Yung-Hsiang Lu

Wenxin Jiang, Nicholas Synovic, Matt Hyatt et al.

Deep Neural Networks (DNNs) are being adopted as components in software systems. Creating and specializing DNNs from scratch has grown increasingly difficult as state-of-the-art architectures grow more complex. Following the path of traditional software engineering, machine learning engineers have begun to reuse large-scale pre-trained models (PTMs) and fine-tune these models for downstream tasks. Prior works have studied reuse practices for traditional software packages to guide software engineers towards better package maintenance and dependency management. We lack a similar foundation of knowledge to guide behaviors in pre-trained model ecosystems. In this work, we present the first empirical investigation of PTM reuse. We interviewed 12 practitioners from the most popular PTM ecosystem, Hugging Face, to learn the practices and challenges of PTM reuse. From this data, we model the decision-making process for PTM reuse. Based on the identified practices, we describe useful attributes for model reuse, including provenance, reproducibility, and portability. Three challenges for PTM reuse are missing attributes, discrepancies between claimed and actual performance, and model risks. We substantiate these identified challenges with systematic measurements in the Hugging Face ecosystem. Our work informs future directions on optimizing deep learning ecosystems by automated measuring useful attributes and potential attacks, and envision future research on infrastructure and standardization for model registries.

Caleb Tung, Abhinav Goel, Fischer Bordwell et al.

Object detectors are vital to many modern computer vision applications. However, even state-of-the-art object detectors are not perfect. On two images that look similar to human eyes, the same detector can make different predictions because of small image distortions like camera sensor noise and lighting changes. This problem is called inconsistency. Existing accuracy metrics do not properly account for inconsistency, and similar work in this area only targets improvements on artificial image distortions. Therefore, we propose a method to use non-artificial video frames to measure object detection consistency over time, across frames. Using this method, we show that the consistency of modern object detectors ranges from 83.2% to 97.1% on different video datasets from the Multiple Object Tracking Challenge. We conclude by showing that applying image distortion corrections like .WEBP Image Compression and Unsharp Masking can improve consistency by as much as 5.1%, with no loss in accuracy.

Caleb Tung, Abhinav Goel, Xiao Hu et al.

Computer vision is often performed using Convolutional Neural Networks (CNNs). CNNs are compute-intensive and challenging to deploy on power-contrained systems such as mobile and Internet-of-Things (IoT) devices. CNNs are compute-intensive because they indiscriminately compute many features on all pixels of the input image. We observe that, given a computer vision task, images often contain pixels that are irrelevant to the task. For example, if the task is looking for cars, pixels in the sky are not very useful. Therefore, we propose that a CNN be modified to only operate on relevant pixels to save computation and energy. We propose a method to study three popular computer vision datasets, finding that 48% of pixels are irrelevant. We also propose the focused convolution to modify a CNN's convolutional layers to reject the pixels that are marked irrelevant. On an embedded device, we observe no loss in accuracy, while inference latency, energy consumption, and multiply-add count are all reduced by about 45%.

Caleb Tung, Nicholas Eliopoulos, Purvish Jajal et al.

Computer vision often uses highly accurate Convolutional Neural Networks (CNNs), but these deep learning models are associated with ever-increasing energy and computation requirements. Producing more energy-efficient CNNs often requires model training which can be cost-prohibitive. We propose a novel, automated method to make a pretrained CNN more energy-efficient without re-training. Given a pretrained CNN, we insert a threshold layer that filters activations from the preceding layers to identify regions of the image that are irrelevant, i.e. can be ignored by the following layers while maintaining accuracy. Our modified focused convolution operation saves inference latency (by up to 25%) and energy costs (by up to 22%) on various popular pretrained CNNs, with little to no loss in accuracy.

Purvish Jajal, Wenxin Jiang, Arav Tewari et al.

