectively. The tile in B2 corresponds to the highlighted tile in Panel A. The distribution of the number of nonzero pixels in the tiles is shown in Panel C. Tiles with less than 200 nonzero pixels are discarded from the SLATS dataset.
• Figure 5: Summary of the CycleGAN zhu2017unpaired model. CycleGAN consists of two pairs of GANs, $(\mathcal{G}_{A \to B}, \mathcal{D}_{B})$ and $(\mathcal{G}_{B \to A}, \mathcal{D}_{A})$. The discriminators, $\mathcal{D}_{A}$ and $\mathcal{D}_{B}$, distinguish translations from real images, while the generators (or translators), $\mathcal{G}_{A \to B}$ and $\mathcal{G}_{B \to A}$, produce realistic translations that are consistent with the source images.