Spiking neural networks (SNNs) have gained considerable interest due to their energy-efficient characteristics, yet lack of a scalable training algorithm has restricted their applicability in practical machine learning problems. The deep neural network-to-SNN conversion approach has been widely studied to broaden the applicability of SNNs. Most previous studies, however, have not fully utilized spatio-temporal aspects of SNNs, which has led to inefficiency in terms of number of spikes and inference latency. In this paper, we present T2FSNN, which introduces the concept of time-to-first-spike coding into deep SNNs using the kernel-based dynamic threshold and dendrite to overcome the aforementioned drawback. In addition, we propose gradient-based optimization and early firing methods to further increase the efficiency of the T2FSNN. According to our results, the proposed methods can reduce inference latency and number of spikes to 22% and less than 1%, compared to those of burst coding, which is the state-of-the-art result on the CIFAR-100.
Adversarial examples cause neural networks to produce incorrect outputs with high confidence. Although adversarial training is one of the most effective forms of defense against adversarial examples, unfortunately, a large gap exists between test accuracy and training accuracy in adversarial training. In this paper, we identify Adversarial Feature Overfitting (AFO), which may cause poor adversarially robust generalization, and we show that adversarial training can overshoot the optimal point in terms of robust generalization, leading to AFO in our simple Gaussian model. Considering these theoretical results, we present soft labeling as a solution to the AFO problem. Furthermore, we propose Adversarial Vertex mixup (AVmixup), a soft-labeled data augmentation approach for improving adversarially robust generalization. We complement our theoretical analysis with experiments on CIFAR10, CIFAR100, SVHN, and Tiny ImageNet, and show that AVmixup significantly improves the robust generalization performance and that it reduces the trade-off between standard accuracy and adversarial robustness.
A structure of a protein has a direct impact on its properties and functions. However, identification of structural similarity directly from amino acid sequences remains as a challenging problem in computational biology. In this paper, we introduce a novel BERT-wise pre-training scheme for a protein sequence representation model called PLUS, which stands for Protein sequence representations Learned Using Structural information. As natural language representation models capture syntactic and semantic information of words from a large unlabeled text corpus, PLUS captures structural information of amino acids from a large weakly labeled protein database. Since the Transformer encoder, BERT's original model architecture, has a severe computational requirement to handle long sequences, we first propose to combine a bidirectional recurrent neural network with the BERT-wise pre-training scheme. PLUS is designed to learn protein representations with two pre-training objectives, i.e., masked language modeling and same family prediction. Then, the pre-trained model can be fine-tuned for a wide range of tasks without training randomly initialized task-specific models from scratch. It obtains new state-of-the-art results on both (1) protein-level and (2) amino-acid-level tasks, outperforming many task-specific algorithms.
When a deep neural network is trained on data with only image-level labeling, the regions activated in each image tend to identify only a small region of the target object. We propose a method of using videos automatically harvested from the web to identify a larger region of the target object by using temporal information, which is not present in the static image. The temporal variations in a video allow different regions of the target object to be activated. We obtain an activated region in each frame of a video, and then aggregate the regions from successive frames into a single image, using a warping technique based on optical flow. The resulting localization maps cover more of the target object, and can then be used as proxy ground-truth to train a segmentation network. This simple approach outperforms existing methods under the same level of supervision, and even approaches relying on extra annotations. Based on VGG-16 and ResNet 101 backbones, our method achieves the mIoU of 65.0 and 67.4, respectively, on PASCAL VOC 2012 test images, which represents a new state-of-the-art.
Most deep learning approaches for text-to-SQL generation are limited to the WikiSQL dataset, which only supports very simple queries. Recently, template-based and sequence-to-sequence approaches were proposed to support complex queries, which contain join queries, nested queries, and other types. However, Finegan-Dollak et al. (2018) demonstrated that both the approaches lack the ability to generate SQL of unseen templates. In this paper, we propose a template-based one-shot learning model for the text-to-SQL generation so that the model can generate SQL of an untrained template based on a single example. First, we classify the SQL template using the Matching Network that is augmented by our novel architecture Candidate Search Network. Then, we fill the variable slots in the predicted template using the Pointer Network. We show that our model outperforms state-of-the-art approaches for various text-to-SQL datasets in two aspects: 1) the SQL generation accuracy for the trained templates, and 2) the adaptability to the unseen SQL templates based on a single example without any additional training.
