We propose a novel deep network architecture for lifelong learning which we refer to as Dynamically Expandable Network (DEN), that can dynamically decide its capacity it trains on sequence of tasks, learn compact overlapping knowledge sharing structure among tasks. DEN is efficiently trained in an online manner by performing selective retraining, expands upon arrival each task with only the necessary number units, and effectively prevents semantic drift splitting/duplicating units...

The goal of few-shot learning is to learn a classifier that generalizes well even when trained with limited number training instances per class. recently introduced meta-learning approaches tackle this problem by generic across large multiclass classification tasks and generalizing the model new task. Yet, such meta-learning, low-data in novel task still remains. In paper, we propose Transductive Propagation Network (TPN), framework for transductive inference classifies entire test set at...

Reducing bit-widths of activations and weights deep networks makes it efficient to compute store them in memory, which is crucial their deployments resource-limited devices, such as mobile phones. However, decreasing with quantization generally yields drastically degraded accuracy. To tackle this problem, we propose learn quantize via a trainable quantizer that transforms discretizes them. Specifically, parameterize the intervals obtain optimal values by directly minimizing task loss...

Dense computer vision tasks such as object detection and segmentation require effective multi-scale feature representation for detecting or classifying objects regions with varying sizes. While Convolutional Neural Networks (CNNs) have been the dominant architectures tasks, recently introduced Vision Transformers (ViTs) aim to replace them a backbone. Similar CNNs, ViTs build simple multi-stage structure (i.e., fine-to-coarse) single-scale patches. In this work, different perspective from...

Visual attributes expose human-defined semantics to object recognition models, but existing work largely restricts their influence mid-level cues during classifier training. Rather than treat as intermediate features, we consider how learning visual properties in concert with categories can regularize the models for both. Given a low-level feature space together attribute-and object-labeled image data, learn shared lower-dimensional representation by optimizing joint loss function that...

Existing adversarial learning approaches mostly use class labels to generate samples that lead incorrect predictions, which are then used augment the training of model for improved robustness. While some recent works propose semi-supervised methods utilize unlabeled data, they still require labels. However, do we really need at all, adversarially robust deep neural networks? In this paper, a novel attack makes confuse instance-level identities perturbed data samples. Further, present...

While semi-supervised learning (SSL) has proven to be a promising way for leveraging unlabeled data when labeled is scarce, the existing SSL algorithms typically assume that training class distributions are balanced. However, these trained under imbalanced can severely suffer generalizing balanced testing criterion, since they utilize biased pseudo-labels of toward majority classes. To alleviate this issue, we formulate convex optimization problem softly refine generated from model, and...

Summary Plants, as a sessile organism, produce various secondary metabolites to interact with the environment. These chemicals have fascinated plant science community because of their ecological significance and notable biological activity. However, predicting complete biosynthetic pathways from target molecules metabolic building blocks remains challenge. Here, we propose retrieval‐augmented dual‐view retrosynthesis (READRetro) practical bio‐retrosynthesis tool predict natural products....

Current uses of tagged images typically exploit only the most explicit information: link between nouns named and objects present somewhere in image. We propose to leverage "unspoken" cues that rest within an ordered list image tags so as improve object localization. define three novel implicit features from image's tags-the relative prominence each signified by its order mention, scale constraints implied unnamed objects, loose spatial links hinted at proximity names on list. By learning a...

As the application of deep learning has expanded to real-world problems with insufficient volume training data, transfer recently gained much attention as means improving performance in such small-data regime. However, when existing methods are applied between heterogeneous architectures and tasks, it becomes more important manage their detailed configurations often requires exhaustive tuning on them for desired performance. To address issue, we propose a novel approach based meta-learning...

In practical settings, a speaker recognition system needs to identify given short utterance, while the enrollment utterance may be relatively long.However, existing models perform poorly with such utterances.To solve this problem, we introduce meta-learning framework for imbalance length pairs.Specifically, use Prototypical Networks and train it support set of long utterances query varying lengths.Further, since optimizing only classes in episode insufficient learning discriminative...

While tasks could come with varying the number of instances and classes in realistic settings, existing meta-learning approaches for few-shot classification assume that per task class is fixed. Due to such restriction, they learn equally utilize meta-knowledge across all tasks, even when largely varies. Moreover, do not consider distributional difference unseen on which may have less usefulness depending relatedness. To overcome these limitations, we propose a novel model adaptively balances...

One of the most crucial challenges in question answering (QA) is scarcity labeled data, since it costly to obtain question-answer pairs for a target text domain with human annotation. An alternative approach tackle problem use automatically generated QA from either context or large amount unstructured texts (e.g. Wikipedia). In this work, we propose hierarchical conditional variational autoencoder (HCVAE) generating given as contexts, while maximizing mutual information between ensure their...

We introduce a method for image retrieval that leverages the implicit information about object importance conveyed by list of keyword tags person supplies an image.We propose unsupervised learning procedure based on Kernel Canonical Correlation Analysis discovers relationship between how humans tag images (e.g., order in which words are mentioned) and relative objects their layout scene.Using this discovered connection, we show to boost accuracy novel queries, such search results may more...