Search

Photonics is a promising technology to accelerate Deep Neural Networks as it can use optical interconnects to reduce data movement energy and it enables low-energy, high-throughput optical-analog computations. To realize these benefits in a full system (accelerator + DRAM), designers must ensure that the benefits of using the electrical, optical, analog, and digital domains exceed the costs of converting data between domains. Designers must also consider system-level energy costs such as data fetch from DRAM. Converting data and accessing DRAM can consume significant energy, so to evaluate and explore the photonic system space, there is a need for a tool that can model these full-system considerations. In this work, we show that similarities between Compute-in-Memory (CiM) and photonics let us use CiM system modeling tools to accurately model photonics systems. Bringing modeling tools to photonics enables evaluation of photonic research in a full-system context, rapid design space exploration, co-design, and comparison between systems. Using our open-source model, we show that cross-domain conversion and DRAM can consume a significant portion of photonic system energy. We then demonstrate optimizations that reduce conversions and DRAM accesses to improve photonic system energy efficiency by up to 3x., Comment: Published at ISPASS 2024

Compute-In-Memory (CiM) is a promising solution to accelerate Deep Neural Networks (DNNs) as it can avoid energy-intensive DNN weight movement and use memory arrays to perform low-energy, high-density computations. These benefits have inspired research across the CiM stack, but CiM research often focuses on only one level of the stack (i.e., devices, circuits, architecture, workload, or mapping) or only one design point (e.g., one fabricated chip). There is a need for a full-stack modeling tool to evaluate design decisions in the context of full systems (e.g., see how a circuit impacts system energy) and to perform rapid early-stage exploration of the CiM co-design space. To address this need, we propose CiMLoop: an open-source tool to model diverse CiM systems and explore decisions across the CiM stack. CiMLoop introduces (1) a flexible specification that lets users describe, model, and map workloads to both circuits and architecture, (2) an accurate energy model that captures the interaction between DNN operand values, hardware data representations, and analog/digital values propagated by circuits, and (3) a fast statistical model that can explore the design space orders-of-magnitude more quickly than other high-accuracy models. Using CiMLoop, researchers can evaluate design choices at different levels of the CiM stack, co-design across all levels, fairly compare different implementations, and rapidly explore the design space., Comment: Available at https://github.com/mit-emze/cimloop. Published in ISPASS 2024

Analog Compute-in-Memory (CiM) accelerators use analog-digital converters (ADCs) to read the analog values that they compute. ADCs can consume significant energy and area, so architecture-level ADC decisions such as ADC resolution or number of ADCs can significantly impact overall CiM accelerator energy and area. Therefore, modeling how architecture-level decisions affect ADC energy and area is critical for performing architecture-level design space exploration of CiM accelerators. This work presents an open-source architecture-level model to estimate ADC energy and area. To enable fast design space exploration, the model uses only architecture-level attributes while abstracting circuit-level details. Our model enables researchers to quickly and easily model key architecture-level tradeoffs in accelerators that use ADCs.

Sparse tensor algebra is a challenging class of workloads to accelerate due to low arithmetic intensity and varying sparsity patterns. Prior sparse tensor algebra accelerators have explored tiling sparse data to increase exploitable data reuse and improve throughput, but typically allocate tile size in a given buffer for the worst-case data occupancy. This severely limits the utilization of available memory resources and reduces data reuse. Other accelerators employ complex tiling during preprocessing or at runtime to determine the exact tile size based on its occupancy. This paper proposes a speculative tensor tiling approach, called overbooking, to improve buffer utilization by taking advantage of the distribution of nonzero elements in sparse tensors to construct larger tiles with greater data reuse. To ensure correctness, we propose a low-overhead hardware mechanism, Tailors, that can tolerate data overflow by design while ensuring reasonable data reuse. We demonstrate that Tailors can be easily integrated into the memory hierarchy of an existing sparse tensor algebra accelerator. To ensure high buffer utilization with minimal tiling overhead, we introduce a statistical approach, Swiftiles, to pick a tile size so that tiles usually fit within the buffer's capacity, but can potentially overflow, i.e., it overbooks the buffers. Across a suite of 22 sparse tensor algebra workloads, we show that our proposed overbooking strategy introduces an average speedup of $52.7\times$ and $2.3\times$ and an average energy reduction of $22.5\times$ and $2.5\times$ over ExTensor without and with optimized tiling, respectively., Comment: 17 pages, 13 figures, in MICRO 2023

