High-data-rate sensors, such as video cameras, are becoming ubiquitous in the Internet of Things. This article describes GigaSight, an Internet-scale repository crowd-sourced content, with strong enforcement privacy preferences and access controls. The GigaSight architecture is a federated system VM-based cloudlets that perform analytics at edge Internet, thus reducing demand for ingress bandwidth into cloud. Denaturing, which owner-specific reduction fidelity content to preserve privacy,...

This paper presents OptNet, a network architecture that integrates optimization problems (here, specifically in the form of quadratic programs) as individual layers larger end-to-end trainable deep networks. These encode constraints and complex dependencies between hidden states traditional convolutional fully-connected often cannot capture. We explore foundations for such an architecture: we show how techniques from sensitivity analysis, bilevel optimization, implicit differentiation can be...

An emerging class of interactive wearable cognitive assistance applications is poised to become one the key demonstrators edge computing infrastructure. In this paper, we design seven such and evaluate their performance in terms latency across a range configurations, mobile hardware, wireless networks, including 4G LTE. We also devise novel multi-algorithm approach that leverages temporal locality reduce end-to-end by 60% 70%, without sacrificing accuracy. Finally, derive target latencies...

Recent work has shown how to embed differentiable optimization problems (that is, whose solutions can be backpropagated through) as layers within deep learning architectures. This method provides a useful inductive bias for certain problems, but existing software is rigid and difficult apply new settings. In this paper, we propose an approach differentiating through disciplined convex programs, subclass of used by domain-specific languages (DSLs) optimization. We introduce parametrized...

Training an agent to solve control tasks directly from high-dimensional images with model-free reinforcement learning (RL) has proven difficult. A promising approach is learn a latent representation together the policy. However, fitting high-capacity encoder using scarce reward signal sample inefficient and leads poor performance. Prior work shown that auxiliary losses, such as image reconstruction, can aid efficient learning. incorporating reconstruction loss into off-policy algorithm often...

The widespread adoption and contextually sensitive nature of smartphone devices has increased concerns over malware. Machine learning classifiers are a current method for detecting malicious applications on systems. This paper presents the evaluation number existing classifiers, using dataset containing thousands real (i.e. not synthetic) applications. We also present our STREAM framework, which was developed to enable rapid large-scale validation mobile malware machine classifiers.

Computational offloading services at the edge of Internet for mobile devices are becoming a reality. Using wide range applications, we explore how such infrastructure improves latency and energy consumption relative to cloud. We present experimental results from WiFi 4G LTE networks that confirm substantial wins computing highly interactive applications.

We present foundations for using Model Predictive Control (MPC) as a differentiable policy class reinforcement learning in continuous state and action spaces. This provides one way of leveraging combining the advantages model-free model-based approaches. Specifically, we differentiate through MPC by KKT conditions convex approximation at fixed point controller. Using this strategy, are able to learn cost dynamics controller via end-to-end learning. Our experiments focus on imitation pendulum...

We present OpenFace, our new open-source face recognition system that approaches state-of-the-art accuracy. Integrating OpenFace with inter-frame tracking, we build RTFace, a mechanism for denaturing video streams selectively blurs faces according to specified policies at full frame rates. This enables privacy management live analytics while providing secure approach handling retrospective policy exceptions. Finally, scalable, privacy-aware architecture large camera networks using RTFace.

This paper presents the input convex neural network architecture. These are scalar-valued (potentially deep) networks with constraints on parameters such that output of is a function (some of) inputs. The allow for efficient inference via optimization over some inputs to given others, and can be applied settings including structured prediction, data imputation, reinforcement learning, others. In this we lay basic groundwork these models, proposing methods inference, analyze their...

Many (but not all) approaches self-qualifying as "meta-learning" in deep learning and reinforcement fit a common pattern of approximating the solution to nested optimization problem. In this paper, we give formalization shared pattern, which call GIMLI, prove its general requirements, derive general-purpose algorithm for implementing similar approaches. Based on analysis algorithm, describe library our design, higher, share with community assist enable future research into these kinds...

With the increasing popularity of machine learning techniques, it has become common to see prediction algorithms operating within some larger process. However, criteria by which we train these often differ from ultimate on evaluate them. This paper proposes an end-to-end approach for probabilistic models in a manner that directly captures task-based objective they will be used, context stochastic programming. We present three experimental evaluations proposed approach: classical inventory...

Unease over data privacy will retard consumer acceptance of IoT deployments. The primary source discomfort is a lack user control raw that streamed directly from sensors to the cloud. This direct consequence over-centralization today's cloud-based hub designs. We propose solution interposes locally-controlled software component called mediator on every sensor stream. Each in same administrative domain as whose being collected, and dynamically enforces current policies owners or mobile users...

VM handoff enables rapid and transparent placement changes to executing code in edge computing use cases where the safety management attributes of encapsulation are important. This versatile primitive offers functionality classic live migration but is highly optimized for edge. Over WAN bandwidths ranging from 5 25 Mbps, migrates a running 8 GB about minute, with downtime few tens seconds. By dynamically adapting varying network bandwidth processing load, more than an order magnitude faster...

Recent cyberattacks have highlighted the risk of physical equipment operating outside designed tolerances to produce catastrophic failures. A related threat is that change design and manufacturing a machine's part, such as an automobile brake component, so it no longer functions properly. These risks stem from lack cyber-physical models identify ongoing attacks well rigorous application known cybersecurity best practices. To protect processes in future, research will be needed on number...

We present Theseus, an efficient application-agnostic open source library for differentiable nonlinear least squares (DNLS) optimization built on PyTorch, providing a common framework end-to-end structured learning in robotics and vision. Existing DNLS implementations are application specific do not always incorporate many ingredients important efficiency. Theseus is application-agnostic, as we illustrate with several example applications that using the same underlying components, such...

A cognitive assistance application combines a wearable device such as Google Glass with cloudlet processing to provide step-by-step guidance on complex task. In this paper, we focus user for narrow and well-defined tasks that require specialized knowledge and/or skills. We describe proof-of-concept implementations four different tasks: assembling 2D Lego models, freehand sketching, playing ping-pong, recommending context-relevant YouTube tutorials. then reflect the difficulties faced in...