Binarized Neural Networks (BNNs) can significantly reduce the inference latency and energy consumption in resource-constrained devices due to their pure-logical computation fewer memory accesses. However, training BNNs is difficult since activation flow encounters degeneration, saturation, gradient mismatch problems. Prior work alleviates these issues by increasing bits adding floating-point scaling factors, thereby sacrificing BNN's efficiency. In this paper, we propose use distribution...

Pruning convolutional filters has demonstrated its effectiveness in compressing ConvNets. Prior art filter pruning requires users to specify a target model complexity (e.g., size or FLOP count) for the resulting architecture. However, determining can be difficult optimizing various embodied AI applications such as autonomous robots, drones, and user-facing applications. First, both accuracy speed of ConvNets affect performance application. Second, application hard assess without evaluating...

In vision-enabled autonomous systems such as robots and cars, video object detection plays a crucial role, both its speed accuracy are important factors to provide reliable operation. The key insight we show in this paper is that not necessarily trade-off when it comes image scaling. Our results re-scaling the lower resolution will sometimes produce better accuracy. Based on observation, propose novel approach, dubbed AdaScale, which adaptively selects input scale improves for detection. To...

Can we reduce the search cost of Neural Architecture Search (NAS) from days down to only few hours? NAS methods automate design Convolutional Networks (ConvNets) under hardware constraints and they have emerged as key components AutoML frameworks. However, problem remains challenging due combinatorially large space significant time (at least 200 GPU-hours). In this work, alleviate less than 3 hours, while achieving state-of-the-art image classification results mobile latency constraints. We...

Deep Neural Networks (DNNs) have been adopted in many systems because of their higher classification accuracy, with custom hardware implementations great candidates for high-speed, accurate inference. While progress achieving large scale, highly DNNs has made, significant energy and area are required due to massive memory accesses computations. Such demands pose a challenge any DNN implementation, yet it is more natural handle platform. To alleviate the increased demand storage energy,...

Application-specific integrated circuit (ASIC) implementations for Deep Neural Networks (DNNs) have been adopted in many systems because of their higher classification speed. However, although they may be characterized by better accuracy, larger DNNs require significant energy and area, thereby limiting wide adoption. The consumption is driven both memory accesses computation. Binarized (BNNs), as a trade-off between accuracy consumption, can achieve great reduction, good large due to its...

Deep Neural Networks (DNNs) have been adopted in many systems because of their higher classification accuracy, with custom hardware implementations great candidates for high-speed, accurate inference. While progress achieving large scale, highly DNNs has made, significant energy and area are required due to massive memory accesses computations. Such demands pose a challenge any DNN implementation, yet it is more natural handle platform. To alleviate the increased demand storage energy,...

In recent years, supervised learning using Convolutional Neural Networks (CNNs) has achieved great success in image classification tasks, and large scale labeled datasets have contributed significantly to this achievement. However, the definition of a label is often application dependent. For example, an cat can be as "cat" or perhaps more specifically "Persian cat." We refer granularity. paper, we conduct extensive experiments various demonstrate analyze how why training based on fine-grain...

To improve the throughput and energy efficiency of Deep Neural Networks (DNNs) on customized hardware, lightweight neural networks constrain weights DNNs to be a limited combination (denoted as k ϵ {1, 2}) powers 2. In such networks, multiply-accumulate operation can replaced with single shift operation, or two shifts an add operation. provide even more design flexibility, for each convolutional filter optimally chosen instead being fixed every filter. this paper, we formulate selection...

Hardware implementations of deep neural networks (DNNs) have been adopted in many systems because their higher classification speed. However, while they may be characterized by better accuracy, larger DNNs require significant energy and area, thereby limiting wide adoption. The consumption is driven both memory accesses computation. Binarized (BNNs), as a tradeoff between accuracy consumption, can achieve great reduction good for large due to regularization effect. BNNs show poor when...

Can we automatically design a Convolutional Network (ConvNet) with the highest image classification accuracy under latency constraint of mobile device? Neural Architecture Search (NAS) for ConvNet is challenging problem due to combinatorially large space and search time (at least 200 GPU-hours). To alleviate this complexity, propose Single-Path NAS, novel differentiable NAS method designing device-efficient ConvNets in less than 4 hours. 1. Novel formulation: our introduces single-path,...

Application-specific integrated circuit (ASIC) implementations for Deep Neural Networks (DNNs) have been adopted in many systems because of their higher classification speed. However, although they may be characterized by better accuracy, larger DNNs require significant energy and area, thereby limiting wide adoption. The consumption is driven both memory accesses computation. Binarized (BNNs), as a trade-off between accuracy consumption, can achieve great reduction, good large due to its...

In recent years, Convolutional Neural Networks (CNNs) have shown superior capability in visual learning tasks. While accuracy-wise CNNs provide unprecedented performance, they are also known to be computationally intensive and energy demanding for modern computer systems. this paper, we propose Virtual Pooling (ViP), a model-level approach improve speed consumption of CNN-based image classification object detection tasks, with provable error bound. We show the efficacy ViP through...

Binarized Neural Networks (BNNs) can significantly reduce the inference latency and energy consumption in resource-constrained devices due to their pure-logical computation fewer memory accesses. However, training BNNs is difficult since activation flow encounters degeneration, saturation, gradient mismatch problems. Prior work alleviates these issues by increasing bits adding floating-point scaling factors, thereby sacrificing BNN's efficiency. In this paper, we propose use distribution...

To improve the throughput and energy efficiency of Deep Neural Networks (DNNs) on customized hardware, lightweight neural networks constrain weights DNNs to be a limited combination (denoted as $k\in\{1,2\}$) powers 2. In such networks, multiply-accumulate operation can replaced with single shift operation, or two shifts an add operation. provide even more design flexibility, $k$ for each convolutional filter optimally chosen instead being fixed every filter. this paper, we formulate...

To improve precipitation predictions and enable accurate rainfall models for smart water networks, it is imperative to account multivariate multiscale dynamics with long-memory temporal relationships, long-range spatial dependencies, low-frequency variability. While prior art has motivated the use of complex networks capture these trends, existing work limited specific climate phenomena (such as El Niño) regions. In this paper we employ a comprehensive assessment respect across multiple...

In recent years, supervised learning using Convolutional Neural Networks (CNNs) has achieved great success in image classification tasks, and large scale labeled datasets have contributed significantly to this achievement. However, the definition of a label is often application dependent. For example, an cat can be as "cat" or perhaps more specifically "Persian cat." We refer granularity. paper, we conduct extensive experiments various demonstrate analyze how why training based on fine-grain...

As the machine learning and systems community strives to achieve higher energy-efficiency through custom DNN accelerators model compression techniques, there is a need for design space exploration framework that incorporates quantization-aware processing elements into accelerator while having accurate fast power, performance, area models. In this work, we present QAPPA, highly parameterized modeling accelerators. Our can facilitate future research on of various choices such as bit precision,...

As the machine learning and systems communities strive to achieve higher energy-efficiency through custom deep neural network (DNN) accelerators, varied bit precision or quantization levels, there is a need for design space exploration frameworks that incorporate quantization-aware processing elements (PE) into accelerator while having accurate fast power, performance, area models. In this work, we present QADAM, highly parameterized modeling framework DNN accelerators. Our can facilitate...