Energy consumption of memory accesses dominates the compute energy in energy-constrained robots which require a compact 3D map of the environment to achieve autonomy. Recent mapping frameworks only focused on reducing the map size while incurring significant memory usage during map construction due to multi-pass processing of each depth image. In this work, we present a memory-efficient continuous occupancy map, named GMMap, that accurately models the 3D environment using a Gaussian Mixture Model (GMM). Memory-efficient GMMap construction is enabled by the single-pass compression of depth images into local GMMs which are directly fused together into a globally-consistent map. By extending Gaussian Mixture Regression to model unexplored regions, occupancy probability is directly computed from Gaussians. Using a low-power ARM Cortex A57 CPU, GMMap can be constructed in real-time at up to 60 images per second. Compared with prior works, GMMap maintains high accuracy while reducing the map size by at least 56%, memory overhead by at least 88%, DRAM access by at least 78%, and energy consumption by at least 69%. Thus, GMMap enables real-time 3D mapping on energy-constrained robots., Comment: 17 pages, 12 figures, 3 tables

Due to complex interactions among various deep neural network (DNN) optimization techniques, modern DNNs can have weights and activations that are dense or sparse with diverse sparsity degrees. To offer a good trade-off between accuracy and hardware performance, an ideal DNN accelerator should have high flexibility to efficiently translate DNN sparsity into reductions in energy and/or latency without incurring significant complexity overhead. This paper introduces hierarchical structured sparsity (HSS), with the key insight that we can systematically represent diverse sparsity degrees by having them hierarchically composed from multiple simple sparsity patterns. As a result, HSS simplifies the underlying hardware since it only needs to support simple sparsity patterns; this significantly reduces the sparsity acceleration overhead, which improves efficiency. Motivated by such opportunities, we propose a simultaneously efficient and flexible accelerator, named HighLight, to accelerate DNNs that have diverse sparsity degrees (including dense). Due to the flexibility of HSS, different HSS patterns can be introduced to DNNs to meet different applications' accuracy requirements. Compared to existing works, HighLight achieves a geomean of up to 6.4x better energy-delay product (EDP) across workloads with diverse sparsity degrees, and always sits on the EDP-accuracy Pareto frontier for representative DNNs, Comment: Accepted to MICRO23

Processing-In-Memory (PIM) accelerators have the potential to efficiently run Deep Neural Network (DNN) inference by reducing costly data movement and by using resistive RAM (ReRAM) for efficient analog compute. Unfortunately, overall PIM accelerator efficiency is limited by energy-intensive analog-to-digital converters (ADCs). Furthermore, existing accelerators that reduce ADC cost do so by changing DNN weights or by using low-resolution ADCs that reduce output fidelity. These strategies harm DNN accuracy and/or require costly DNN retraining to compensate. To address these issues, we propose the RAELLA architecture. RAELLA adapts the architecture to each DNN; it lowers the resolution of computed analog values by encoding weights to produce near-zero analog values, adaptively slicing weights for each DNN layer, and dynamically slicing inputs through speculation and recovery. Low-resolution analog values allow RAELLA to both use efficient low-resolution ADCs and maintain accuracy without retraining, all while computing with fewer ADC converts. Compared to other low-accuracy-loss PIM accelerators, RAELLA increases energy efficiency by up to 4.9$\times$ and throughput by up to 3.3$\times$. Compared to PIM accelerators that cause accuracy loss and retrain DNNs to recover, RAELLA achieves similar efficiency and throughput without expensive DNN retraining., Comment: 16 pages; 15 figures; Accepted at ISCA 2023 (the International Symposium on Computer Architecture)

