DRAM-based accelerators have shown their potential in addressing the memory wall challenge of the traditional von Neumann architecture. Such accelerators exploit charge sharing or logic circuits for simple logic operations. As a result, they require many cycles for more complex operations such as a multi-bit multiply-accumulate (MAC) operation, resulting in significant data access and movement and potentially worsening power efficiency.To overcome these limitations, this paper presents MAC-DO, an efficient and low-power DRAM-based accelerator. Compared to previous DRAM-based accelerators, a MAC-DO cell, consisting of two 1T1C DRAM cells, innately supports a multi-bit MAC operation within a single cycle, significantly improving power efficiency while maintaining good linearity and compatibility with existing 1T1C DRAM cell and array structures. This achievement is facilitated by a novel analog computation method utilizing charge steering. As a result, MAC-DO efficiently can accelerate convolutions based on output stationary mapping, supporting the majority of computations performed in deep neural networks.Our evaluation using transistor-level simulation shows that a test MAC-DO array with 16×16 MAC-DO cells achieves 120.96 TOPS/W and 97.07% Top-1 accuracy for MNIST dataset without retraining.

Although supervised person re-identification (Re-ID) methods have shown impressive performance, they suffer from a poor generalization capability on unseen domains. Therefore, generalizable Re-ID has recently attracted growing attention. Many existing methods have employed an instance normalization technique to reduce style variations, but the loss of discriminative information could not be avoided. In this paper, we propose a novel generalizable Re-ID framework, named Meta Batch-Instance Normalization (MetaBIN). Our main idea is to generalize normalization layers by simulating unsuccessful generalization scenarios beforehand in the meta-learning pipeline. To this end, we combine learnable batch-instance normalization layers with meta-learning and investigate the challenging cases caused by both batch and instance normalization layers. Moreover, we diversify the virtual simulations via our meta-train loss accompanied by a cyclic inner-updating manner to boost generalization capability. After all, the MetaBIN framework prevents our model from overfitting to the given source styles and improves the generalization capability to unseen domains without additional data augmentation or complicated network design. Extensive experimental results show that our model outperforms the state-of-the-art methods on the large-scale domain generalization Re-ID benchmark and the cross-domain Re-ID problem. The source code is available at: https://github.com/bismex/MetaBIN.

In this paper, we attack a few-shot open-set recognition (FSOSR) problem, which is a combination of few-shot learning (FSL) and open-set recognition (OSR). It aims to quickly adapt a model to a given small set of labeled samples while rejecting unseen class samples. Since OSR requires rich data and FSL considers closed-set classification, existing OSR and FSL methods show poor performances in solving FSOSR problems. The previous FSOSR method follows the pseudo-unseen class sample-based methods, which collect pseudo-unseen samples from the other dataset or synthesize samples to model unseen class representations. However, this approach is heavily dependent on the composition of the pseudo samples. In this paper, we propose a novel unknown class sample detector, named SnaTCHer, that does not require pseudo-unseen samples. Based on the transformation consistency, our method measures the difference between the transformed prototypes and a modified prototype set. The modified set is composed by replacing a query feature and its predicted class prototype. SnaTCHer rejects samples with large differences to the transformed prototypes. Our method alters the unseen class distribution estimation problem to a relative feature transformation problem, independent of pseudo-unseen class samples. We investigate our SnaTCHer with various prototype transformation methods and observe that our method consistently improves unseen class sample detection performance without closed-set classification reduction.

Capsule networks consider spatial relationships in an input image. The relationship-based feature propagation in capsule networks shows promising results. However, a large number of trainable parameters limit their widespread use. In this paper, we propose Decomposed Capsule Network (DCN) to reduce the number of training parameters in the primary capsule generation stage. Our DCN represents a capsule as a combination of basis vectors. Generating basis vectors and their coefficients notably reduce the total number of training parameters. Moreover, we introduce an extension of the DCN architecture, named Multi-scale Decomposed Capsule Network (MDCN). The MDCN architecture integrates features from multiple scales to synthesize capsules with fewer parameters. Our proposed networks show better performance on the Fashion-MNIST dataset and the CIFAR10 dataset with fewer parameters than the original network.

Filtering noisy labels is crucial for robust training of deep neural networks. To train networks with noisy labels, sampling methods have been introduced, which sample the reliable instances to update networks using only sampled data. Since they rarely employ the non-sampled data for training, these methods have a fundamental limitation that they reduce the amount of the training data. To alleviate this problem, our approach aims to fully utilize the whole dataset by leveraging the information of the sampled data. To this end, we propose a novel graph-based learning framework that enables networks to propagate the label information of the sampled data to adjacent data, whether they are sampled or not. Also, we propose a novel self-training strategy to utilize the non-sampled data without labels and to regularize the network update using the information of the sampled data. Our method outperforms state-of-the-art sampling methods.

We introduce a novel unsupervised domain adaptation approach for object detection. We aim to alleviate the imperfect translation problem of pixel-level adaptations, and the source-biased discriminativity problem of feature-level adaptations simultaneously. Our approach is composed of two stages, i.e., Domain Diversification (DD) and Multi-domain-invariant Representation Learning (MRL). At the DD stage, we diversify the distribution of the labeled data by generating various distinctive shifted domains from the source domain. At the MRL stage, we apply adversarial learning with a multi-domain discriminator to encourage feature to be indistinguishable among the domains. DD addresses the source-biased discriminativity, while MRL mitigates the imperfect image translation. We construct a structured domain adaptation framework for our learning paradigm and introduce a practical way of DD for implementation. Our method outperforms the state-of-the-art methods by a large margin of 3%~11% in terms of mean average precision (mAP) on various datasets.

Visual object counting aims to count the number of objects in a given image or video. Among many object counting methods, the counting by density estimation method draws attention because of its capability of counting and its object localization ability. The method utilizes the object density map of the image with multiple objects. The density is estimated by a regression model which learns the mapping between the local features of the given image and the density map generated from its corresponding object locations. Unlike conventional methods that only rely on object locations for density map generation, in this paper, we show that the system performance can be increased by considering object shapes as well as object locations. To this end, we propose two approaches to generate the ground truth density map from the object locations. Both methods generate the density map which reflects structural features of the objects. We show that the regression models trained with density maps which reflect the object shape outperforms the models trained with density maps generated by the conventional density map generation method on several challenging benchmarks. In other words, we observe that it is essential to generate the ground truth density map according to object shape in the image.