Sachin Katti

Pete Warden, Matthew Stewart, Brian Plancher et al.

Machine learning sensors represent a paradigm shift for the future of embedded machine learning applications. Current instantiations of embedded machine learning (ML) suffer from complex integration, lack of modularity, and privacy and security concerns from data movement. This article proposes a more data-centric paradigm for embedding sensor intelligence on edge devices to combat these challenges. Our vision for "sensor 2.0" entails segregating sensor input data and ML processing from the wider system at the hardware level and providing a thin interface that mimics traditional sensors in functionality. This separation leads to a modular and easy-to-use ML sensor device. We discuss challenges presented by the standard approach of building ML processing into the software stack of the controlling microprocessor on an embedded system and how the modularity of ML sensors alleviates these problems. ML sensors increase privacy and accuracy while making it easier for system builders to integrate ML into their products as a simple component. We provide examples of prospective ML sensors and an illustrative datasheet as a demonstration and hope that this will build a dialogue to progress us towards sensor 2.0.

Hang Qiu, Ioanna Vavelidou, Jian Li et al.

Benefiting from expanding cloud infrastructure, deep neural networks (DNNs) today have increasingly high performance when trained in the cloud. Researchers spend months of effort competing for an extra few percentage points of model accuracy. However, when these models are actually deployed on edge devices in practice, very often, the performance can abruptly drop over 10% without obvious reasons. The key challenge is that there is not much visibility into ML inference execution on edge devices, and very little awareness of potential issues during the edge deployment process. We present ML-EXray, an end-to-end framework, which provides visibility into layer-level details of the ML execution, and helps developers analyze and debug cloud-to-edge deployment issues. More often than not, the reason for sub-optimal edge performance does not only lie in the model itself, but every operation throughout the data flow and the deployment process. Evaluations show that ML-EXray can effectively catch deployment issues, such as pre-processing bugs, quantization issues, suboptimal kernels, etc. Using ML-EXray, users need to write less than 15 lines of code to fully examine the edge deployment pipeline. Eradicating these issues, ML-EXray can correct model performance by up to 30%, pinpoint error-prone layers, and guide users to optimize kernel execution latency by two orders of magnitude. Code and APIs will be released as an open-source multi-lingual instrumentation library and a Python deployment validation library.

Zain Asgar, Michelle Nguyen, Sachin Katti

AI agents are emerging as a dominant workload in a wide range of applications, promising to be the vehicle that delivers the promised benefits of AI to enterprises and consumers. Unlike conventional software or static inference, agentic workloads are dynamic and structurally complex. Often these agents are directed graphs of compute and IO operations that span multi-modal data input and conversion), data processing and context gathering (e.g vector DB lookups), multiple LLM inferences, tool calls, etc. To scale AI agent usage, we need efficient and scalable deployment and agent-serving infrastructure. To tackle this challenge, in this paper, we present a system design for dynamic orchestration of AI agent workloads on heterogeneous compute infrastructure spanning CPUs and accelerators, both from different vendors and across different performance tiers within a single vendor. The system delivers several building blocks: a framework for planning and optimizing agentic AI execution graphs using cost models that account for compute, memory, and bandwidth constraints of different HW; a MLIR based representation and compilation system that can decompose AI agent execution graphs into granular operators and generate code for different HW options; and a dynamic orchestration system that can place the granular components across a heterogeneous compute infrastructure and stitch them together while meeting an end-to-end SLA. Our design performs a systems level TCO optimization and preliminary results show that leveraging a heterogeneous infrastructure can deliver significant TCO benefits. A preliminary surprising finding is that for some workloads a heterogeneous combination of older generation GPUs with newer accelerators can deliver similar TCO as the latest generation homogenous GPU infrastructure design, potentially extending the life of deployed infrastructure.

Taras Sereda, Tom St. John, Burak Bartan et al.

GPU kernels are critical for ML performance but difficult to optimize across diverse accelerators. We present KForge, a platform-agnostic framework built on two collaborative LLM-based agents: a generation agent that produces and iteratively refines programs through compilation and correctness feedback, and a performance analysis agent that interprets profiling data to guide optimization. This agent-based architecture requires only a single-shot example to target new platforms. We make three key contributions: (1) introducing an iterative refinement system where the generation agent and performance analysis agent collaborate through functional and optimization passes, interpreting diverse profiling data (from programmatic APIs to GUI-based tools) to generate actionable recommendations that guide program synthesis for arbitrary accelerators; (2) demonstrating that the generation agent effectively leverages cross-platform knowledge transfer, where a reference implementation from one architecture substantially improves generation quality for different hardware targets; and (3) validating the platform-agnostic nature of our approach by demonstrating effective program synthesis across fundamentally different parallel computing platforms: NVIDIA CUDA and Apple Metal.