Exploration tasks are essential to many emerging robotics applications, ranging from search and rescue to space exploration. The planning problem for exploration requires determining the best locations for future measurements that will enhance the fidelity of the map, for example, by reducing its total entropy. A widely-studied technique involves computing the Mutual Information (MI) between the current map and future measurements, and utilizing this MI metric to decide the locations for future measurements. However, computing MI for reasonably-sized maps is slow and power hungry, which has been a bottleneck towards fast and efficient robotic exploration. In this paper, we introduce a new hardware accelerator architecture for MI computation that features a low-latency, energy-efficient MI compute core and an optimized memory subsystem that provides sufficient bandwidth to keep the cores fully utilized. The core employs interleaving to counter the recursive algorithm, and workload balancing and numerical approximations to reduce latency and energy consumption. We demonstrate this optimized architecture with a Field-Programmable Gate Array (FPGA) implementation, which can compute MI for all cells in an entire 201-by-201 occupancy grid ({\em e.g.}, representing a 20.1m-by-20.1m map at 0.1m resolution) in 1.55 ms while consuming 1.7 mJ of energy, thus finally rendering MI computation for the whole map real time and at a fraction of the energy cost of traditional compute platforms. For comparison, this particular FPGA implementation running on the Xilinx Zynq-7000 platform is two orders of magnitude faster and consumes three orders of magnitude less energy per MI map compute, when compared to a baseline GPU implementation running on an NVIDIA GeForce GTX 980 platform. The improvements are more pronounced when compared to CPU implementations of equivalent algorithms.

Video conferencing systems suffer from poor user experience when network conditions deteriorate because current video codecs simply cannot operate at extremely low bitrates. Recently, several neural alternatives have been proposed that reconstruct talking head videos at very low bitrates using sparse representations of each frame such as facial landmark information. However, these approaches produce poor reconstructions in scenarios with major movement or occlusions over the course of a call, and do not scale to higher resolutions. We design Gemino, a new neural compression system for video conferencing based on a novel high-frequency-conditional super-resolution pipeline. Gemino upsamples a very low-resolution version of each target frame while enhancing high-frequency details (e.g., skin texture, hair, etc.) based on information extracted from a single high-resolution reference image. We use a multi-scale architecture that runs different components of the model at different resolutions, allowing it to scale to resolutions comparable to 720p, and we personalize the model to learn specific details of each person, achieving much better fidelity at low bitrates. We implement Gemino atop aiortc, an open-source Python implementation of WebRTC, and show that it operates on 1024x1024 videos in real-time on a Titan X GPU, and achieves 2.2-5x lower bitrate than traditional video codecs for the same perceptual quality., Comment: 13 pages, 5 appendix

Through a series of federal initiatives and orders, the U.S. Government has been making a concerted effort to ensure American leadership in AI. These broad strategy documents have influenced organizations such as the United States Department of the Air Force (DAF). The DAF-MIT AI Accelerator is an initiative between the DAF and MIT to bridge the gap between AI researchers and DAF mission requirements. Several projects supported by the DAF-MIT AI Accelerator are developing public challenge problems that address numerous Federal AI research priorities. These challenges target priorities by making large, AI-ready datasets publicly available, incentivizing open-source solutions, and creating a demand signal for dual use technologies that can stimulate further research. In this article, we describe these public challenges being developed and how their application contributes to scientific advances.

In recent years, many accelerators have been proposed to efficiently process sparse tensor algebra applications (e.g., sparse neural networks). However, these proposals are single points in a large and diverse design space. The lack of systematic description and modeling support for these sparse tensor accelerators impedes hardware designers from efficient and effective design space exploration. This paper first presents a unified taxonomy to systematically describe the diverse sparse tensor accelerator design space. Based on the proposed taxonomy, it then introduces Sparseloop, the first fast, accurate, and flexible analytical modeling framework to enable early-stage evaluation and exploration of sparse tensor accelerators. Sparseloop comprehends a large set of architecture specifications, including various dataflows and sparse acceleration features (e.g., elimination of zero-based compute). Using these specifications, Sparseloop evaluates a design's processing speed and energy efficiency while accounting for data movement and compute incurred by the employed dataflow as well as the savings and overhead introduced by the sparse acceleration features using stochastic tensor density models. Across representative accelerators and workloads, Sparseloop achieves over 2000 times faster modeling speed than cycle-level simulations, maintains relative performance trends, and achieves 0.1% to 8% average error. With a case study, we demonstrate Sparseloop's ability to help reveal important insights for designing sparse tensor accelerators (e.g., it is important to co-design orthogonal design aspects)., Comment: Update website link, update UOP format description

