We propose the use of deep convolutional neural networks (DCNNs) for human detection and activity classification based on Doppler radar. Previously, proposed schemes these problems remained in conventional supervised learning paradigm that relies design handcrafted features. Whereas attained high accuracy, requirement domain knowledge each problem limits scalability schemes. In this letter, we present an alternative approach. apply DCNN, one most successful algorithms, directly to a raw...

We consider class incremental learning (CIL) problem, in which a agent continuously learns new classes from incrementally arriving training data batches and aims to predict well on all the learned so far. The main challenge of problem is catastrophic forgetting, for exemplar-memory based CIL methods, it generally known that forgetting commonly caused by classification score bias injected due imbalance between old (in exemplar-memory). While several methods have been proposed correct such...

Accurate classification of human aquatic activities using radar has a variety potential applications such as rescue operations and border patrols. Nevertheless, the on water not been extensively studied, unlike case dry ground, due to its unique challenge. Namely, only is cross section small, but micro-Doppler signatures are much noisier drops waves. In this paper, we first investigate whether discriminative could be obtained for through simulation study. Then, show how can effectively...

Recently, recurrent neural networks (RNN) have achieved the state-of-the-art performance in several applications that deal with temporal data, e.g., speech recognition, handwriting recognition and machine translation. While ability of handling long-term dependency data is key for success RNN, combating over-fitting training models a critical issue achieving cutting-edge particularly when depth size network increase. To end, there been some attempts to apply dropout, popular regularization...

Non-intrusive load monitoring (NILM), also known as energy disaggregation, is a blind source separation problem where household’s aggregate electricity consumption broken down into usages of individual appliances. In this way, the cost and trouble installing many measurement devices over numerous household appliances can be avoided, only one device needs to installed. The has been well-known since Hart’s seminal paper in 1992, recently significant performance improvements have achieved by...

We introduce a new neural network-based continual learning algorithm, dubbed as Uncertainty-regularized Continual Learning (UCL), which builds on traditional Bayesian online framework with variational inference. focus two significant drawbacks of the recently proposed regularization-based methods: a) considerable additional memory cost for determining per-weight regularization strengths and b) absence gracefully forgetting scheme, can prevent performance degradation in tasks. In this paper,...

With the explosive growth of online news readership, recommending interesting articles to users has become extremely important. While existing Web services such as Yahoo! and Digg attract users' initial clicks by leveraging various kinds signals, how engage algorithmically after their visit is largely under-explored. In this paper, we study problem post-click recommendation. Given that a user perused current article, our idea automatically identify "related" which would like read afterwards....

We consider the correction of errors from nucleotide sequences produced by next-generation targeted amplicon sequencing. The sequencing (NGS) platforms can provide a great deal data thanks to their high throughput, but associated error rates often tend be high. Denoising in high-throughput has thus become crucial process for boosting reliability downstream analyses. Our methodology, named DUDE-Seq, is derived general setting reconstructing finite-valued source corrupted discrete memoryless...

We ask whether the neural network interpretation methods can be fooled via adversarial model manipulation, which is defined as a fine-tuning step that aims to radically alter explanations without hurting accuracy of original models, e.g., VGG19, ResNet50, and DenseNet121. By incorporating results directly in penalty term objective function for fine-tuning, we show state-of-the-art saliency map based interpreters, LRP, Grad-CAM, SimpleGrad, easily with our manipulation. propose two types...

Since the recent advent of regulations for data protection (e.g., General Data Protection Regulation), there has been increasing demand in deleting information learned from sensitive pre-trained models without retraining scratch. The inherent vulnerability neural networks towards adversarial attacks and unfairness also calls a robust method to remove or correct an instance-wise fashion, while retaining predictive performance across remaining data. To this end, we consider unlearning, which...

We propose a new image denoising algorithm, dubbed as Fully Convolutional Adaptive Image DEnoiser (FC-AIDE), that can learn from an offline supervised training set with fully convolutional neural network well adaptively fine-tune the model for each given noisy image. significantly extend framework of recently proposed Neural AIDE, which formulates denoiser to be context-based pixelwise mappings and utilizes unbiased estimator MSE such denoisers. The two main contributions we make are; 1)...

We propose a novel regularization-based continual learning method, dubbed as Adaptive Group Sparsity based Continual Learning (AGS-CL), using two group sparsity-based penalties. Our method selectively employs the penalties when each node its importance, which is adaptively updated after new task. By utilizing proximal gradient descent for learning, exact sparsity and freezing of model guaranteed, thus, learner can explicitly control capacity continues. Furthermore, critical detail, we...

We consider the challenging blind denoising problem for Poisson-Gaussian noise, in which no additional information about clean images or noise level parameters is available. Particularly, when only "single" noisy are available training a denoiser, performance of existing methods was not satisfactory. Recently, pixelwise affine image denoiser (BP-AIDE) proposed and significantly improved above setting, to extent that it competitive with denoisers utilized information. However, BP-AIDE...

Fairness is becoming an increasingly crucial issue for computer vision, especially in the human-related decision systems. However, achieving algorithmic fairness, which makes a model produce indiscriminative outcomes against protected groups, still unresolved problem. In this paper, we devise systematic approach reduces biases via feature distillation visual recognition tasks, dubbed as MMD-based Fair Distillation (MFD). While technique has been widely used general to improve prediction...

Abstract Accurate prediction of the brain’s task reactivity from resting-state functional magnetic resonance imaging (fMRI) data remains a significant challenge in neuroscience. Traditional statistical approaches often fail to capture complex, nonlinear spatiotemporal patterns brain function. This study introduces SwiFUN (Swin fMRI UNet Transformer), novel deep-learning framework designed predict 3D activation maps directly scans. leverages advanced techniques such as shifted window-based...

Continual learning (CL) research typically assumes highly constrained exemplar memory resources. However, in many real-world scenarios-especially the era of large foundation models-memory is abundant, while GPU computational costs are primary bottleneck. In this work, we investigate CL a novel setting where ample (i.e., sufficient memory). Unlike prior methods designed for strict constraints, propose simple yet effective approach that directly operates model's weight space through...

<title>Abstract</title> Efficient and accurate DNA sequence classification is a crucial task in genomic data analysis. In this work, we construct lightweight classifier based on the LZ78 lossless universal compressor, optimize its performance through hyperparameter tuning. This outperforms state-of-the-art DNABERT-2 model Genomic Understanding Evaluation suite, while drastically reducing computational costs. Unlike DNABERT-2, which requires two weeks of multi-GPU training, our can be trained...

In the isointense stage, accurate volumetric image segmentation is a challenging task due to low contrast between tissues. this paper, we propose novel very deep network architecture based on densely convolutional for brain segmentation. The proposed provides dense connection layers that aims improve information flow in network. By concatenating features map of fine and coarse blocks, it allows capturing multi-scale contextual information. Experimental results demonstrate significant...