Embedded deep neural networks: “The cost of everything and the value of nothing”

Classification of ships and icebergs in the Arctic in satellite images is an important problem. We study how to train deep neural networks for improving the discrimination of ships and icebergs in multispectral satellite images. We also analyze synthetic-aperture radar (SAR) images for comparison. The annotated datasets of ships and icebergs are collected from multispectral Sentinel-2 data and taken from the C-CORE dataset of Sentinel-1 SAR images. Convolutional Neural Networks with a range of hyperparameters are tested and optimized. Classification accuracies are considerably better for deep neural networks than for support vector machines. Deeper neural nets improve the accuracy per epoch but at the cost of longer processing time. Extending the datasets with semi-supervised data from Greenland improves the accuracy considerably whereas data augmentation by rotating and flipping the images has little effect. The resulting classification accuracies for ships and icebergs are 86% for the SAR data and 96% for the MSI data due to the better resolution and more multispectral bands. The size and quality of the datasets are essential for training the deep neural networks, and methods to improve them are discussed. The reduced false alarm rates and exploitation of multisensory data are important for Arctic search and rescue services.

Download Full-text

Pruning for Hardware-Based Deep Spiking Neural Networks Using Gated Schottky Diode as Synaptic Devices

Journal of Nanoscience and Nanotechnology ◽

10.1166/jnn.2020.18772 ◽

2020 ◽

Vol 20 (11) ◽

pp. 6603-6608 ◽

Cited By ~ 1

Author(s):

Sung-Tae Lee ◽

Suhwan Lim ◽

Jong-Ho Bae ◽

Dongseok Kwon ◽

Hyeong-Su Kim ◽

...

Keyword(s):

Neural Networks ◽

Deep Neural Networks ◽

Schottky Diodes ◽

Computational Cost ◽

Spiking Neural Networks ◽

Training Procedure ◽

Learning Tasks ◽

L1 Regularization ◽

The Cost ◽

High Computational Cost

Deep learning represents state-of-the-art results in various machine learning tasks, but for applications that require real-time inference, the high computational cost of deep neural networks becomes a bottleneck for the efficiency. To overcome the high computational cost of deep neural networks, spiking neural networks (SNN) have been proposed. Herein, we propose a hardware implementation of the SNN with gated Schottky diodes as synaptic devices. In addition, we apply L1 regularization for connection pruning of the deep spiking neural networks using gated Schottky diodes as synap-tic devices. Applying L1 regularization eliminates the need for a re-training procedure because it prunes the weights based on the cost function. The compressed hardware-based SNN is energy efficient while achieving a classification accuracy of 97.85% which is comparable to 98.13% of the software deep neural networks (DNN).

Download Full-text

Particle-Based Fluid Surface Rendering with Neural Networks

10.24132/csrn.2021.3101.26 ◽

2021 ◽

Author(s):

Viktória Burkus ◽

Attila Kárpáti ◽

László Szécsi

Keyword(s):

Neural Networks ◽

Deep Neural Networks ◽

Ground Truth ◽

Fluid Simulation ◽

Image Manipulation ◽

Surface Rendering ◽

Fluid Surface ◽

Image Pairs ◽

The Cost

Surface reconstruction for particle-based fluid simulation is a computational challenge on par with the simula- tion itself. In real-time applications, splatting-style rendering approaches based on forward rendering of particle impostors are prevalent, but they suffer from noticeable artifacts. In this paper, we present a technique that combines forward rendering simulated features with deep-learning image manipulation to improve the rendering quality of splatting-style approaches to be perceptually similar to ray tracing solutions, circumventing the cost, complexity, and limitations of exact fluid surface rendering by replacing it with the flat cost of a neural network pass. Our solution is based on the idea of training generative deep neural networks with image pairs consisting of cheap particle impostor renders and ground truth high quality ray-traced images.

Download Full-text

CrowdTC: Crowd-powered Learning for Text Classification

ACM Transactions on Knowledge Discovery from Data ◽

10.1145/3457216 ◽

2021 ◽

Vol 16 (1) ◽

pp. 1-23

Author(s):

Keyu Yang ◽

Yunjun Gao ◽

Lei Liang ◽

Song Bian ◽

Lu Chen ◽

...