Vision Transformer (ViT) demonstrates that Transformer for natural language processing can be applied to computer vision tasks and result in comparable performance to convolutional neural networks (CNN), which have been studied and adopted in computer vision for years. This naturally raises the question of how the performance of ViT can be advanced with design techniques of CNN. To this end, we propose to incorporate two techniques and present ViT-ResNAS, an efficient multi-stage ViT architecture designed with neural architecture search (NAS). First, we propose residual spatial reduction to decrease sequence lengths for deeper layers and utilize a multi-stage architecture. When reducing lengths, we add skip connections to improve performance and stabilize training deeper networks. Second, we propose weight-sharing NAS with multi-architectural sampling. We enlarge a network and utilize its sub-networks to define a search space. A super-network covering all sub-networks is then trained for fast evaluation of their performance. To efficiently train the super-network, we propose to sample and train multiple sub-networks with one forward-backward pass. After that, evolutionary search is performed to discover high-performance network architectures. Experiments on ImageNet demonstrate that ViT-ResNAS achieves better accuracy-MACs and accuracy-throughput trade-offs than the original DeiT and other strong baselines of ViT. Code is available at https://github.com/yilunliao/vit-search.

Neural architecture search (NAS) typically consists of three main steps: training a super-network, training and evaluating sampled deep neural networks (DNNs), and training the discovered DNN. Most of the existing efforts speed up some steps at the cost of a significant slowdown of other steps or sacrificing the support of non-differentiable search metrics. The unbalanced reduction in the time spent per step limits the total search time reduction, and the inability to support non-differentiable search metrics limits the performance of discovered DNNs. In this paper, we present NetAdaptV2 with three innovations to better balance the time spent for each step while supporting non-differentiable search metrics. First, we propose channel-level bypass connections that merge network depth and layer width into a single search dimension to reduce the time for training and evaluating sampled DNNs. Second, ordered dropout is proposed to train multiple DNNs in a single forward-backward pass to decrease the time for training a super-network. Third, we propose the multi-layer coordinate descent optimizer that considers the interplay of multiple layers in each iteration of optimization to improve the performance of discovered DNNs while supporting non-differentiable search metrics. With these innovations, NetAdaptV2 reduces the total search time by up to $5.8\times$ on ImageNet and $2.4\times$ on NYU Depth V2, respectively, and discovers DNNs with better accuracy-latency/accuracy-MAC trade-offs than state-of-the-art NAS works. Moreover, the discovered DNN outperforms NAS-discovered MobileNetV3 by 1.8% higher top-1 accuracy with the same latency. The project website is http://netadapt.mit.edu., Comment: Accepted by CVPR 2021

We aid in neurocognitive monitoring outside the hospital environment by enabling app-based measurements of visual reaction time (saccade latency) and error rate in a cohort of subjects spanning the adult age spectrum. Methods: We developed an iOS app to record subjects with the frontal camera during pro- and anti-saccade tasks. We further developed automated algorithms for measuring saccade latency and error rate that take into account the possibility that it might not always be possible to determine the eye movement from app-based recordings. Results: To measure saccade latency on a tablet, we ensured that the absolute timing error between on-screen task presentation and the camera recording is within 5 ms. We collected over 235,000 eye movements in 80 subjects ranging in age from 20 to 92 years, with 96% of recorded eye movements either declared good or directional errors. Our error detection code achieved a sensitivity of 0.97 and a specificity of 0.97. Confirming prior reports, we observed a positive correlation between saccade latency and age while the relationship between error rate and age was not significant. Finally, we observed significant intra- and inter-subject variations in saccade latency and error rate distributions, which highlights the importance of individualized tracking of these visual digital biomarkers. Conclusion and Significance: Our system and algorithms allow ubiquitous tracking of saccade latency and error rate, which opens up the possibility of quantifying patient state on a finer timescale in a broader population than previously possible., Comment: 11 pages, 16 figures, journal

As deep neural network (DNN) models grow ever-larger, they can achieve higher accuracy and solve more complex problems. This trend has been enabled by an increase in available compute power; however, efforts to continue to scale electronic processors are impeded by the costs of communication, thermal management, power delivery and clocking. To improve scalability, we propose a digital optical neural network (DONN) with intralayer optical interconnects and reconfigurable input values. The near path-length-independence of optical energy consumption enables information locality between a transmitter and arbitrarily arranged receivers, which allows greater flexibility in architecture design to circumvent scaling limitations. In a proof-of-concept experiment, we demonstrate optical multicast in the classification of 500 MNIST images with a 3-layer, fully-connected network. We also analyze the energy consumption of the DONN and find that optical data transfer is beneficial over electronics when the spacing of computational units is on the order of >10 micrometers., Comment: 19 pages (15 main and 4 supplementary), 11 figures (6 main and 5 supplementary)