Software engineers develop, fine-tune, and deploy deep learning (DL) models using a variety of development frameworks and runtime environments. DL model converters move models between frameworks and to runtime environments. Conversion errors compromise model quality and disrupt deployment. However, the failure characteristics of DL model converters are unknown, adding risk when using DL interoperability technologies. This paper analyzes failures in DL model converters. We survey software engineers about DL interoperability tools, use cases, and pain points (N=92). Then, we characterize failures in model converters associated with the main interoperability tool, ONNX (N=200 issues in PyTorch and TensorFlow). Finally, we formulate and test two hypotheses about structural causes for the failures we studied. We find that the node conversion stage of a model converter accounts for ~75% of the defects and 33% of reported failure are related to semantically incorrect models. The cause of semantically incorrect models is elusive, but models with behaviour inconsistencies share operator sequences. Our results motivate future research on making DL interoperability software simpler to maintain, extend, and validate. Research into behavioural tolerances and architectural coverage metrics could be fruitful.

Nick John Eliopoulos, Purvish Jajal, James C. Davis et al.

This paper investigates how to efficiently deploy vision transformers on edge devices for small workloads. Recent methods reduce the latency of transformer neural networks by removing or merging tokens, with small accuracy degradation. However, these methods are not designed with edge device deployment in mind: they do not leverage information about the latency-workload trends to improve efficiency. We address this shortcoming in our work. First, we identify factors that affect ViT latency-workload relationships. Second, we determine token pruning schedule by leveraging non-linear latency-workload relationships. Third, we demonstrate a training-free, token pruning method utilizing this schedule. We show other methods may increase latency by 2-30%, while we reduce latency by 9-26%. For similar latency (within 5.2% or 7ms) across devices we achieve 78.6%-84.5% ImageNet1K accuracy, while the state-of-the-art, Token Merging, achieves 45.8%-85.4%.

Purvish Jajal, Nick John Eliopoulos, Benjamin Shiue-Hal Chou et al.

We propose Vision Token Turing Machines (ViTTM), an efficient, low-latency, memory-augmented Vision Transformer (ViT). Our approach builds on Neural Turing Machines and Token Turing Machines, which were applied to NLP and sequential visual understanding tasks. ViTTMs are designed for non-sequential computer vision tasks such as image classification and segmentation. Our model creates two sets of tokens: process tokens and memory tokens; process tokens pass through encoder blocks and read-write from memory tokens at each encoder block in the network, allowing them to store and retrieve information from memory. By ensuring that there are fewer process tokens than memory tokens, we are able to reduce the inference time of the network while maintaining its accuracy. On ImageNet-1K, the state-of-the-art ViT-B has median latency of 529.5ms and 81.0% accuracy, while our ViTTM-B is 56% faster (234.1ms), with 2.4 times fewer FLOPs, with an accuracy of 82.9%. On ADE20K semantic segmentation, ViT-B achieves 45.65mIoU at 13.8 frame-per-second (FPS) whereas our ViTTM-B model acheives a 45.17 mIoU with 26.8 FPS (+94%).

Benjamin Shiue-Hal Chou, Purvish Jajal, Nicholas John Eliopoulos et al.

Beginner musicians often struggle to identify specific errors in their performances, such as playing incorrect notes or rhythms. There are two limitations in existing tools for music error detection: (1) Existing approaches rely on automatic alignment; therefore, they are prone to errors caused by small deviations between alignment targets.; (2) There is a lack of sufficient data to train music error detection models, resulting in over-reliance on heuristics. To address (1), we propose a novel transformer model, Polytune, that takes audio inputs and outputs annotated music scores. This model can be trained end-to-end to implicitly align and compare performance audio with music scores through latent space representations. To address (2), we present a novel data generation technique capable of creating large-scale synthetic music error datasets. Our approach achieves a 64.1% average Error Detection F1 score, improving upon prior work by 40 percentage points across 14 instruments. Additionally, compared with existing transcription methods repurposed for music error detection, our model can handle multiple instruments. Our source code and datasets are available at https://github.com/ben2002chou/Polytune.

Ojas Chaturvedi, Kayshav Bhardwaj, Tanay Gondil et al.

This paper presents the results of the 2025 Automatic Music Transcription (AMT) Challenge, an online competition to benchmark progress in multi-instrument transcription. Eight teams submitted valid solutions; two outperformed the baseline MT3 model. The results highlight both advances in transcription accuracy and the remaining difficulties in handling polyphony and timbre variation. We conclude with directions for future challenges: broader genre coverage and stronger emphasis on instrument detection.