Manabu Nakanoya, Junha Im, Hang Qiu et al.

Autonomous vehicles (AVs) must interact with a diverse set of human drivers in heterogeneous geographic areas. Ideally, fleets of AVs should share trajectory data to continually re-train and improve trajectory forecasting models from collective experience using cloud-based distributed learning. At the same time, these robots should ideally avoid uploading raw driver interaction data in order to protect proprietary policies (when sharing insights with other companies) or protect driver privacy from insurance companies. Federated learning (FL) is a popular mechanism to learn models in cloud servers from diverse users without divulging private local data. However, FL is often not robust -- it learns sub-optimal models when user data comes from highly heterogeneous distributions, which is a key hallmark of human-robot interactions. In this paper, we present a novel variant of personalized FL to specialize robust robot learning models to diverse user distributions. Our algorithm outperforms standard FL benchmarks by up to 2x in real user studies that we conducted where human-operated vehicles must gracefully merge lanes with simulated AVs in the standard CARLA and CARLO AV simulators.

Jiangnan Cheng, Marco Pavone, Sachin Katti et al.

Sharing forecasts of network timeseries data, such as cellular or electricity load patterns, can improve independent control applications ranging from traffic scheduling to power generation. Typically, forecasts are designed without knowledge of a downstream controller's task objective, and thus simply optimize for mean prediction error. However, such task-agnostic representations are often too large to stream over a communication network and do not emphasize salient temporal features for cooperative control. This paper presents a solution to learn succinct, highly-compressed forecasts that are co-designed with a modular controller's task objective. Our simulations with real cellular, Internet-of-Things (IoT), and electricity load data show we can improve a model predictive controller's performance by at least $25\%$ while transmitting $80\%$ less data than the competing method. Further, we present theoretical compression results for a networked variant of the classical linear quadratic regulator (LQR) control problem.

Sandeep Chinchali, Evgenya Pergament, Manabu Nakanoya et al.

Today's robotic fleets are increasingly measuring high-volume video and LIDAR sensory streams, which can be mined for valuable training data, such as rare scenes of road construction sites, to steadily improve robotic perception models. However, re-training perception models on growing volumes of rich sensory data in central compute servers (or the "cloud") places an enormous time and cost burden on network transfer, cloud storage, human annotation, and cloud computing resources. Hence, we introduce HarvestNet, an intelligent sampling algorithm that resides on-board a robot and reduces system bottlenecks by only storing rare, useful events to steadily improve perception models re-trained in the cloud. HarvestNet significantly improves the accuracy of machine-learning models on our novel dataset of road construction sites, field testing of self-driving cars, and streaming face recognition, while reducing cloud storage, dataset annotation time, and cloud compute time by between 65.7-81.3%. Further, it is between 1.05-2.58x more accurate than baseline algorithms and scalably runs on embedded deep learning hardware. We provide a suite of compute-efficient perception models for the Google Edge Tensor Processing Unit (TPU), an extended technical report, and a novel video dataset to the research community at https://sites.google.com/view/harvestnet.

Manabu Nakanoya, Sandeep Chinchali, Alexandros Anemogiannis et al.

Today, even the most compute-and-power constrained robots can measure complex, high data-rate video and LIDAR sensory streams. Often, such robots, ranging from low-power drones to space and subterranean rovers, need to transmit high-bitrate sensory data to a remote compute server if they are uncertain or cannot scalably run complex perception or mapping tasks locally. However, today's representations for sensory data are mostly designed for human, not robotic, perception and thus often waste precious compute or wireless network resources to transmit unimportant parts of a scene that are unnecessary for a high-level robotic task. This paper presents an algorithm to learn task-relevant representations of sensory data that are co-designed with a pre-trained robotic perception model's ultimate objective. Our algorithm aggressively compresses robotic sensory data by up to 11x more than competing methods. Further, it achieves high accuracy and robust generalization on diverse tasks including Mars terrain classification with low-power deep learning accelerators, neural motion planning, and environmental timeseries classification.

Eyal Cidon, Evgenya Pergament, Zain Asgar et al.

The same machine learning model running on different edge devices may produce highly-divergent outputs on a nearly-identical input. Possible reasons for the divergence include differences in the device sensors, the device's signal processing hardware and software, and its operating system and processors. This paper presents the first methodical characterization of the variations in model prediction across real-world mobile devices. We demonstrate that accuracy is not a useful metric to characterize prediction divergence, and introduce a new metric, instability, which captures this variation. We characterize different sources for instability, and show that differences in compression formats and image signal processing account for significant instability in object classification models. Notably, in our experiments, 14-17% of images produced divergent classifications across one or more phone models. We evaluate three different techniques for reducing instability. In particular, we adapt prior work on making models robust to noise in order to fine-tune models to be robust to variations across edge devices. We demonstrate our fine-tuning techniques reduce instability by 75%.