Depth information is useful for many applications. Active depth sensors are appealing because they obtain dense and accurate depth maps. However, due to issues that range from power constraints to multi-sensor interference, these sensors cannot always be continuously used. To overcome this limitation, we propose an algorithm that estimates depth maps using concurrently collected images and a previously measured depth map for dynamic scenes, where both the camera and objects in the scene may be independently moving. To estimate depth in these scenarios, our algorithm models the dynamic scene motion using independent and rigid motions. It then uses the previous depth map to efficiently estimate these rigid motions and obtain a new depth map. Our goal is to balance the acquisition of depth between the active depth sensor and computation, without incurring a large computational cost. Thus, we leverage the prior depth information to avoid computationally expensive operations like dense optical flow estimation or segmentation used in similar approaches. Our approach can obtain dense depth maps at up to real-time (30 FPS) on a standard laptop computer, which is orders of magnitude faster than similar approaches. When evaluated using RGB-D datasets of various dynamic scenes, our approach estimates depth maps with a mean relative error of 2.5% while reducing the active depth sensor usage by over 90%.

This paper describes various design considerations for deep neural networks that enable them to operate efficiently and accurately on processing-in-memory accelerators. We highlight important properties of these accelerators and the resulting design considerations using experiments conducted on various state-of-the-art deep neural networks with the large-scale ImageNet dataset., Comment: Accepted by IEDM 2019

Exploration tasks are embedded in many robotics applications, such as search and rescue and space exploration. Information-based exploration algorithms aim to find the most informative trajectories by maximizing an information-theoretic metric, such as the mutual information between the map and potential future measurements. Unfortunately, most existing information-based exploration algorithms are plagued by the computational difficulty of evaluating the Shannon mutual information metric. In this paper, we consider the fundamental problem of evaluating Shannon mutual information between the map and a range measurement. First, we consider 2D environments. We propose a novel algorithm, called the Fast Shannon Mutual Information (FSMI). The key insight behind the algorithm is that a certain integral can be computed analytically, leading to substantial computational savings. Second, we consider 3D environments, represented by efficient data structures, e.g., an OctoMap, such that the measurements are compressed by Run-Length Encoding (RLE). We propose a novel algorithm, called FSMI-RLE, that efficiently evaluates the Shannon mutual information when the measurements are compressed using RLE. For both the FSMI and the FSMI-RLE, we also propose variants that make different assumptions on the sensor noise distribution for the purpose of further computational savings. We evaluate the proposed algorithms in extensive experiments. In particular, we show that the proposed algorithms outperform existing algorithms that compute Shannon mutual information as well as other algorithms that compute the Cauchy-Schwarz Quadratic mutual information (CSQMI). In addition, we demonstrate the computation of Shannon mutual information on a 3D map for the first time.

Machine learning (ML) techniques are enjoying rapidly increasing adoption. However, designing and implementing the systems that support ML models in real-world deployments remains a significant obstacle, in large part due to the radically different development and deployment profile of modern ML methods, and the range of practical concerns that come with broader adoption. We propose to foster a new systems machine learning research community at the intersection of the traditional systems and ML communities, focused on topics such as hardware systems for ML, software systems for ML, and ML optimized for metrics beyond predictive accuracy. To do this, we describe a new conference, MLSys, that explicitly targets research at the intersection of systems and machine learning with a program committee split evenly between experts in systems and ML, and an explicit focus on topics at the intersection of the two.

Depth sensing is a critical function for robotic tasks such as localization, mapping and obstacle detection. There has been a significant and growing interest in depth estimation from a single RGB image, due to the relatively low cost and size of monocular cameras. However, state-of-the-art single-view depth estimation algorithms are based on fairly complex deep neural networks that are too slow for real-time inference on an embedded platform, for instance, mounted on a micro aerial vehicle. In this paper, we address the problem of fast depth estimation on embedded systems. We propose an efficient and lightweight encoder-decoder network architecture and apply network pruning to further reduce computational complexity and latency. In particular, we focus on the design of a low-latency decoder. Our methodology demonstrates that it is possible to achieve similar accuracy as prior work on depth estimation, but at inference speeds that are an order of magnitude faster. Our proposed network, FastDepth, runs at 178 fps on an NVIDIA Jetson TX2 GPU and at 27 fps when using only the TX2 CPU, with active power consumption under 10 W. FastDepth achieves close to state-of-the-art accuracy on the NYU Depth v2 dataset. To the best of the authors' knowledge, this paper demonstrates real-time monocular depth estimation using a deep neural network with the lowest latency and highest throughput on an embedded platform that can be carried by a micro aerial vehicle., Comment: Accepted for presentation at ICRA 2019. 8 pages, 6 figures, 7 tables