Paolo Wang, Michael Zhang, Shrinand Perumal et al.

Posture is a critical factor for beginning instrumental learners. Most students receive instruction only once a week, and during the intervals between lessons they have little or no feedback on their physical posture. As a result, posture often deteriorates, increasing the risk of musculoskeletal injury and inefficient technique. Recent advances in computer vision and machine learning make it possible to evaluate posture without the constant presence of a human expert. However, current solutions have been extremely limited in availability and convenience due to their reliance on computationally expensive hardware or multi-sensor setups. We present Cello Evaluator, a real-time postural feedback system for practicing cellists. Through this optimization for on-device computer vision inference, we provide access to cellist postural evaluation to anyone with a current generation Android phone and thus reduces the postural feedback voids within individual practice. To validate our mobile application, we conduct a heuristic evaluation consisting of cellist and UX experts. Overall feedback from the evaluation found the app to be user friendly and helpful.

Purvish Jajal, Nick John Eliopoulos, Benjamin Shiue-Hal Chou et al.

Diffusion models are state-of-the-art generative models in various domains, yet their samples often fail to satisfy downstream objectives such as safety constraints or domain-specific validity. Existing techniques for alignment require gradients, internal model access, or large computational budgets. We introduce an inference-time alignment framework based on evolutionary algorithms. We treat diffusion models as black-boxes and search their latent space to maximize alignment objectives. Our method enables efficient inference-time alignment for both differentiable and non-differentiable alignment objectives across a range of diffusion models. On the DrawBench and Open Image Preferences benchmark, our EA methods outperform state-of-the-art gradient-based and gradient-free inference-time methods. In terms of memory consumption, we require 55% to 76% lower GPU memory than gradient-based methods. In terms of running-time, we are 72% to 80% faster than gradient-based methods. We achieve higher alignment scores over 50 optimization steps on Open Image Preferences than gradient-based and gradient-free methods.

Purvish Jajal, Nick John Eliopoulos, Benjamin Shiue-Hal Chou et al.

Modern transformer architectures achieve remarkable performance across tasks and domains but remain rigid in how they allocate computation at inference time. Real-world deployment often requires models to adapt to diverse hardware and latency constraints, yet most approaches to dynamic computation focus on a single axis -- such as reducing the number of tokens. We present a novel capability: AdaPerceiver, the first transformer architecture with unified adaptivity across depth, width, and tokens within a single model. We propose an architecture that supports adaptivity along these axes. We couple this with an efficient joint training regime that ensures the model maintains performance across its various configurations. We evaluate AdaPerceiver on image classification, semantic segmentation, and depth estimation tasks. On image classification, AdaPerceiver expands the accuracy-throughput Pareto front. It achieves 85.4% accuracy while yielding 36% higher throughput than FlexiViT-L. On dense prediction, AdaPerceiver matches ViT-H/14 while having $\sim$26x fewer encoder FLOPs (floating-point operations) on semantic segmentation and depth estimation. Finally, we show how AdaPerceiver equipped with a policy can maintain ImageNet1K accuracy ($\pm0.1$ percentage points) while reducing FLOPs by $24-33$%.

Benjamin Shiue-Hal Chou, Purvish Jajal, Nick John Eliopoulos et al.

Music learners can greatly benefit from tools that accurately detect errors in their practice. Existing approaches typically compare audio recordings to music scores using heuristics or learnable models. This paper introduces \textit{LadderSym}, a novel Transformer-based method for music error detection. \textit{LadderSym} is guided by two key observations about the state-of-the-art approaches: (1) late fusion limits inter-stream alignment and cross-modality comparison capability; and (2) reliance on score audio introduces ambiguity in the frequency spectrum, degrading performance in music with concurrent notes. To address these limitations, \textit{LadderSym} introduces (1) a two-stream encoder with inter-stream alignment modules to improve audio comparison capabilities and error detection F1 scores, and (2) a multimodal strategy that leverages both audio and symbolic scores by incorporating symbolic representations as decoder prompts, reducing ambiguity and improving F1 scores. We evaluate our method on the \textit{MAESTRO-E} and \textit{CocoChorales-E} datasets by measuring the F1 score for each note category. Compared to the previous state of the art, \textit{LadderSym} more than doubles F1 for missed notes on \textit{MAESTRO-E} (26.8\% $\rightarrow$ 56.3\%) and improves extra note detection by 14.4 points (72.0\% $\rightarrow$ 86.4\%). Similar gains are observed on \textit{CocoChorales-E}. This work introduces general insights about comparison models that could inform sequence evaluation tasks for reinforcement Learning, human skill assessment, and model evaluation.

