Publications

Oobleck enables resilient distributed training of large DNN models with guaranteed fault tolerance. It takes a planning-execution co-design approach, where it first generates a set of heterogeneous pipeline templates and instantiates at least f+1 logically equivalent pipeline replicas to tolerate any f simultaneous failures. During execution, it relies on already-replicated model states across the replicas to provide fast recovery. Oobleck provably guarantees that some combination of the initially created pipeline templates can be used to cover all available resources after f or fewer simultaneous failures, thereby avoiding resource idling at all times. Evaluation on large DNN models with billions of parameters shows that Oobleck provides consistently high throughput, and it outperforms state-of-the-art fault tolerance solutions like Bamboo and Varuna by up to 13.9x.

Cycling is an efficient and effective way to improve one's overall fitness level, such as cardiovascular fitness, stamina, lower body strength, and body fat percentage. To improve fitness performance, real-time cycling fitness tracking can not only allow cyclists to better control their energy outputs but also help push workout intensity and keep users accountable for their fitness progress. However, existing bike sensors (e.g., the ones mounted to bike's wheel hub or crank arm) are only limited to measuring cycling cadence and speed. Although several recent studies relying on on-body sensors or cameras can provide more fine-grained information (e.g., riding position and knee joint angle), they would either require inconvenient setups or raise serious privacy concerns. To circumvent these limitations, in this paper, we propose SmarCyPad, an innovative smart seat pad that can continuously and unobtrusively track five cycling-specific metrics, including cadence, per-leg stability, leg strength balance, riding position, and knee joint angle of the cyclist. Specifically, we embed conductive fabric sensors in the seat pad to sense the pressure applied to the bike's seat exerted by the cyclist's gluteal muscles. A series of signal processing algorithms are developed to estimate the pedaling period from the sensed pressure signal and further derive the cycling cadence, per-leg stability, and leg strength balance. Additionally, we leverage a deep learning model to detect the cyclist's riding position and reconstruct the cyclist's knee joint angles via linear regression. The sensors and the system prototype are manufactured from scratch leveraging off-the-shelf materials, and the total cost is less than $50. Extensive experiments involving 15 participants demonstrate that SmarCyPad can accurately estimate the cycling cadence with an average error of 1.13 rounds per minute, quantify the cycling stability for each leg, detect cycling imbalance, distinguish five riding positions with an accuracy of 96.60%, and continuously track the knee joint angle with an average mean error as low as 9.58 degrees.

Surface emissivity (ε) plays a critical role in Earth’s radiation budget and climate. The Polar Radiant Energy in the Far-InfraRed Experiment (PREFIRE) and the Far-Infrared Outgoing Radiation Understanding and Monitoring (FORUM) satellite missions aim to estimate surface spectral emissivities in mid-IR and far-IR regions. This study presents a neural network (NN)-based surface spectral emissivity retrieval algorithm under clear-sky that offers comparable performance to optimal-estimation (OE)-based methods but reduces computation time by a factor of 10^5. The NN-based method has achieved a mean relative retrieval error (|∆ε|) of 0.0028 with a standard deviation of 0.0013. Shapley values are employed to interpret the algorithm’s results and assess the relative importance of input features. The results derived from the Shapley values analysis are in good agreement with the physical understanding. The study highlights the effectiveness of the proposed neural network approach for surface emissivity estimation in forthcoming satellite missions.

Deep learning has experienced significant growth in recent years, resulting in increased energy consumption and carbon emission from the use of GPUs for training deep neural networks (DNNs). Answering the call for sustainability, conventional solutions have attempted to move training jobs to locations or time frames with lower carbon intensity. However, moving jobs to other locations may not always be feasible due to large dataset sizes or data regulations. Moreover, postponing training can negatively impact application service quality because the DNNs backing the service are not updated in a timely fashion. In this work, we present a practical solution that reduces the carbon footprint of DNN training without migrating or postponing jobs. Specifically, our solution observes real-time carbon intensity shifts during training and controls the energy consumption of GPUs, thereby reducing carbon footprint while maintaining training performance. Furthermore, in order to proactively adapt to shifting carbon intensity, we propose a lightweight machine learning algorithm that predicts the carbon intensity of the upcoming time frame. Our solution, Chase, reduces the total carbon footprint of training ResNet-50 on ImageNet by 13.6% while only increasing training time by 2.5%.

Semantic segmentation is a challenging computer vision task demanding a significant amount of pixel-level annotated data. Producing such data is a time-consuming and costly process, especially for domains with a scarcity of experts, such as medicine or forensic anthropology. While numerous semi-supervised approaches have been developed to make the most from the limited labeled data and ample amount of unlabeled data, domain-specific real-world datasets often have characteristics that both reduce the effectiveness of off-the-shelf state-of-the-art methods and also provide opportunities to create new methods that exploit these characteristics. We propose and evaluate a semi-supervised method that reuses available labels for unlabeled images of a dataset by exploiting existing similarities, while dynamically weighting the impact of these reused labels in the training process. We evaluate our method on a large dataset of human decomposition images and find that our method, while conceptually simple, outperforms state-of-the-art consistency and pseudo-labeling-based methods for the segmentation of this dataset. This paper includes graphic content of human decomposition.

Annotating images for semantic segmentation requires intense manual labor and is a time-consuming and expensive task especially for domains with a scarcity of experts, such as Forensic Anthropology. We leverage the evolving nature of images depicting the decay process in human decomposition data to design a simple yet effective pseudo-pixel-level label generation technique to reduce the amount of effort for manual annotation of such images. We first identify sequences of images with a minimum variation that are most suitable to share the same or similar annotation using an unsupervised approach. Given one user-annotated image in each sequence, we propagate the annotation to the remaining images in the sequence by merging it with annotations produced by a state-of-the-art CAM-based pseudo label generation technique. To evaluate the quality of our pseudo-pixel-level labels, we train two semantic segmentation models with VGG and ResNet backbones on images labeled using our pseudo labeling method and those of a state-of-the-art method. The results indicate that using our pseudo-labels instead of those generated using the state-of-the-art method in the training process improves the mean-IoU and the frequency-weighted-IoU of the VGG and ResNet-based semantic segmentation models by 3.36%, 2.58%, 10.39%, and 12.91% respectively.