We present a single-shot, bottom-up approach for whole image parsing. Whole image parsing, also known as Panoptic Segmentation, generalizes the tasks of semantic segmentation for 'stuff' classes and instance segmentation for 'thing' classes, assigning both semantic and instance labels to every pixel in an image. Recent approaches to whole image parsing typically employ separate standalone modules for the constituent semantic and instance segmentation tasks and require multiple passes of inference. Instead, the proposed DeeperLab image parser performs whole image parsing with a significantly simpler, fully convolutional approach that jointly addresses the semantic and instance segmentation tasks in a single-shot manner, resulting in a streamlined system that better lends itself to fast processing. For quantitative evaluation, we use both the instance-based Panoptic Quality (PQ) metric and the proposed region-based Parsing Covering (PC) metric, which better captures the image parsing quality on 'stuff' classes and larger object instances. We report experimental results on the challenging Mapillary Vistas dataset, in which our single model achieves 31.95% (val) / 31.6% PQ (test) and 55.26% PC (val) with 3 frames per second (fps) on GPU or near real-time speed (22.6 fps on GPU) with reduced accuracy., Comment: 20 pages. The code of the proposed Parsing Covering metric is available at http://deeperlab.mit.edu

Depth sensing is useful in a variety of applications that range from augmented reality to robotics. Time-of-flight (TOF) cameras are appealing because they obtain dense depth measurements with minimal latency. However, for many battery-powered devices, the illumination source of a TOF camera is power hungry and can limit the battery life of the device. To address this issue, we present an algorithm that lowers the power for depth sensing by reducing the usage of the TOF camera and estimating depth maps using concurrently collected images. Our technique also adaptively controls the TOF camera and enables it when an accurate depth map cannot be estimated. To ensure that the overall system power for depth sensing is reduced, we design our algorithm to run on a low power embedded platform, where it outputs 640x480 depth maps at 30 frames per second. We evaluate our approach on several RGB-D datasets, where it produces depth maps with an overall mean relative error of 0.96% and reduces the usage of the TOF camera by 85%. When used with commercial TOF cameras, we estimate that our algorithm can lower the total power for depth sensing by up to 73%.

This paper presents Navion, an energy-efficient accelerator for visual-inertial odometry (VIO) that enables autonomous navigation of miniaturized robots (e.g., nano drones), and virtual/augmented reality on portable devices. The chip uses inertial measurements and mono/stereo images to estimate the drone's trajectory and a 3D map of the environment. This estimate is obtained by running a state-of-the-art VIO algorithm based on non-linear factor graph optimization, which requires large irregularly structured memories and heterogeneous computation flow. To reduce the energy consumption and footprint, the entire VIO system is fully integrated on chip to eliminate costly off-chip processing and storage. This work uses compression and exploits both structured and unstructured sparsity to reduce on-chip memory size by 4.1$\times$. Parallelism is used under tight area constraints to increase throughput by 43%. The chip is fabricated in 65nm CMOS, and can process 752$\times$480 stereo images from EuRoC dataset in real-time at 20 frames per second (fps) consuming only an average power of 2mW. At its peak performance, Navion can process stereo images at up to 171 fps and inertial measurements at up to 52 kHz, while consuming an average of 24mW. The chip is configurable to maximize accuracy, throughput and energy-efficiency trade-offs and to adapt to different environments. To the best of our knowledge, this is the first fully integrated VIO system in an ASIC.

