Existing machine learning inference-serving systems largely rely on hardware scaling by adding more devices or using more powerful accelerators to handle increasing query demands. However, hardware scaling might not be feasible for fixed-size edge clusters or private clouds due to their limited hardware resources. A viable alternate solution is accuracy scaling, which adapts the accuracy of ML models instead of hardware resources to handle varying query demands. This work studies the design of a high-throughput inference-serving system with accuracy scaling that can meet throughput requirements while maximizing accuracy. To achieve the goal, this work proposes to identify the right amount of accuracy scaling by jointly optimizing three sub-problems: how to select model variants, how to place them on heterogeneous devices, and how to assign query workloads to each device. It also proposes a new adaptive batching algorithm to handle variations in query arrival times and minimize SLO violations. Based on the proposed techniques, we build an inference-serving system called Proteus and empirically evaluate it on real-world and synthetic traces. We show that Proteus reduces accuracy drop by up to 3× and latency timeouts by 2--10× with respect to baseline schemes, while meeting throughput requirements.

The role of direct cell-to-cell spread in viral infections-where virions spread between host and susceptible cells without needing to be secreted into the extracellular environment-has come to be understood as essential to the dynamics of medically significant viruses like hepatitis C and influenza. Recent work in both the experimental and mathematical modelling literature has attempted to quantify the prevalence of cell-to-cell infection compared to the conventional free virus route using a variety of methods and experimental data. However, estimates are subject to significant uncertainty and moreover rely on data collected by inhibiting one mode of infection by either chemical or physical factors, which may influence the other mode of infection to an extent which is difficult to quantify. In this work, we conduct a simulation-estimation study to probe the practical identifiability of the proportion of cell-to-cell infection, using two standard mathematical models and synthetic data that would likely be realistic to obtain in the laboratory. We show that this quantity cannot be estimated using non-spatial data alone, and that the collection of data which describes the spatial structure of the infection is necessary to infer the proportion of cell-to-cell infection. Our results provide guidance for the design of relevant experiments and mathematical tools for accurately inferring the prevalence of cell-to-cell infection in in vitro and in vivo contexts.

There is growing interest in U.S. cities to shift resources towards community-led solutions to crime and disorder. However, there is a simultaneous need to provide community organizations with access to real-time data to facilitate decision making, to which only the police normally have access. In this work we present a low-cost gunshot detection system with localization that has been developed for community-based violence interruption. The distributed real-time gunshot detection sensor network is linked to a mobile phone-based alert and tasking system for exclusive use by civilian gang interventionists. Here we present details on the system architecture and gunshot detection model, which consists of an Audio Spectrogram Transformer (AST) neural network. We then combine gradient maps of the input to the AST for time of arrival identification with a Bayesian maximum a posteriori estimation procedure to identify the location of gunshots. We conduct several experiments using simulated data, open data from the commercial ShotSpotter detection system in Pittsburgh, and data collected using our devices during live-fire experiments at the Indianapolis Metropolitan Police Department (IMPD) gun firing range. We then discuss potential applications of the system and directions for future research.

Efficient GPU scheduling is the key to minimizing the execution time of the Deep Learning (DL) training workloads. DL training system schedulers typically allocate a fixed number of GPUs to each job, which inhibits high resource utilization and often extends the overall training time. The recent introduction of schedulers that can dynamically reallocate GPUs has achieved better cluster efficiency. This dynamic nature, however, introduces additional overhead by terminating and restarting jobs or requires modification to the DL training frameworks.We propose and develop an efficient, non-intrusive GPU scheduling framework that employs a combination of an adaptive GPU scheduler and an elastic GPU allocation mechanism to reduce the completion time of DL training workloads and improve resource utilization. Specifically, the adaptive GPU scheduler includes a scheduling algorithm that uses training job progress information to determine the most efficient allocation and reallocation of GPUs for incoming and running jobs at any given time. The elastic GPU allocation mechanism works in concert with the scheduler. It offers a lightweight and nonintrusive method to reallocate GPUs based on a “SideCar” process that temporarily stops and restarts the job's DL training process with a different number of GPUs. We implemented the scheduling framework as plugins in Kubernetes and conducted evaluations on two 16-GPU clusters with multiple training jobs based on TensorFlow. Results show that our proposed scheduling framework reduces the overall execution time and the average job completion time by up to 45% and 63%, respectively, compared to the Kubernetes default scheduler. Compared to a termination based scheduler, our framework reduces the overall execution time and the average job completion time by up to 20% and 37%, respectively.

Ornaments are a way to describe changes in datatype definitions reorganizing, adding, or dropping some pieces of data so that functions operating on the bare definition can be partially and sometimes totally lifted into functions operating on the ornamented structure. We propose an extension of ML with higher-order ornaments, demonstrate its expressiveness with a few typical examples, including code refactoring, study the metatheoretical properties of ornaments, and describe their elaboration process. We formalize ornamentation via an a posteriori abstraction of the bare code, returning a generic term, which lives in a meta-language above ML. The lifted code is obtained by application of the generic term to well-chosen arguments, followed by staged reduction, and some remaining simplifications. We use logical relations to closely relate the lifted code to the bare code.

Abstract—The charge-storage subharmonic parametric oscillator may be switched from one state to the other within one cycle of pump voltage. This property is used in the design and construction of a 6-element delay line. This delay line is shown to perform predictably when the elements are connected in a loop to give a self- switching arrangement.