Unsupervised domain adaptation aiming to learn a specific task for one domain using another domain data has emerged to address the labeling issue in supervised learning, especially because it is difficult to obtain massive amounts of labeled data in practice. The existing methods have succeeded by reducing the difference between the embedded features of both domains, but the performance is still unsatisfactory compared to the supervised learning scheme. This is attributable to the embedded features that lay around each other but do not align perfectly and establish clearly separable clusters. We propose a novel domain adaptation method based on label propagation and cycle consistency to let the clusters of the features from the two domains overlap exactly and become clear for high accuracy. Specifically, we introduce cycle consistency to enforce the relationship between each cluster and exploit label propagation to achieve the association between the data from the perspective of the manifold structure instead of a one-to-one relation. Hence, we successfully formed aligned and discriminative clusters. We present the empirical results of our method for various domain adaptation scenarios and visualize the embedded features to prove that our method is critical for better domain adaptation.
Over the past decade, deep neural networks (DNNs) have become a de-facto standard for solving machine learning problems. As we try to solve more advanced problems, growing demand for computing and power resources are inevitable, nearly impossible to employ DNNs on embedded systems, where available resources are limited. Given these circumstances, spiking neural networks (SNNs) are attracting widespread interest as the third generation of neural network, due to event-driven and low-powered nature. However, SNNs come at the cost of significant performance degradation largely due to complex dynamics of SNN neurons and non-differential spike operation. Thus, its application has been limited to relatively simple tasks such as image classification. In this paper, we investigate the performance degradation of SNNs in the much more challenging task of object detection. From our in-depth analysis, we introduce two novel methods to overcome a significant performance gap: channel-wise normalization and signed neuron with imbalanced threshold. Consequently, we present a spiked-based real-time object detection model, called Spiking-YOLO that provides near-lossless information transmission in a shorter period of time for deep SNN. Our experiments show that the Spiking-YOLO is able to achieve comparable results up to 97% of the original YOLO on a non-trivial dataset, PASCAL VOC.
The main obstacle to weakly supervised semantic image segmentation is the difficulty of obtaining pixel-level information from coarse image-level annotations. Most methods based on image-level annotations use localization maps obtained from the classifier, but these only focus on the small discriminative parts of objects and do not capture precise boundaries. FickleNet explores diverse combinations of locations on feature maps created by generic deep neural networks. It selects hidden units randomly and then uses them to obtain activation scores for image classification. FickleNet implicitly learns the coherence of each location in the feature maps, resulting in a localization map which identifies both discriminative and other parts of objects. The ensemble effects are obtained from a single network by selecting random hidden unit pairs, which means that a variety of localization maps are generated from a single image. Our approach does not require any additional training steps and only adds a simple layer to a standard convolutional neural network; nevertheless it outperforms recent comparable techniques on the Pascal VOC 2012 benchmark in both weakly and semi-supervised settings.
Steganography is the science of unnoticeably concealing a secret message within a certain image, called a cover image. The cover image with the secret message is called a stego image. Steganography is commonly used for illegal purposes such as terrorist activities and pornography. To thwart covert communications and transactions, attacking algorithms against steganography, called steganalysis, exist. Currently, there are many studies implementing deep learning to the steganography algorithm. However, conventional steganalysis is no longer effective for deep learning based steganography algorithms. Our framework is the first one to disturb covert communications and transactions via the recent deep learning-based steganography algorithms. We first extract a sophisticated pixel distribution of the potential stego image from the auto-regressive model induced by deep learning. Using the extracted pixel distributions, we detect whether an image is the stego or not at the pixel level. Each pixel value is adjusted as required and the adjustment induces an effective removal of the secret image. Because the decoding method of deep learning-based steganography algorithms is approximate (lossy), which is different from the conventional steganography, we propose a new quantitative metric that is more suitable for measuring the accurate effect. We evaluate our method using three public benchmarks in comparison with a conventional steganalysis method and show up to a 20% improvement in terms of decoding rate.