A recent trend in DNN development is to extend the reach of deep learning applications to platforms that are more resource and energy constrained, e.g., mobile devices. These endeavors aim to reduce the DNN model size and improve the hardware processing efficiency, and have resulted in DNNs that are much more compact in their structures and/or have high data sparsity. These compact or sparse models are different from the traditional large ones in that there is much more variation in their layer shapes and sizes, and often require specialized hardware to exploit sparsity for performance improvement. Thus, many DNN accelerators designed for large DNNs do not perform well on these models. In this work, we present Eyeriss v2, a DNN accelerator architecture designed for running compact and sparse DNNs. To deal with the widely varying layer shapes and sizes, it introduces a highly flexible on-chip network, called hierarchical mesh, that can adapt to the different amounts of data reuse and bandwidth requirements of different data types, which improves the utilization of the computation resources. Furthermore, Eyeriss v2 can process sparse data directly in the compressed domain for both weights and activations, and therefore is able to improve both processing speed and energy efficiency with sparse models. Overall, with sparse MobileNet, Eyeriss v2 in a 65nm CMOS process achieves a throughput of 1470.6 inferences/sec and 2560.3 inferences/J at a batch size of 1, which is 12.6x faster and 2.5x more energy efficient than the original Eyeriss running MobileNet. We also present an analysis methodology called Eyexam that provides a systematic way of understanding the performance limits for DNN processors as a function of specific characteristics of the DNN model and accelerator design; it applies these characteristics as sequential steps to increasingly tighten the bound on the performance limits., Comment: accepted for publication in IEEE Journal on Emerging and Selected Topics in Circuits and Systems. This extended version on arXiv also includes Eyexam in the appendix

This work proposes an algorithm, called NetAdapt, that automatically adapts a pre-trained deep neural network to a mobile platform given a resource budget. While many existing algorithms simplify networks based on the number of MACs or weights, optimizing those indirect metrics may not necessarily reduce the direct metrics, such as latency and energy consumption. To solve this problem, NetAdapt incorporates direct metrics into its adaptation algorithm. These direct metrics are evaluated using empirical measurements, so that detailed knowledge of the platform and toolchain is not required. NetAdapt automatically and progressively simplifies a pre-trained network until the resource budget is met while maximizing the accuracy. Experiment results show that NetAdapt achieves better accuracy versus latency trade-offs on both mobile CPU and mobile GPU, compared with the state-of-the-art automated network simplification algorithms. For image classification on the ImageNet dataset, NetAdapt achieves up to a 1.7$\times$ speedup in measured inference latency with equal or higher accuracy on MobileNets (V1&V2)., Comment: Accepted by ECCV 2018

Deep neural networks (DNNs) are currently widely used for many artificial intelligence (AI) applications including computer vision, speech recognition, and robotics. While DNNs deliver state-of-the-art accuracy on many AI tasks, it comes at the cost of high computational complexity. Accordingly, techniques that enable efficient processing of DNNs to improve energy efficiency and throughput without sacrificing application accuracy or increasing hardware cost are critical to the wide deployment of DNNs in AI systems. This article aims to provide a comprehensive tutorial and survey about the recent advances towards the goal of enabling efficient processing of DNNs. Specifically, it will provide an overview of DNNs, discuss various hardware platforms and architectures that support DNNs, and highlight key trends in reducing the computation cost of DNNs either solely via hardware design changes or via joint hardware design and DNN algorithm changes. It will also summarize various development resources that enable researchers and practitioners to quickly get started in this field, and highlight important benchmarking metrics and design considerations that should be used for evaluating the rapidly growing number of DNN hardware designs, optionally including algorithmic co-designs, being proposed in academia and industry. The reader will take away the following concepts from this article: understand the key design considerations for DNNs; be able to evaluate different DNN hardware implementations with benchmarks and comparison metrics; understand the trade-offs between various hardware architectures and platforms; be able to evaluate the utility of various DNN design techniques for efficient processing; and understand recent implementation trends and opportunities., Comment: Based on tutorial on DNN Hardware at eyeriss.mit.edu/tutorial.html

Computer vision enables a wide range of applications in robotics/drones, self-driving cars, smart Internet of Things, and portable/wearable electronics. For many of these applications, local embedded processing is preferred due to privacy and/or latency concerns. Accordingly, energy-efficient embedded vision hardware delivering real-time and robust performance is crucial. While deep learning is gaining popularity in several computer vision algorithms, a significant energy consumption difference exists compared to traditional hand-crafted approaches. In this paper, we provide an in-depth analysis of the computation, energy and accuracy trade-offs between learned features such as deep Convolutional Neural Networks (CNN) and hand-crafted features such as Histogram of Oriented Gradients (HOG). This analysis is supported by measurements from two chips that implement these algorithms. Our goal is to understand the source of the energy discrepancy between the two approaches and to provide insight about the potential areas where CNNs can be improved and eventually approach the energy-efficiency of HOG while maintaining its outstanding performance accuracy.