Abhinav Goel, Caleb Tung, Xiao Hu et al.

Computer vision on low-power edge devices enables applications including search-and-rescue and security. State-of-the-art computer vision algorithms, such as Deep Neural Networks (DNNs), are too large for inference on low-power edge devices. To improve efficiency, some existing approaches parallelize DNN inference across multiple edge devices. However, these techniques introduce significant communication and synchronization overheads or are unable to balance workloads across devices. This paper demonstrates that the hierarchical DNN architecture is well suited for parallel processing on multiple edge devices. We design a novel method that creates a parallel inference pipeline for computer vision problems that use hierarchical DNNs. The method balances loads across the collaborating devices and reduces communication costs to facilitate the processing of multiple video frames simultaneously with higher throughput. Our experiments consider a representative computer vision problem where image recognition is performed on each video frame, running on multiple Raspberry Pi 4Bs. With four collaborating low-power edge devices, our approach achieves 3.21X higher throughput, 68% less energy consumption per device per frame, and 58% decrease in memory when compared with existing single-device hierarchical DNNs.

Vishnu Banna, Akhil Chinnakotla, Zhengxin Yan et al.

Machine learning techniques are becoming a fundamental tool for scientific and engineering progress. These techniques are applied in contexts as diverse as astronomy and spam filtering. However, correctly applying these techniques requires careful engineering. Much attention has been paid to the technical potential; relatively little attention has been paid to the software engineering process required to bring research-based machine learning techniques into practical utility. Technology companies have supported the engineering community through machine learning frameworks such as TensorFLow and PyTorch, but the details of how to engineer complex machine learning models in these frameworks have remained hidden. To promote best practices within the engineering community, academic institutions and Google have partnered to launch a Special Interest Group on Machine Learning Models (SIGMODELS) whose goal is to develop exemplary implementations of prominent machine learning models in community locations such as the TensorFlow Model Garden (TFMG). The purpose of this report is to define a process for reproducing a state-of-the-art machine learning model at a level of quality suitable for inclusion in the TFMG. We define the engineering process and elaborate on each step, from paper analysis to model release. We report on our experiences implementing the YOLO model family with a team of 26 student researchers, share the tools we developed, and describe the lessons we learned along the way.

Abhinav Goel, Caleb Tung, Xiao Hu et al.

Low-power computer vision on embedded devices has many applications. This paper describes a low-power technique for the object re-identification (reID) problem: matching a query image against a gallery of previously seen images. State-of-the-art techniques rely on large, computationally-intensive Deep Neural Networks (DNNs). We propose a novel hierarchical DNN architecture that uses attribute labels in the training dataset to perform efficient object reID. At each node in the hierarchy, a small DNN identifies a different attribute of the query image. The small DNN at each leaf node is specialized to re-identify a subset of the gallery: only the images with the attributes identified along the path from the root to a leaf. Thus, a query image is re-identified accurately after processing with a few small DNNs. We compare our method with state-of-the-art object reID techniques. With a 4% loss in accuracy, our approach realizes significant resource savings: 74% less memory, 72% fewer operations, and 67% lower query latency, yielding 65% less energy consumption.

Ryan Dailey, Aniesh Chawla, Andrew Liu et al.

Reduction in the cost of Network Cameras along with a rise in connectivity enables entities all around the world to deploy vast arrays of camera networks. Network cameras offer real-time visual data that can be used for studying traffic patterns, emergency response, security, and other applications. Although many sources of Network Camera data are available, collecting the data remains difficult due to variations in programming interface and website structures. Previous solutions rely on manually parsing the target website, taking many hours to complete. We create a general and automated solution for aggregating Network Camera data spread across thousands of uniquely structured web pages. We analyze heterogeneous web page structures and identify common characteristics among 73 sample Network Camera websites (each website has multiple web pages). These characteristics are then used to build an automated camera discovery module that crawls and aggregates Network Camera data. Our system successfully extracts 57,364 Network Cameras from 237,257 unique web pages.

