Publications

International Publication Award (IPA) Recipients

Awarded by the UP System, the International Publication Award (IPA) for journal articles is given to faculty members, researchers, project researchers, and thesis students who publish in international peer-reviewed (Thomson Reuters-listed or SCOPUS-listed) journals. The IPA for books/book chapters is given to faculty members, REPS, and professors emeriti whose books or book chapters are published by international academic publishers and university presses.

IPA 2023

August 2023

Title: Performance Analysis of Multistage Cross-Coupled Differential-Drive Rectifiers using Simulations on 65nm CMOS Process

Authors: Herlan Kester Benitez, Maria Theresa De Leon, John Richard Hizon, Marc Rosales

Abstract: In this paper, the effect of varying transistor widths, rectifier stages, and threshold voltage on the performance of cross-coupled differential-drive rectifiers using a 65nm CMOS process are observed through simulations with a resistive load. Improving conversion efficiency has mostly been the focus for radio frequency (RF) rectifiers used in RF energy harvesting (RFEH) where the input power is expected to be small. Input impedance is also another concern since the matching network between the antenna and rectifier may introduce power loss due to reflections. However, it has been difficult to accurately model and predict the performance and characteristics of these rectifiers because of their non-linearity. Instead of relying on calculations, data from simulations is used to summarize the effects of various design decisions and come up with a suitable design given target specifications. Results show that the same output specifications can be achieved with less rectifier stages by using higher threshold transistors. This design achieves approximately the same conversion efficiency at the same input power level with a higher input resistance which helps increase the voltage gain for impedance matching networks for the rectifier.

Link

Title: Depth Pruning with Auxiliary Networks for Tinyml

Authors: Josen Daniel De Leon, Rowel Atienza

Abstract: Pruning is a neural network optimization technique that sacrifices accuracy in exchange for lower computational requirements. Pruning has been useful when working with extremely constrained environments in tinyML. Unfortunately, special hardware requirements and limited study on its effectiveness on already compact models prevent its wider adoption. Depth pruning is a form of pruning that requires no specialized hardware but suffers from a large accuracy falloff. To improve this, we propose a modification that utilizes a highly efficient auxiliary network as an effective interpreter of intermediate feature maps. Our results show a parameter reduction of 93% on the MLPerfTiny Visual Wakewords (VWW) task and 28% on the Keyword Spotting (KWS) task with accuracy cost of 0.65% and 1.06% respectively. When evaluated on a Cortex-M0 microcontroller, our proposed method reduces the VWW model size by 4.7x and latency by 1.6x while counter intuitively gaining 1% accuracy. KWS model size on Cortex-M0 was also reduced by 1.2x and latency by 1.2x at the cost of 2.21% accuracy.

Link

Title: Improving Model Generalization by Agreement of Learned Representations from Data Augmentation

Authors: Herlan Kester Benitez, Maria Theresa De Leon, John Richard Hizon, Marc Rosales

Abstract: Data augmentation reduces the generalization error by forcing a model to learn invariant representations given different transformations of the input image. In computer vision, on top of the standard image processing functions, data augmentation techniques based on regional dropout such as CutOut, MixUp, and CutMix and policy-based selection such as AutoAugment demonstrated state-of-the-art (SOTA) results. With an increasing number of data augmentation algorithms being proposed, the focus is always on optimizing the input-output mapping while not realizing that there might be an untapped value in the transformed images with the same label. We hypothesize that by forcing the representations of two transformations to agree, we can further reduce the model generalization error. We call our proposed method Agreement Maximization or simply AgMax. With this simple constraint applied during training, empirical results show that data augmentation algorithms can further improve the classification accuracy of ResNet50 on ImageNet by up to 1.5%, WideResNet40-2 on CIFAR10 by up to 0.7%, WideResNet40-2 on CIFAR100 by up to 1.6%, and LeNet5 on Speech Commands Dataset by up to 1.4%. Experimental results further show that unlike other regularization terms such as label smoothing, AgMax can take advantage of the data augmentation to consistently improve model generalization by a significant margin. On downstream tasks such as object detection and segmentation on PascalVOC and COCO, AgMax pre-trained models outperforms other data augmentation methods by as much as 1.0mAP (box) and 0.5mAP (mask). Code is available at https://github.com/roatienza/agmax.

Link

Title: Finding congestion modes of future grids with diverse power exchanges

Authors: Adonis Tio, David Hill, Jin Ma

Abstract: Future power exchange scenarios will become more diverse as more renewable generation, energy storage, and flex-ible loads get integrated into market-dispatched power grids. Transmission network expansion planning (TNEP) practices must be updated to consider the resulting diversity of power in-jection scenarios in the future. In previous work, we quantified grid adequacy to host diverse power exchanges by characterizing the power flow infeasible set. We did this in the context of TNEP to help in the preliminary assessment of existing grids and in the identification of network expansion interventions that help improve grid adequacy. In this work, we present another approach to characterize the infeasible set using what we call grid congestion modes. We visualize how these modes arise from the constrained DC power flow problem and present a method to find them in small- to medium-sized power grids. Case studies using a six-bus test system and a reduced approximate model of the Philippine Luzon grid illustrate the benefits of this approach in TNEP: (1) give insights on grid bottlenecks when power ex-change scenarios become more diverse and (2) identify power exchange scenarios that trigger different congestion patterns that can be used in TNEP.

Link

July 2023

Title: Object Detection in Aerial Images with Attention-based Regression Loss

Authors: Chandler Timm Doloriel, Rhandley Cajote

Abstract: Object detection is a computer vision technique used to identify objects that are usually present in natural scenes. However, the methods used for this case are not easily transferable to detect objects in aerial images. Objects in aerial images are mostly arbitrary-oriented, small, and in complex backgrounds compared to upright and well-focused objects in natural scenes. To effectively detect objects in aerial images, we propose a new regression loss function based on the attention mechanism through attention weights. Using the relative position of the attention weights to the bounding box, the foreground is given more attention, which highlights the target object and effectively suppresses the noise and background. Preliminary experiments are conducted on an attention-based object detector using the DOTA dataset to test the capability of attention mechanism in extracting the contextual information of objects, especially in complex environments.

Link

Title: Improving the detection of small oriented objects in aerial images

Authors: Chandler Timm Doloriel, Rhandley Cajote

Abstract: Small oriented objects that represent tiny pixel-area in large-scale aerial images are difficult to detect due to their size and orientation. Existing oriented aerial detectors have shown promising results but are mainly focused on orientation modeling with less regard to the size of the objects. In this work, we proposed a method to accurately detect small oriented objects in aerial images by enhancing the classification and regression tasks of the oriented object detection model. We designed the Attention-Points Network consisting of two losses: Guided-Attention Loss (GALoss) and Box-Points Loss (BPLoss). GALoss uses an instance segmentation mask as ground-truth to learn the attention features needed to improve the detection of small objects. These attention features are then used to predict box points for BPLoss, which determines the points’ position relative to the target oriented bounding box. Experimental results show the effectiveness of our Attention-Points Network on a standard oriented aerial dataset with small object instances (DOTA-v1.S) and on a maritime-related dataset (HRSC2016). The code is publicly available.

Link

 

You can also visit individual Faculty pages to see the full list of their publications.