Machine learning plays a critical role in extracting meaningful information out of the zetabytes of sensor data collected every day. For some applications, the goal is to analyze and understand the data to identify trends (e.g., surveillance, portable/wearable electronics); in other applications, the goal is to take immediate action based the data (e.g., robotics/drones, self-driving cars, smart Internet of Things). For many of these applications, local embedded processing near the sensor is preferred over the cloud due to privacy or latency concerns, or limitations in the communication bandwidth. However, at the sensor there are often stringent constraints on energy consumption and cost in addition to throughput and accuracy requirements. Furthermore, flexibility is often required such that the processing can be adapted for different applications or environments (e.g., update the weights and model in the classifier). In many applications, machine learning often involves transforming the input data into a higher dimensional space, which, along with programmable weights, increases data movement and consequently energy consumption. In this paper, we will discuss how these challenges can be addressed at various levels of hardware design ranging from architecture, hardware-friendly algorithms, mixed-signal circuits, and advanced technologies (including memories and sensors)., Comment: Published as an invited conference paper at CICC 2017

Deep convolutional neural networks (CNNs) are indispensable to state-of-the-art computer vision algorithms. However, they are still rarely deployed on battery-powered mobile devices, such as smartphones and wearable gadgets, where vision algorithms can enable many revolutionary real-world applications. The key limiting factor is the high energy consumption of CNN processing due to its high computational complexity. While there are many previous efforts that try to reduce the CNN model size or amount of computation, we find that they do not necessarily result in lower energy consumption, and therefore do not serve as a good metric for energy cost estimation. To close the gap between CNN design and energy consumption optimization, we propose an energy-aware pruning algorithm for CNNs that directly uses energy consumption estimation of a CNN to guide the pruning process. The energy estimation methodology uses parameters extrapolated from actual hardware measurements that target realistic battery-powered system setups. The proposed layer-by-layer pruning algorithm also prunes more aggressively than previously proposed pruning methods by minimizing the error in output feature maps instead of filter weights. For each layer, the weights are first pruned and then locally fine-tuned with a closed-form least-square solution to quickly restore the accuracy. After all layers are pruned, the entire network is further globally fine-tuned using back-propagation. With the proposed pruning method, the energy consumption of AlexNet and GoogLeNet are reduced by 3.7x and 1.6x, respectively, with less than 1% top-5 accuracy loss. Finally, we show that pruning the AlexNet with a reduced number of target classes can greatly decrease the number of weights but the energy reduction is limited. Energy modeling tool and energy-aware pruned models available at http://eyeriss.mit.edu/energy.html, Comment: Published as a conference paper at CVPR 2017

This paper presents a programmable, energy-efficient and real-time object detection accelerator using deformable parts models (DPM), with 2x higher accuracy than traditional rigid body models. With 8 deformable parts detection, three methods are used to address the high computational complexity: classification pruning for 33x fewer parts classification, vector quantization for 15x memory size reduction, and feature basis projection for 2x reduction of the cost of each classification. The chip is implemented in 65nm CMOS technology, and can process HD (1920x1080) images at 30fps without any off-chip storage while consuming only 58.6mW (0.94nJ/pixel, 1168 GOPS/W). The chip has two classification engines to simultaneously detect two different classes of objects. With a tested high throughput of 60fps, the classification engines can be time multiplexed to detect even more than two object classes. It is energy scalable by changing the pruning factor or disabling the parts classification.

State-of-the-art super-resolution (SR) algorithms require significant computational resources to achieve real-time throughput (e.g., 60Mpixels/s for HD video). This paper introduces FAST (Free Adaptive Super-resolution via Transfer), a framework to accelerate any SR algorithm applied to compressed videos. FAST exploits the temporal correlation between adjacent frames such that SR is only applied to a subset of frames; SR pixels are then transferred to the other frames. The transferring process has negligible computation cost as it uses information already embedded in the compressed video (e.g., motion vectors and residual). Adaptive processing is used to retain accuracy when the temporal correlation is not present (e.g., occlusions). FAST accelerates state-of-the-art SR algorithms by up to 15x with a visual quality loss of 0.2dB. FAST is an important step towards real-time SR algorithms for ultra-HD displays and energy constrained devices (e.g., phones and tablets).