Isha Ghodgaonkar, Subhankar Chakraborty, Vishnu Banna et al.

In order to contain the COVID-19 pandemic, countries around the world have introduced social distancing guidelines as public health interventions to reduce the spread of the disease. However, monitoring the efficacy of these guidelines at a large scale (nationwide or worldwide) is difficult. To make matters worse, traditional observational methods such as in-person reporting is dangerous because observers may risk infection. A better solution is to observe activities through network cameras; this approach is scalable and observers can stay in safe locations. This research team has created methods that can discover thousands of network cameras worldwide, retrieve data from the cameras, analyze the data, and report the sizes of crowds as different countries issued and lifted restrictions (also called ''lockdown''). We discover 11,140 network cameras that provide real-time data and we present the results across 15 countries. We collect data from these cameras beginning April 2020 at approximately 0.5TB per week. After analyzing 10,424,459 images from still image cameras and frames extracted periodically from video, the data reveals that the residents in some countries exhibited more activity (judged by numbers of people and vehicles) after the restrictions were lifted. In other countries, the amounts of activities showed no obvious changes during the restrictions and after the restrictions were lifted. The data further reveals whether people stay ''social distancing'', at least 6 feet apart. This study discerns whether social distancing is being followed in several types of locations and geographical locations worldwide and serve as an early indicator whether another wave of infections is likely to occur soon.

Abhinav Goel, Caleb Tung, Sara Aghajanzadeh et al.

Deep Neural Networks (DNNs) can achieve state-of-the-art accuracy in many computer vision tasks, such as object counting. Object counting takes two inputs: an image and an object query and reports the number of occurrences of the queried object. To achieve high accuracy on such tasks, DNNs require billions of operations, making them difficult to deploy on resource-constrained, low-power devices. Prior work shows that a significant number of DNN operations are redundant and can be eliminated without affecting the accuracy. To reduce these redundancies, we propose a hierarchical DNN architecture for object counting. This architecture uses a Region Proposal Network (RPN) to propose regions-of-interest (RoIs) that may contain the queried objects. A hierarchical classifier then efficiently finds the RoIs that actually contain the queried objects. The hierarchy contains groups of visually similar object categories. Small DNNs are used at each node of the hierarchy to classify between these groups. The RoIs are incrementally processed by the hierarchical classifier. If the object in an RoI is in the same group as the queried object, then the next DNN in the hierarchy processes the RoI further; otherwise, the RoI is discarded. By using a few small DNNs to process each image, this method reduces the memory requirement, inference time, energy consumption, and number of operations with negligible accuracy loss when compared with the existing object counters.

Abhinav Goel, Caleb Tung, Yung-Hsiang Lu et al.

Deep neural networks (DNNs) are successful in many computer vision tasks. However, the most accurate DNNs require millions of parameters and operations, making them energy, computation and memory intensive. This impedes the deployment of large DNNs in low-power devices with limited compute resources. Recent research improves DNN models by reducing the memory requirement, energy consumption, and number of operations without significantly decreasing the accuracy. This paper surveys the progress of low-power deep learning and computer vision, specifically in regards to inference, and discusses the methods for compacting and accelerating DNN models. The techniques can be divided into four major categories: (1) parameter quantization and pruning, (2) compressed convolutional filters and matrix factorization, (3) network architecture search, and (4) knowledge distillation. We analyze the accuracy, advantages, disadvantages, and potential solutions to the problems with the techniques in each category. We also discuss new evaluation metrics as a guideline for future research.

Sergei Alyamkin, Matthew Ardi, Alexander C. Berg et al.

Computer vision has achieved impressive progress in recent years. Meanwhile, mobile phones have become the primary computing platforms for millions of people. In addition to mobile phones, many autonomous systems rely on visual data for making decisions and some of these systems have limited energy (such as unmanned aerial vehicles also called drones and mobile robots). These systems rely on batteries and energy efficiency is critical. This article serves two main purposes: (1) Examine the state-of-the-art for low-power solutions to detect objects in images. Since 2015, the IEEE Annual International Low-Power Image Recognition Challenge (LPIRC) has been held to identify the most energy-efficient computer vision solutions. This article summarizes 2018 winners' solutions. (2) Suggest directions for research as well as opportunities for low-power computer vision.