Keyword(s):

Neural Network ◽

Neural Networks ◽

Text Classification ◽

Deep Neural Networks ◽

Semantic Information ◽

Human Beings ◽

Hybrid Neural Network ◽

Public Datasets ◽

Almost All ◽

The Cost

Text classification is a fundamental task in content analysis. Nowadays, deep learning has demonstrated promising performance in text classification compared with shallow models. However, almost all the existing models do not take advantage of the wisdom of human beings to help text classification. Human beings are more intelligent and capable than machine learning models in terms of understanding and capturing the implicit semantic information from text. In this article, we try to take guidance from human beings to classify text. We propose Crowd-powered learning for Text Classification (CrowdTC for short). We design and post the questions on a crowdsourcing platform to extract keywords in text. Sampling and clustering techniques are utilized to reduce the cost of crowdsourcing. Also, we present an attention-based neural network and a hybrid neural network to incorporate the extracted keywords as human guidance into deep neural networks. Extensive experiments on public datasets confirm that CrowdTC improves the text classification accuracy of neural networks by using the crowd-powered keyword guidance.

Download Full-text

Localizer-to-Mapper Knowledge Transfer: Real-time Diagnosis of Deep SLAM in Everyday Navigation

10.31224/osf.io/qb6p7 ◽

2021 ◽

Author(s):

kanji tanaka

Keyword(s):

Neural Networks ◽

Fault Diagnosis ◽

Real Time ◽

Deep Neural Networks ◽

Black Box ◽

Input Output ◽

Diagnosis Algorithm ◽

The Cost ◽

Time Diagnosis

In this study, we address a novel "domain-shift localization (DSL)" problem, by which "user robots" of a deep SLAM system localize a domain-shifted region in the robot workspace during their daily navigation. Furthermore, we present a case study pertaining to a simple deep SLAM system comprising a visual place classifier and visual odometry as exteroceptive and proprioceptive modules, respectively. Such a DSL method enables "mapper robots" to focus on available resources (e.g., time, energy, and computation) in the domainshifted region, rather than the entire workspace, thereby significantly reducing the cost of per-domain DNN maintenance. Unlike conventional scenarios of SLAM diagnosis, the deep SLAM system comprises deep neural networks (DNNs) with black-box characteristics, which render it difficult to directly diagnose the internal signals of SLAM modules. Hence, we present a novel diagnosis algorithm that does not rely on internal signals but uses only the input/output signals of DNNs as input to DSL. Experiments demonstrate that, compared with a vanilla deep SLAM system that does not reflect fault diagnosis, the proposed deep SLAM framework can achieve a path that is more similar to the actual measured GPS path. This study is a first step towards a real-time approach to diagnose aging of deep SLAM systems.

Download Full-text

Direct Quantization for Training Highly Accurate Low Bit-width Deep Neural Networks

Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence ◽

10.24963/ijcai.2020/292 ◽

2020 ◽

Author(s):

Tuan Hoang ◽

Thanh-Toan Do ◽

Tam V. Nguyen ◽

Ngai-Man Cheung

Keyword(s):

Neural Networks ◽

Cost Function ◽

Image Classification ◽

Convolutional Neural Networks ◽

Gradient Descent ◽

Deep Neural Networks ◽

State Of The Art ◽

Deep Convolutional Neural Networks ◽

Novel Method ◽

The Cost

This paper proposes two novel techniques to train deep convolutional neural networks with low bit-width weights and activations. First, to obtain low bit-width weights, most existing methods obtain the quantized weights by performing quantization on the full-precision network weights. However, this approach would result in some mismatch: the gradient descent updates full-precision weights, but it does not update the quantized weights. To address this issue, we propose a novel method that enables direct updating of quantized weights with learnable quantization levels to minimize the cost function using gradient descent. Second, to obtain low bit-width activations, existing works consider all channels equally. However, the activation quantizers could be biased toward a few channels with high-variance. To address this issue, we propose a method to take into account the quantization errors of individual channels. With this approach, we can learn activation quantizers that minimize the quantization errors in the majority of channels. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on the image classification task, using AlexNet, ResNet and MobileNetV2 architectures on CIFAR-100 and ImageNet datasets.

Download Full-text