About Me

Edwin Arkel Rios
PhD Candidate in EECS
National Yang Ming Chiao Tung University
Hsinchu, Taiwan

edwinarkel.rios@gmail.com
[GitHub] [Resume]

My name's Edwin and I come from Panama. I received the B.S. degree in Energy Engineering from National Cheng Kung University (NCKU), Tainan, Taiwan, in 2019. I enrolled at National Yang Ming Chiao Tung University (NYCU) to pursue an M.S. and then a Ph.D. degree in Electrical Engineering and Computer Science (EECS), in 2019 and 2021, respectively, where I have been co-supervised by Professor Bo-Cheng Lai (賴伯承) and Professor Min-Chun Hu (胡敏君).

My research is focused on the design of efficient deep learning models. In particular, I have done work on time-series for heart-rate monitoring, model compression using knowledge distillation and self-supervised learning, and efficient fine-grained image recognition systems incorporating parameter-efficient transfer learning, token pruning for vision transformers, and efficient discriminative feature selection mechanisms. More recently I have worked on low-bit quantization of speech and audio language models.

Publications

How to Choose Your Teacher for Fine Grained Image Recognition. CVPR FGVC Workshop 26
On the Role of the Teacher in Knowledge Distillation for Fine Grained Image Recognition. CVGIP 25
Systematic study of teacher selection for knowledge distillation on fine-grained datasets, where the usual rule of picking the most accurate teacher breaks down.
Experiments across teacher-student pairs, datasets and distillation losses show which teacher properties predict student accuracy.
[Paper]
A Large-Scale Study on the Accuracy vs Cost Trade-offs of Training and Evaluation Settings in Fine-Grained Image Recognition. CVPR FGVC Workshop 26
Over 2,000 experiments across 16 backbones, 17 datasets and 6 training and evaluation settings to measure what FGIR-specific settings buy in accuracy and what they cost in training and inference time.
Results distilled into a decision chart for choosing a backbone and setting under a given system constraint; 825 pretrained checkpoints released.
[Source code] [Paper]
An Empirical Analysis of Weights Post-Training Quantization for Hyperspectral Image Classification. CVGIP 26 (Merit Paper Award)
First systematic post-training quantization benchmark for Transformer- and Mamba-based hyperspectral image classifiers: 2,820 runs over bit widths, group sizes, quantization libraries, label budgets and datasets.
4-bit weights are near-lossless and the accuracy-cost sweet spot; Mamba models are the most sensitive below that, and Martian scenes are less sensitive than Earth ones.
Switchable-Precision Universally Slimmable Networks. IEEE SPL 26
Revisiting the Backbone, Pretraining and Transferability for Hierarchical Fine-Grained Image Recognition. CVGIP 25 (Excellent Paper Award)
A single network that runs at any width and at several weight precisions (SPL), and a study of how backbone and pretraining choices transfer to hierarchical fine-grained recognition (CVGIP, equal contribution).
Cross-Layer Cache Aggregation for Token Reduction in Ultra-Fine-Grained Image Recognition. ICASSP 25
Plug-and-play modules to avoid information loss when applying token pruning methods for datasets with small inter-class differences.
Thorough experiments across diverse datasets, backbones, token reduction methods, image sizes, and keep rates show our method enables significant cost savings while maintaining high accuracy.
[Source code] [Paper]
Down-Sampling Inter-Layer Adapter for Parameter and Computation Efficient Ultra-Fine-Grained Image Recognition. ECCV EFM Workshop 24
Addresses attention collapse under wide domain gap adaptation settings by proposing a down-sampling adapter module.
Similar accuracy but 123x fewer trainable parameters compared to SOTA.
[Source code] [Paper]
Global-Local Similarity for Efficient Fine-Grained Image Recognition with Vision Transformers. ISCAS 25
Similarity between global and local representations as the criterion for discriminative region selection in FGIR.
Thorough experiments across 10 FGIR datasets covering a variety of settings including image size and AugReg combinations.
Our proposed model obtains the best accuracy on 8 of 10 datasets at much lower cost than the best on the others.
[Source code] [Paper (arXiv)]
Anime Character Recognition using Intermediate Features Aggregation. ISCAS 22
Intermediate features aggregation classification head for reducing sensitivity to hyperparameters in ViTs at low computational cost.
Experiments on how model architecture and hyperparameters affect classification performance.
[Source code] [Paper]
IFACD: Intermediate Features Augmented Contrastive Distillation. ICLR CSS Workshop 22
Extend contrastive learning framework (in particular SimCLR-like) to support features from intermediate layers of a particular image as additional positive pairs and intermediate features from other images in the batch as negative pairs, reducing reliance on augmentation pairs and large batch size requirements.
Apply proposed contrastive loss in a knowledge distillation framework where intermediate features are extracted using a frozen teacher and the student learns using a combined contrastive, distillation, and supervised cross-entropy loss.
Experiments with a variety of teacher-student combinations show our method obtains favorable accuracy at a competitive computational cost compared to alternatives.
[Source code] [Recorded Presentation (Starts at 2:47:50)] [Slides]
DLPrPPG: Development and Design of Deep Learning Platform for Remote Photoplethysmography. ISCAS 22
Parametric Study of Performance of Remote Photopletysmography System. ISCAS 21
Systematic studies of remote photoplethysmography (RPPG) design hyperparameters (face detection method, region of interest selection scheme, sliding window size, preprocessing and postprocessing algorithm choice) and their effects on heart rate prediction accuracy in relation to computational cost using C++, Python and OpenCV.
[Paper ISCAS 2022] [Paper ISCAS 2021]

Projects

Vision Transformers for Anime Character Face Recognition
Image classification of anime character faces using the state of the art for image recognition: vision transformers.
Revamped a dataset for this task, DAF:re: almost 500K images and more than 3K classes.
Conducted experiments on both CNN and self-attention based models, and explored for the first time the hyperparameter tuning (mini-batch and image size) of vision transformers for out-of-domain data.
[Source code] [Paper on DAF:re]
DeblurGAN for Deblurring of Videos
Done in Python, using OpenCV and PyTorch.
Final project for Digital Image Processing course (108-1).
Also included slides for midterm presentation covering the DeblurGANv2 paper.
[Final Project Poster] [DeblurGANv2 Paper Slides]
Computer Vision for Autonomous Driving
Done in Python, making use of TensorFlow, Keras, OpenCV, and ROS.
Final project for Deep Learning course (108-1).
Also included slides for midterm presentation covering the VGG network paper.
[VGG Paper Slides] [Final Project Slides]