Bugra Gedik

Mannat Singh, Laura Gustafson, Aaron Adcock et al.

Model pre-training is a cornerstone of modern visual recognition systems. Although fully supervised pre-training on datasets like ImageNet is still the de-facto standard, recent studies suggest that large-scale weakly supervised pre-training can outperform fully supervised approaches. This paper revisits weakly-supervised pre-training of models using hashtag supervision with modern versions of residual networks and the largest-ever dataset of images and corresponding hashtags. We study the performance of the resulting models in various transfer-learning settings including zero-shot transfer. We also compare our models with those obtained via large-scale self-supervised learning. We find our weakly-supervised models to be very competitive across all settings, and find they substantially outperform their self-supervised counterparts. We also include an investigation into whether our models learned potentially troubling associations or stereotypes. Overall, our results provide a compelling argument for the use of weakly supervised learning in the development of visual recognition systems. Our models, Supervised Weakly through hashtAGs (SWAG), are available publicly.

Mark Zhao, Niket Agarwal, Aarti Basant et al.

Datacenter-scale AI training clusters consisting of thousands of domain-specific accelerators (DSA) are used to train increasingly-complex deep learning models. These clusters rely on a data storage and ingestion (DSI) pipeline, responsible for storing exabytes of training data and serving it at tens of terabytes per second. As DSAs continue to push training efficiency and throughput, the DSI pipeline is becoming the dominating factor that constrains the overall training performance and capacity. Innovations that improve the efficiency and performance of DSI systems and hardware are urgent, demanding a deep understanding of DSI characteristics and infrastructure at scale. This paper presents Meta's end-to-end DSI pipeline, composed of a central data warehouse built on distributed storage and a Data PreProcessing Service that scales to eliminate data stalls. We characterize how hundreds of models are collaboratively trained across geo-distributed datacenters via diverse and continuous training jobs. These training jobs read and heavily filter massive and evolving datasets, resulting in popular features and samples used across training jobs. We measure the intense network, memory, and compute resources required by each training job to preprocess samples during training. Finally, we synthesize key takeaways based on our production infrastructure characterization. These include identifying hardware bottlenecks, discussing opportunities for heterogeneous DSI hardware, motivating research in datacenter scheduling and benchmark datasets, and assimilating lessons learned in optimizing DSI infrastructure.

Fuat Basik, Bugra Gedik, Hakan Ferhatosmanoglu et al.

Faster and more cost-efficient, crowdsourced delivery is needed to meet the growing customer demands of many industries, including online shopping, on-demand local delivery, and on-demand transportation. The power of crowdsourced delivery stems from the large number of workers potentially available to provide services and reduce costs. It has been shown in social psychology literature that fairness is key to ensuring high worker participation. However, existing assignment solutions fall short on modeling the dynamic fairness metric. In this work, we introduce a new assignment strategy for crowdsourced delivery tasks. This strategy takes fairness towards workers into consideration, while maximizing the task allocation ratio. Since redundant assignments are not possible in delivery tasks, we first introduce a 2-phase allocation model that increases the reliability of a worker to complete a given task. To realize the effectiveness of our model in practice, we present both offline and online versions of our proposed algorithm called F-Aware. Given a task-to-worker bipartite graph, F-Aware assigns each task to a worker that minimizes unfairness, while allocating tasks to use worker capacities as much as possible. We present an evaluation of our algorithms with respect to running time, task allocation ratio (TAR), as well as unfairness and assignment ratio. Experiments show that F-Aware runs around 10^7 x faster than the TAR-optimal solution and allocates 96.9% of the tasks that can be allocated by it. Moreover, it is shown that, F-Aware is able to provide a much fair distribution of tasks to workers than the best competitor algorithm.