The ability to learn from incrementally arriving data is essential for any life-long learning system. However, standard deep neural networks forget the knowledge about the old tasks, a phenomenon called catastrophic forgetting, when trained on incrementally arriving data. We discuss the biases in current Generative Adversarial Networks (GAN) based approaches that learn the classifier by knowledge distillation from previously trained classifiers. These biases cause the trained classifier to perform poorly. We propose an approach to remove these biases by distilling knowledge from the classifier of AC-GAN. Experiments on MNIST and CIFAR10 show that this method is comparable to current state of the art rehearsal based approaches. The code for this paper is available at https://bit.ly/incremental-learning

Herein, we present a system for hyperspectral image segmentation that utilizes multiple class--based denoising autoencoders which are efficiently trained. Moreover, we present a novel hyperspectral data augmentation method for labelled HSI data using linear mixtures of pixels from each class, which helps the system with edge pixels which are almost always mixed pixels. Finally, we utilize a deep neural network and morphological hole-filling to provide robust image classification. Results run on the Salinas dataset verify the high performance of the proposed algorithm.

Neural sequence-to-sequence (seq2seq) approaches have proven to be successful in grammatical error correction (GEC). Based on the seq2seq framework, we propose a novel fluency boost learning and inference mechanism. Fluency boosting learning generates diverse error-corrected sentence pairs during training, enabling the error correction model to learn how to improve a sentence's fluency from more instances, while fluency boosting inference allows the model to correct a sentence incrementally with multiple inference steps. Combining fluency boost learning and inference with convolutional seq2seq models, our approach achieves the state-of-the-art performance: 75.72 (F_{0.5}) on CoNLL-2014 10 annotation dataset and 62.42 (GLEU) on JFLEG test set respectively, becoming the first GEC system that reaches human-level performance (72.58 for CoNLL and 62.37 for JFLEG) on both of the benchmarks.

Deep Learning is well known to be the representative method of artificial intelligence. The representation learning can discover the good set of features to input patterns and calculate the representation itself. Many kinds of structures and learning methods have been developed to achieve the great success. It is often said that Deep Learning should include the hierarchical model deeply, or the discovering of optimal structure and its parameters of Convolutional Neural Network (CNN) [1] is important. This issue pointed out by many researchers is right definitely, however, the effort to find the optimal structure and the parameters is very expensive and the calculation cost becomes high. To realize high level representation at low calculation cost, the self-organizing mechanism to adjust the structure itself and parameters simultaneously should be required with the statistical learning method. We have developed the structural learning method of Restricted Boltzmann Machine (RBM) [2] by neuron generation/annihilation algorithm [3].

Recently, Deep Learning has been applied in the techniques of artificial intelligence. Especially, Deep Learning performed good results in the field of image recognition. Most new Deep Learning architectures are naturally developed in image recognition. For this reason, not only the numerical data and text data but also the time-series data are transformed to the image data format. Multi-modal data consists of two or more kinds of data such as picture and text. The arrangement in a general method is formed in the squared array with no specific aim. In this paper, the data arrangement are modified according to the similarity of input-output pattern in Adaptive Structural Learning method of Deep Belief Network. The similarity of output signals of hidden neurons is made by the order rearrangement of hidden neurons. The experimental results for the data rearrangement in squared array showed the shortening time required for DBN learning.

Recently, the market on deep learning including not only software but also hardware is developing rapidly. Big data is collected through IoT devices and the industry world will analyze them to improve their manufacturing process. Deep Learning has the hierarchical network architecture to represent the complicated features of input patterns. Although deep learning can show the high capability of classification, prediction, and so on, the implementation on GPU devices are required. We may meet the trade-off between the higher precision by deep learning and the higher cost with GPU devices. We can success the knowledge extraction from the trained deep learning with high classification capability. The knowledge that can realize faster inference of pre-trained deep network is extracted as IF-THEN rules from the network signal flow given input data. Some experiment results with benchmark tests for time series data sets showed the effectiveness of our proposed method related to the computational speed.

The automatic analysis of ultrasound sequences can substantially improve the efficiency of clinical diagnosis. In this work we present our attempt to automate the challenging task of measuring the vascular diameter of the fetal abdominal aorta from ultrasound images. We propose a neural network architecture consisting of three blocks: a convolutional layer for the extraction of imaging features, a Convolution Gated Recurrent Unit (C-GRU) for enforcing the temporal coherence across video frames and exploiting the temporal redundancy of a signal, and a regularized loss function, called \textit{CyclicLoss}, to impose our prior knowledge about the periodicity of the observed signal. We present experimental evidence suggesting that the proposed architecture can reach an accuracy substantially superior to previously proposed methods, providing an average reduction of the mean squared error from $0.31 mm^2$ (state-of-art) to $0.09 mm^2$, and a relative error reduction from $8.1\%$ to $5.3\%$. The mean execution speed of the proposed approach of 289 frames per second makes it suitable for real time clinical use.

We proposed a modified regularized dual averaging method for training sparse deep convolutional neural networks. The regularized dual averaging method has been proven to be effective in obtaining sparse solutions in convex optimization problems, but not applied to deep learning fields before. We analyzed the new version in convex conditions and prove the convergence of it. The modified method can obtain more sparse solutions than traditional sparse optimization methods such as proximal-SGD, while keeping almost the same accuracy as stochastic gradient method with momentum on certain datasets.

Raven's Progressive Matrices are one of the widely used tests in evaluating the human test taker's fluid intelligence. Analogously, this paper introduces geometric generalization based zero-shot learning tests to measure the rapid learning ability and the internal consistency of deep generative models. Our empirical research analysis on state-of-the-art generative models discern their ability to generalize concepts across classes. In the process, we introduce Infinite World, an evaluable, scalable, multi-modal, light-weight dataset and Zero-Shot Intelligence Metric ZSI. The proposed tests condenses human-level spatial and numerical reasoning tasks to its simplistic geometric forms. The dataset is scalable to a theoretical limit of infinity, in numerical features of the generated geometric figures, image size and in quantity. We systematically analyze state-of-the-art model's internal consistency, identify their bottlenecks and propose a pro-active optimization method for few-shot and zero-shot learning.

Specialized deep learning hardware is starting to become commonplace in the datacenter, and on the edge. Much of the progress in ML system has been fueled by hardware specialization, which allows faster training and inference at lower energy costs [7, 2, 3, 6]. Using specialized hardware, though, requires deep learning frameworks to be redesigned around novel hardware architectural features and interfaces. Google's TensorFlow XLA and TPU [7] system stack is an excellent example of a complete framework built to take full advantage of hardware acceleration. However at the time of this writing, not much of the TPU hardware design, low-level programming interface, code-generation, or operator libraries are made transparent and available for researchers to maintain, experiment, and customize. In order to understand how hardware specialization is transforming the deep learning system landscape, it is essential to give systems, compilers, and machine learning researchers access to a complete system stack that transparently exposes all of its layers, including the hardware architecture of the accelerator itself, and its low-level programming interface. We present the VTA stack, a complete deep-learning system stack built with TVM [1] around VTA, a generic deep learning hardware accelerator design. VTA aims to serve as a blueprint to deep learning, systems, and compilers researchers who want to understand how hardware accelerators dictate new constraints across the system stack, and how the latter can be better co-designed with hardware. The VTA stack was designed with the following design objectives: - Provide a common deep learning system stack for hardware, compilers, and systems researchers alike to incorporate state-of-the-art optimizations and co-design techniques.