Yung-Hsiang Lu, George K. Thiruvathukal, Ahmed S. Kaseb et al.

Millions of network cameras have been deployed worldwide. Real-time data from many network cameras can offer instant views of multiple locations with applications in public safety, transportation management, urban planning, agriculture, forestry, social sciences, atmospheric information, and more. This paper describes the real-time data available from worldwide network cameras and potential applications. Second, this paper outlines the CAM2 System available to users at https://www.cam2project.net/. This information includes strategies to discover network cameras and create the camera database, user interface, and computing platforms. Third, this paper describes many opportunities provided by data from network cameras and challenges to be addressed.

Chen Feng, Tao Sheng, Zhiyu Liang et al.

The IEEE Low-Power Image Recognition Challenge (LPIRC) is an annual competition started in 2015 that encourages joint hardware and software solutions for computer vision systems with low latency and power. Track 1 of the competition in 2018 focused on the innovation of software solutions with fixed inference engine and hardware. This decision allows participants to submit models online and not worry about building and bringing custom hardware on-site, which attracted a historically large number of submissions. Among the diverse solutions, the winning solution proposed a quantization-friendly framework for MobileNets that achieves an accuracy of 72.67% on the holdout dataset with an average latency of 27ms on a single CPU core of Google Pixel2 phone, which is superior to the best real-time MobileNet models at the time.

Chittayong Surakitbanharn, Calvin Yau, Guizhen Wang et al.

Physical media (like surveillance cameras) and social media (like Instagram and Twitter) may both be useful in attaining on-the-ground information during an emergency or disaster situation. However, the intersection and reliability of both surveillance cameras and social media during a natural disaster are not fully understood. To address this gap, we tested whether social media is of utility when physical surveillance cameras went off-line during Hurricane Irma in 2017. Specifically, we collected and compared geo-tagged Instagram and Twitter posts in the state of Florida during times and in areas where public surveillance cameras went off-line. We report social media content and frequency and content to determine the utility for emergency managers or first responders during a natural disaster.

Caleb Tung, Matthew R. Kelleher, Ryan J. Schlueter et al.

Computer vision relies on labeled datasets for training and evaluation in detecting and recognizing objects. The popular computer vision program, YOLO ("You Only Look Once"), has been shown to accurately detect objects in many major image datasets. However, the images found in those datasets, are independent of one another and cannot be used to test YOLO's consistency at detecting the same object as its environment (e.g. ambient lighting) changes. This paper describes a novel effort to evaluate YOLO's consistency for large-scale applications. It does so by working (a) at large scale and (b) by using consecutive images from a curated network of public video cameras deployed in a variety of real-world situations, including traffic intersections, national parks, shopping malls, university campuses, etc. We specifically examine YOLO's ability to detect objects in different scenarios (e.g., daytime vs. night), leveraging the cameras' ability to rapidly retrieve many successive images for evaluating detection consistency. Using our camera network and advanced computing resources (supercomputers), we analyzed more than 5 million images captured by 140 network cameras in 24 hours. Compared with labels marked by humans (considered as "ground truth"), YOLO struggles to consistently detect the same humans and cars as their positions change from one frame to the next; it also struggles to detect objects at night time. Our findings suggest that state-of-the art vision solutions should be trained by data from network camera with contextual information before they can be deployed in applications that demand high consistency on object detection.

Sergei Alyamkin, Matthew Ardi, Achille Brighton et al.

The Low-Power Image Recognition Challenge (LPIRC, https://rebootingcomputing.ieee.org/lpirc) is an annual competition started in 2015. The competition identifies the best technologies that can classify and detect objects in images efficiently (short execution time and low energy consumption) and accurately (high precision). Over the four years, the winners' scores have improved more than 24 times. As computer vision is widely used in many battery-powered systems (such as drones and mobile phones), the need for low-power computer vision will become increasingly important. This paper summarizes LPIRC 2018 by describing the three different tracks and the winners' solutions.