Tianshu Chu, Sandeep Chinchali, Sachin Katti

This paper considers multi-agent reinforcement learning (MARL) in networked system control. Specifically, each agent learns a decentralized control policy based on local observations and messages from connected neighbors. We formulate such a networked MARL (NMARL) problem as a spatiotemporal Markov decision process and introduce a spatial discount factor to stabilize the training of each local agent. Further, we propose a new differentiable communication protocol, called NeurComm, to reduce information loss and non-stationarity in NMARL. Based on experiments in realistic NMARL scenarios of adaptive traffic signal control and cooperative adaptive cruise control, an appropriate spatial discount factor effectively enhances the learning curves of non-communicative MARL algorithms, while NeurComm outperforms existing communication protocols in both learning efficiency and control performance.

Sandeep Chinchali, Apoorva Sharma, James Harrison et al.

Today's robotic systems are increasingly turning to computationally expensive models such as deep neural networks (DNNs) for tasks like localization, perception, planning, and object detection. However, resource-constrained robots, like low-power drones, often have insufficient on-board compute resources or power reserves to scalably run the most accurate, state-of-the art neural network compute models. Cloud robotics allows mobile robots the benefit of offloading compute to centralized servers if they are uncertain locally or want to run more accurate, compute-intensive models. However, cloud robotics comes with a key, often understated cost: communicating with the cloud over congested wireless networks may result in latency or loss of data. In fact, sending high data-rate video or LIDAR from multiple robots over congested networks can lead to prohibitive delay for real-time applications, which we measure experimentally. In this paper, we formulate a novel Robot Offloading Problem --- how and when should robots offload sensing tasks, especially if they are uncertain, to improve accuracy while minimizing the cost of cloud communication? We formulate offloading as a sequential decision making problem for robots, and propose a solution using deep reinforcement learning. In both simulations and hardware experiments using state-of-the art vision DNNs, our offloading strategy improves vision task performance by between 1.3-2.6x of benchmark offloading strategies, allowing robots the potential to significantly transcend their on-board sensing accuracy but with limited cost of cloud communication.

Manikanta Kotaru, Guy Satat, Ramesh Raskar et al.

Most computer vision systems and computational photography systems are visible light based which is a small fraction of the electromagnetic (EM) spectrum. In recent years radio frequency (RF) hardware has become more widely available, for example, many cars are equipped with a RADAR, and almost every home has a WiFi device. In the context of imaging, RF spectrum holds many advantages compared to visible light systems. In particular, in this regime, EM energy effectively interacts in different ways with matter. This property allows for many novel applications such as privacy preserving computer vision and imaging through absorbing and scattering materials in visible light such as walls. Here, we expand many of the concepts in computational photography in visible light to RF cameras. The main limitation of imaging with RF is the large wavelength that limits the imaging resolution when compared to visible light. However, the output of RF cameras is usually processed by computer vision and perception algorithms which would benefit from multi-modal sensing of the environment, and from sensing in situations in which visible light systems fail. To bridge the gap between computational photography and RF imaging, we expand the concept of light-field to RF. This work paves the way to novel computational sensing systems with RF.

Assaf Eisenman, Maxim Naumov, Darryl Gardner et al.

Typical large-scale recommender systems use deep learning models that are stored on a large amount of DRAM. These models often rely on embeddings, which consume most of the required memory. We present Bandana, a storage system that reduces the DRAM footprint of embeddings, by using Non-volatile Memory (NVM) as the primary storage medium, with a small amount of DRAM as cache. The main challenge in storing embeddings on NVM is its limited read bandwidth compared to DRAM. Bandana uses two primary techniques to address this limitation: first, it stores embedding vectors that are likely to be read together in the same physical location, using hypergraph partitioning, and second, it decides the number of embedding vectors to cache in DRAM by simulating dozens of small caches. These techniques allow Bandana to increase the effective read bandwidth of NVM by 2-3x and thereby significantly reduce the total cost of ownership.

Manikanta Kotaru, Sachin Katti

Today, experiencing virtual reality (VR) is a cumbersome experience which either requires dedicated infrastructure like infrared cameras to track the headset and hand-motion controllers (e.g., Oculus Rift, HTC Vive), or provides only 3-DoF (Degrees of Freedom) tracking which severely limits the user experience (e.g., Samsung Gear). To truly enable VR everywhere, we need position tracking to be available as a ubiquitous service. This paper presents WiCapture, a novel approach which leverages commodity WiFi infrastructure, which is ubiquitous today, for tracking purposes. We prototype WiCapture using off-the-shelf WiFi radios and show that it achieves an accuracy of 0.88 cm compared to sophisticated infrared based tracking systems like the Oculus, while providing much higher range, resistance to occlusion, ubiquity and ease of deployment.