Convolutional neural networks have gained a remarkable success in computer vision. However, most usable network architectures are hand-crafted and usually require expertise elaborate design. In this paper, we provide block-wise generation pipeline called BlockQNN which automatically builds high-performance using the Q-Learning paradigm with epsilon-greedy exploration strategy. The optimal block is constructed by learning agent trained sequentially to choose component layers. We stack...

An approximate analogical reasoning schema (AARS) which exhibits the advantages of fuzzy set theory and in expert systems development is described. The AARS avoids going through conceptually complicated compositional rule inference. It uses a similarity measure sets as well threshold to determine whether should be fired modification function inferred from deduce consequent. Some numerical examples illustrate operation are presented. Finally, proposed compared with conventional existing...

Convolutional neural networks have gained a remarkable success in computer vision. However, most popular network architectures are hand-crafted and usually require expertise elaborate design. In this paper, we provide block-wise generation pipeline called BlockQNN which automatically builds high-performance using the Q-Learning paradigm with epsilon-greedy exploration strategy. The optimal block is constructed by learning agent trained to choose component layers sequentially. We stack...

Data augmentation (DA) has been widely utilized to improve generalization in training deep neural networks. Recently, human-designed data gradually replaced by automatically learned policy. Through finding the best policy well-designed search space of augmentation, AutoAugment can significantly validation accuracy on image classification tasks. However, this approach is not computationally practical for large-scale problems. In paper, we develop an adversarial method arrive at a...

Convolution operator is the core of convolutional neural networks (CNNs) and occupies most computation cost. To make CNNs more efficient, many methods have been proposed to either design lightweight or compress models. Although some efficient network structures proposed, such as MobileNet ShuffleNet, we find that there still exists redundant information between convolution kernels. address this issue, propose a novel dynamic method adaptively generate kernels based on image contents....

In this paper, we propose an inverse reinforcement learning method for architecture search (IRLAS), which trains agent to learn network structures that are topologically inspired by human-designed network. Most existing approaches totally neglect the topological characteristics of architectures, results in complicated with a high inference latency. Motivated fact networks elegant topology fast speed, mirror stimuli function biological cognition theory extract abstract knowledge expert...

This paper considers using deep neural networks for handwritten Chinese character recognition (HCCR) with arbitrary position, scale, and orientations. To solve this problem, we combine the recently proposed spatial transformer network (STN) residual (DRN). The STN acts like a shape normalization procedure. Different from traditional heuristic methods, is learned directly data. Furthermore, DRN makes training of very to be both efficient effective. With combination DRN, whole model can...

Convolutional neural networks have gained a remarkable success in computer vision. However, most usable network architectures are hand-crafted and usually require expertise elaborate design. In this paper, we provide block-wise generation pipeline called BlockQNN which automatically builds high-performance using the Q-Learning paradigm with epsilon-greedy exploration strategy. The optimal block is constructed by learning agent trained sequentially to choose component layers. We stack...

There are many decisions which usually made heuristically both in single object tracking (SOT) and multiple (MOT). Existing methods focus on tackling decision-making problems special tasks without a unified framework. In this paper, we propose decision controller (DC) is generally applicable to SOT MOT tasks. The learns an optimal policy with deep reinforcement learning algorithm that maximizes long term performance supervision. To prove the generalization ability of DC, apply it challenging...

Convolutional neural networks have gained a remarkable success in computer vision. However, most usable network architectures are hand-crafted and usually require expertise elaborate design. In this paper, we provide block-wise generation pipeline called BlockQNN which automatically builds high-performance using the Q-Learning paradigm with epsilon-greedy exploration strategy. The optimal block is constructed by learning agent trained to choose component layers sequentially. We stack...

Convolutional Neural Networks(CNNs) are both computation and memory intensive which hindered their deployment in mobile devices. Inspired by the relevant concept neural science literature, we propose Synaptic Pruning: a data-driven method to prune connections between input output feature maps with newly proposed class of parameters called Strength. Strength is designed capture importance connection based on amount information it transports. Experiment results show effectiveness our approach....

