Face recognition in the presence of age and quality variations poses a formidable challenge. While recent margin-based loss functions have shown promise in addressing these variations individually, real-world scenarios such as selfie versus ID face matching often involve simultaneous variations of both age and quality. In response, we propose a comprehensive framework aimed at mitigating the impact of these variations while preserving vital identity-related information crucial for accurate face recognition. The proposed adaptive margin-based loss function AQUAFace exhibits adaptiveness towards hard samples characterized by significant age and quality variations. This loss function is meticulously designed to prioritize the preservation of identity-related features while simultaneously mitigating the adverse effects of age and quality variations on face recognition accuracy. To validate the effectiveness of our approach, we focus on the specific task of selfie versus ID document matching. Our results demonstrate that AQUAFace effectively handles age and quality differences, leading to enhanced recognition performance. Additionally, we explore the benefits of fine-tuning the recognition model with synthetic data, further boosting performance. As a result, our proposed model, AQUAFace, achieves state-of-the-art performance on six benchmark datasets (CALFW, CPLFW, CFP-FP, AgeDB, IJB-C, and TinyFace), each exhibiting diverse age and quality variations.

This paper delves into the challenging task of selfie vs ID face verification which involves matching high-resolution selfies with low-resolution faces extracted from scanned ID documents. Existing face verification models often face performance degradation when confronted with this task, mainly due to disparities in data distributions, such as age-difference, degradation due to scanning, and difference in appearance. To address this issue and enhance performance, the paper explores the implementation of facial quality assessment and hard-pair mining techniques. In addition, the paper investigates the potential of synthetic data for training face verification models tailored for this specific task. The integration of synthetic data as an alternative training source is explored to improve robustness and overcome legal and privacy concerns arising from authentic datasets. By combining hard pair mining, facial quality assessment, and the utilization of synthetic data, this paper presents a comprehensive framework that aims to achieve improved face verification results in the complex scenario of selfie vs ID matching. The goal is to optimize the models’ performance and enhance their ability to accurately match selfies with the corresponding ID images, even under challenging conditions.

An attention guided convolutional neural network (CNN) for the classification of breast cancer histopathology images is proposed. Neural networks are generally applied as black box models and often the network's decisions are difficult to interpret. Making the decision process transparent, and hence reliable is important for a computer-assisted diagnosis (CAD) system. Moreover, it is crucial that the network's decision be based on histopathological features that are in agreement with a human expert. To this end, we propose to use additional region-level supervision for the classification of breast cancer histopathology images using CNN, where the regions of interest (RoI) are localized and used to guide the attention of the classification network simultaneously. The proposed supervised attention mechanism specifically activates neurons in diagnostically relevant regions while suppressing activations in irrelevant and noisy areas. The class activation maps generated by the proposed method correlate well with the expectations of an expert pathologist. Moreover, the proposed method surpasses the state-of-the-art on the BACH microscopy test dataset (part A) with a significant margin.

In this paper, a memristive artificial neural circuit imitating the excitatory chemical synaptic transmission of biological synapse is designed. The proposed memristor-based neural circuit exhibits synaptic plasticity, one of the important neurochemical foundations for learning and memory, which is demonstrated via the efficient imitation of short-term facilitation and long-term potentiation. Moreover, the memristive artificial circuit also mimics the distinct biological attributes of strong stimulation and deficient synthesis of neurotransmitters. The proposed artificial neural model is designed in SPICE, and the biological functionalities are demonstrated via various simulations. The simulation results obtained with the proposed artificial synapse are similar to the biological features of chemical synaptic transmission and synaptic plasticity.

When multiple memristors are connected to each other, the composite behavior of the devices becomes complicated and is difficult to predict, due to the polarity dependent nonlinear variation in the memristance of individual memristor. In this paper, we investigate the relationships among flux, charge and memristance of diverse composite memristors, using the HP TiO2model, and analyze the characteristics of complex memristor circuits. It is assumed that all memristor circuits operate at a stable composite memristance state, in which the composite flux curve does not vary and the memristor circuits act as a single memristive system, regardless of input current or voltage. Such study will be conducted for serial and parallel memristor circuits.

Analog hardware architecture of a memristor bridge synapse-based multilayer neural network and its learning scheme is proposed. The use of memristor bridge synapse in the proposed architecture solves one of the major problems, regarding nonvolatile weight storage in analog neural network implementations. To compensate for the spatial nonuniformity and nonideal response of the memristor bridge synapse, a modified chip-in-the-loop learning scheme suitable for the proposed neural network architecture is also proposed. In the proposed method, the initial learning is conducted in software, and the behavior of the software-trained network is learned by the hardware network by learning each of the single-layered neurons of the network independently. The forward calculation of the single-layered neuron learning is implemented on circuit hardware, and followed by a weight updating phase assisted by a host computer. Unlike conventional chip-in-the-loop learning, the need for the readout of synaptic weights for calculating weight updates in each epoch is eliminated by virtue of the memristor bridge synapse and the proposed learning scheme. The hardware architecture along with the successful implementation of proposed learning on a three-bit parity network, and on a car detection network is also presented.