A severe bandwidth mismatch between incoming sensor data rate and wireless backhaul bandwidth often exists on unmanned probes when collecting new training data for machine learning (ML). To overcome this mismatch, we describe a self-improving ML-based transmission system called Hawk. Starting from a weak model that is trained on just a few examples, it seamlessly pipelines semi-supervised learning, active learning, and transfer learning, with asynchronous bandwidth-sensitive data transmission to a distant human for labeling. When a significant number of true positives (TPs) have been labeled, Hawk trains an improved model to replace the old model. This iterative workflow, called Live Learning, continues until a sufficient number of TPs have been collected. For very rare events on challenging datasets, and bandwidths as low as 12 kbps, a team of 7 probes using Hawk discovers up to 87% of the TPs that could have been discovered via full preview, transmission and labeling of all mission data. Hawk also uses diversity sampling and few-shot learning.
We propose hardware acceleration for a new edge computing abstraction called a Silent Witness. This abstraction embodies a severe asymmetry in the ease of write versus read operations. Surveillance data from one or more video cameras is continuously encrypted and recorded, but the decrypting, processing, or transmission of that data only occurs under stringent privacy controls. For the new search workloads of such a system, decode-enabled storage alleviates the scalability bottleneck imposed by frequent decoding of data. Our experiments show throughput improvements up to 3.5x for typical search workloads of a Silent Witness.
Finding true positives (TPs) to construct a training set for a new class of interest in machine learning (ML) is often a challenge. The novelty of the class suggests that cloud archives are unlikely to be helpful. We observe that most video data collected for surveillance and briefly stored at the edge before being overwritten is currently unused. To efficiently harness this untapped resource, we describe Delphi, a privacy-sensitive interactive labeling system that continuously improves labeling productivity through background learning. Our experimental results confirm the value of Delphi for training set construction from edge-sourced data.