Michael Zink

Title: Edge Support for Multimedia Streaming
Abstract: In this presentation, I will give an overview of our research in edge support for multimedia streaming. Our approach has the goal to reduce cloud-edge bandwidth in the backhaul network and to lower average end-to-end latency for 360° video streaming applications. This goal is achieved by leveraging edge-based, optimized upscaling techniques in the form of distributed super-resolution (SR). In addition, I will present how SR and downscaling can be combined with caching at the edge to further reduce bandwidth consumption in the backhaul network and to reduce end-to-end latency. The third part of my presentation will highlight how edge support can improve the Quality of Experience for live streaming.

Mohammad Hajiesmaili

Title: Robust Cooperative Learning with Heterogeneous Edge Agents
Abstract: Future machine learning will become increasingly distributed because of the large-scale nature of learning tasks and the need to implement it across a heterogeneous set of edge devices. In this talk, we will explore the power of cooperation in enhancing the performance of integrated learning among multiple edge devices. We leverage the multi-arm bandit framework and extend it to a scenario where multiple heterogeneous agents, each representing a different edge node, cooperate to solve an integrated learning task. The model captures the heterogeneity at the edge in two ways. First, agents have limited access to a local subset of action space. Second, agents are asynchronous; they make decisions at different rates. The eventual learning goal is to find the optimal global action among all agents or the best local action for each agent. The agents can cooperate with additional communication costs and/or delays to achieve this goal. We propose several algorithms for different problem settings and analyze their communication complexity and learning performance in terms of regret. We also shed light on multiple application scenarios that our mathematical models could capture.

Hava Siegelmann

Title: Temporally Aware Networks and Lifelong Learning
Abstract: Lifelong Learning is a cutting edge of artificial intelligence – encompassing computational methods that allow systems to learn in runtime and incorporate learning for application in new, unanticipated situations. Yet SOTA Lifelong Learning, like classical learning, is limited in its accuracy for temporal prediction from short and incomplete measurements. This is where our new technology appears with such capabilities.

Andrew McGregor

Title: Learning Graphs via Sampling, Sketching, and Interventions 
Abstract: Graphs are a natural abstraction in many types of problems, e.g., analyzing real-world networks such as social networks or learning causal relationships. However analyzing these graphs can be challenging for a variety of reasons. These graphs may be massive and not fit into the memory of a single machine, may be changing over time, or may be determined by data that is costly to acquire. In this talk, we review recent algorithmic work on learning and analyzing graphs given specific access models such as the ability to sample subgraphs, perform linear measurements, and perform various types of combinatorial queries such as independent set queries and interventions.

Ina Fiterau

Title: Multimodal Learning for Intelligent Sensing at the Edge
Abstract: The most significant growing component of electric energy consumption in the information and communication technology industry (ICT) comes from Data Centers and the communication of raw data to and from them. It is projected that by 2025 the amount of data stored in these centers will reach 187 Zettabytes (1021 Bytes). By 2030, data storage and associated data communication will be the fastest growing component of ICT’s energy usage which is projected to reach 10-20% of the global electric energy causing an enormous impact on CO2 emissions. The most promising among the proposed solutions is to design intelligent edge sensor platforms that can locally process raw data and use learning to transform data into critical information. Nature’s frugality in tying capacity to need provides lessons we can and should utilize in designing intelligent and energy-efficient edge devices. We offer to realize an electronic equivalent of a biological architecture for developing a multi-sensory intelligent platform that includes, at a minimum, an electromagnetic sensor for communications, a GPS, a navigation radar, and a camera. Following nature’s approach, we propose a layered, hierarchical architecture wherein analog data is preprocessed as close to the individual sensors as possible and fused at the next level of hierarchy using attention focus imposed both by internal (self-focus) and external conditions. We will build on our recent success in co-designing analog hardware with embedded deep machine learning algorithms to develop the intelligent sensor platform described above to demonstrate multiple goals; information processing resulting in a 100,000:1 reduction in bps, capability for real-time response, and 1,000:1 reduction in consumed energy. The proposed research can lead to new directions in neuro-design and empower a new generation of co-designed analog circuits with embedded deep learning algorithms that perform in memory analog processing and attention-based information fusion.

Cameron Musco

Title: Sample Constrained Treatment Effect Estimation 
Abstract: We study the design of efficient randomized controlled trials, to accurately estimate the effect of some treatment on a population of individuals. In particular, we study sample-constrained treatment effect estimation, where we must select just a small subset of the individuals to experiment on, due to computational or other resource constraints. Algorithms for partitioning the full population into treatment and control groups for experimentation, or for choosing a single representative subset, have been well-studied. The key challenge in our setting is in jointly choosing a representative subset and a partition for that set. We study both individual and average treatment effect estimation, under a linear effects model. We give provably efficient experimental designs and corresponding estimators, by identifying connections to discrepancy minimization and leverage-score-based sampling used in randomized numerical linear algebra. Our theoretical results obtain a smooth transition to known guarantees when the subset size equals the population size. We also empirically demonstrate the performance of our algorithms in several applications.

Hui Guan

Title: Towards accurate and efficient edge computing via multi-task learning  
Abstract:  AI-powered applications increasingly adopt Deep Neural Networks (DNNs) for solving many prediction tasks (e.g., semantic segmentation, object detection), leading to more than one DNNs running on resource-constrained devices. Supporting many models simultaneously on a device is challenging due to the linearly increased computation, energy, and storage costs. An effective approach to address the problem is multi-task learning (MTL) where a set of tasks are learned jointly to allow some parameter sharing among tasks. MTL creates multi-task models based on DNN architectures called backbone models and has shown significantly reduced inference costs and improved generalization performance in many machine learning applications. In this talk, we will introduce our recent efforts on leveraging MTL to improve inference efficiency and prediction accuracy for edge computing. The talk will first introduce multi-task architecture design systems that can automatically identify resource-efficient multi-task models with low inference costs and high task accuracy. It will then discuss multi-task learning in federated learning settings and demonstrates the effectiveness of MTL in improving prediction accuracy.