Automatic neural architecture search techniques are becoming increasingly important in machine learning area. Especially, weight sharing methods have shown remarkable potentials on searching good network architectures with few computational resources. However, existing mainly suffer limitations strategies: these either uniformly train all paths to convergence which introduces conflicts between branches and wastes a large amount of computation unpromising candidates, or selectively different...

The choice of activation functions is crucial for modern deep neural networks. Popular hand-designed like Rectified Linear Unit(ReLU) and its variants show promising performance in various tasks models. Swish, the automatically discovered function, has been proposed outperforms ReLU on many challenging datasets. However, it two main drawbacks. First, tree-based search space highly discrete restricted, which difficult searching. Second, sample-based searching method inefficient, making...

It is very common to use a regular grid like Tian-zi-ge or Mi-zi-ge help writing in Chinese handwriting environment, especially education and postal area. Although helpful for writing, it disaster recognition. This paper focuses on handwritten character blind inpainting with spot. To solve this problem, we the recently proposed conditional generative adversarial nets (GANs). Different from traditional engineering based method line detection edge detection, GANs learn map between target...

The choice of activation functions is crucial to deep neural networks. ReLU a popular hand-designed function. Swish, the automatically searched function, outperforms on many challenging datasets. However, search method has two main drawbacks. First, tree-based space highly discrete and restricted, which difficult search. Second, sample-based inefficient in finding specialized for each dataset or architecture. To overcome these drawbacks, we propose new function called Piecewise Linear Unit...

In this paper, we propose an inverse reinforcement learning method for architecture search (IRLAS), which trains agent to learn network structures that are topologically inspired by human-designed network. Most existing approaches totally neglect the topological characteristics of architectures, results in complicated with a high inference latency. Motivated fact networks elegant topology fast speed, mirror stimuli function biological cognition theory extract abstract knowledge expert...

Vector Quantization (VQ) is an appealing model compression method to obtain a tiny with less accuracy loss. While methods better codebooks and codes under fixed clustering dimensionality have been extensively studied, optimizations via the reduction of subvector are not carefully considered. This paper reports our recent progress on combination vector quantization, proposing Low-Rank Representation (LR <sup xmlns:mml="http://www.w3.org/1998/Math/MathML"...

In this paper, we propose a Collaboration of Experts (CoE) framework to pool together the expertise multiple networks towards common aim. Each expert is an individual network with on unique portion dataset, which enhances collective capacity. Given sample, selected by delegator, simultaneously outputs rough prediction support early termination. To fulfill framework, three modules impel each model play its role, namely weight generation module (WGM), label (LGM) and variance calculation...

Neural network architecture design mostly focuses on the new convolutional operator or special topological structure of block, little attention is drawn to configuration stacking each called Block Stacking Style (BSS). Recent studies show that BSS may also have an unneglectable impact networks, thus we efficient algorithm search it automatically. The proposed method, AutoBSS, a novel AutoML based Bayesian optimization by iteratively refining and clustering Code (BSSC), which can find optimal...

Weight decay is a widely used technique for training Deep Neural Networks(DNN). It greatly affects generalization performance but the underlying mechanisms are not fully understood. Recent works show that layers followed by normalizations, weight mainly effective learning rate. However, despite normalizations have been extensively adopted in modern DNNs, such as final fully-connected layer do satisfy this precondition. For these layers, effects of still unclear. In paper, we comprehensively...

An Approximate Reasoning Schema (ARS) is proposed which exhibits the advantage of Fuzzy Sets Theory in Expert Systems' Pattern Matching phase. This schema avoids going through conceptually complicated Compositional Rule Inference, has been employed implementation almost every expert system with approximate reasoning capability. The further extended to Interval-Valued Set's where additional uncertainty can be handled a formal model. Finally, our ARS implemented by using Texas Instruments...

The choice of activation functions is crucial for modern deep neural networks. Popular hand-designed like Rectified Linear Unit(ReLU) and its variants show promising performance in various tasks models. Swish, the automatically discovered function, has been proposed outperforms ReLU on many challenging datasets. However, it two main drawbacks. First, tree-based search space highly discrete restricted, which difficult searching. Second, sample-based searching method inefficient, making...