Published in Empirical Methods in Natural Language Processing (EMNLP) (Short), 2024
We quantify performance gaps between training on captions that come from native German perception and captions that have been either machine-translated or human-translated from English into German. To address these gaps, we further propose and evaluate caption augmentation strategies.
Published in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024
We analyze how context in class text representations of VL models affects geographical robustness in object recognition. We propose to ensure geographical robustness in soft prompting through a new geography knowledge regularization method.
Published in IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024
We analyze if attributes in captions that supervise VL models result in useful representations for fine-grained tasks. We find models have limited fine-grained utility, so we propose a negative sampling strategy in contrastive learning to improve attribute sensitivity.
Published in AAAI Workshop on Practical Deep Learning in the Wild (Practical-DL), 2023
We propose strategies to improve the out-of-domain robustness of contrastively pretrained object detectors. In particular, we augment contrastive views to learn representations that are more robust to appearance-based and out-of-context domain shifts.
Heartbeat Classification on the Loihi Neuromorphic Processor
Published in IEEE Computer Society Annual Symposium on VLSI (ISVLSI), 2021
We conduct a heartbeat classification case study on the Loihi neuromorphic processor, investigating accuracy and latency optimization strategies in artificial-to-spiking neural network conversion. We present benchmarking insights with regards to spiking 1D-CNN versus artificial 1D-CNN performance on state-of-the-art neural network accelerators.
Beyond Floating-Point Ops: CNN Performance Prediction with Critical Datapath Length
Published in IEEE High-Performance Extreme Computing Conference (HPEC), 2020
We investigate how neural network metrics are related to performance on CPU and GPU hardware.
Story Understanding in Video Advertisements
Published in British Machine Vision Conference (BMVC), 2018
We make progress in how to understand the dynamic structure of video advertisements. In particular, we present approaches to effectively predict climax and sentiment in ads.