Krish Malik

Standard ViTs sum across image channels at the embedding layer, an assumption that holds for RGB natural images but breaks down for calorimeter data: the ECAL and HCAL channels here are two physically separate subdetectors, at different radii, built to capture different particle types. Summing them away throws out exactly the structure that makes jet classification possible. DepthViT addresses this by embedding each channel independently and computing attention across channels rather than across spatial patches. The channel-wise attention mechanism is adapted from an unsupervised anomaly-detection architecture (Julson et al., Cerium Laboratories / University of Alabama / CMS HCAL Collaboration) into a fully supervised, five-class jet classifier — a substantial change in kind, not just domain. Only the depth-wise embedding and channel-attention operators are carried over from the original; the classification head, and the mechanism to recover cross-patch spatial communication (Hierarchical Attention Pooling, or HAP), are new to this project.