Here is a compilation of the most remarkable AI research articles. It is a carefully selected compilation of the latest advancements in artificial intelligence and data science, arranged chronologically and accompanied by a link to a more detailed article.

Contextual Diversity for Active Learning

The utilization of deep convolutional neural networks (CNNs) in numerous practical applications is limited due to the necessity for extensively annotated datasets. Active learning (AL) can solve the problem. These techniques enable the selection of a subset of data that achieves the highest accuracy after fine-tuning while working within a specific annotation budget. Advanced AL methods commonly depend on metrics of visual diversity or prediction uncertainty, but they need help adequately capture spatial context differences. Conversely, contemporary CNN designs extensively utilize spatial context to achieve precise predictions.

Given the lack of ground-truth labels, the researchers propose the concept of contextual variety to measure the confusion caused by classes that occur together in space. Contextual Diversity (CD) is based on the critical discovery that the probability vector predicted by a Convolutional Neural Network (CNN) for a specific area of interest usually includes information from a wider input area. Based on this discovery, the researchers utilize the suggested CD measure in two active learning frameworks: 

(1) a strategy based on core sets and 

(2) a policy based on reinforcement learning to select active frames.

Their comprehensive empirical assessment establishes cutting-edge outcomes for active learning on standard Semantic Segmentation, Object Detection, and Image Classification datasets. Their ablation studies demonstrate the distinct benefits of employing contextual variation in the context of active learning.

Attributional Robustness Training using Input-Gradient Spatial Alignment

Interpretability is a developing field of study in reliable machine learning. Ensuring the secure implementation of a machine learning system requires that both the prediction and its accompanying explanation are dependable and resilient. Recent findings have demonstrated that explanations can be easily altered by introducing visually imperceptible modifications to the input while maintaining the model's prediction unchanged. This study focuses on the issue of attributional robustness, which refers to models with solid and reliable explanations. 

The researchers demonstrate an upper limit for attributional vulnerability, determined by the spatial correlation between the input image and its explanation map. The authors provide a training approach that acquires resilient characteristics by lowering the upper limit by utilising soft-margin triplet loss. Their robust attribution training methodology (ART) surpasses the current state-of-the-art measure of attributional robustness by around 6-18% on various widely used datasets, such as SVHN, CIFAR-10, and GTSRB. The researchers demonstrate the effectiveness of the proposed resilient training strategy (\textit{ART}) in improving weakly supervised object localization. They achieve the highest performance on the CUB-200 dataset, setting a new benchmark.

An Inference Algorithm for Multi-label MRF-MAP Problems with Clique Size 100

This study presents an algorithm that offers optimal solutions for submodular higher-order multi-label MRF-MAP energy functions. The technique is capable of handling computer vision issues of a practical nature, accommodating up to 16 labels and cliques with a size of 100. The approach employs a conversion technique that converts a multi-label problem into a 2-label problem on a significantly bigger clique. Only some algorithms utilizing this transformation could address issues exceeding 16 labels on cliques with a size of 4. The proposed algorithm aims to optimize the resulting issue with two labels using the Min Norm Point algorithm based on a submodular polyhedron. The task is arduous due to the extensive number of invalid states in the modified problem's state space. When dealing with polyhedral-based algorithms, the existence of invalid states presents a difficulty. 

In addition to numerical instability, this change significantly increases the dimension of the polyhedral space, making it impractical to employ established algorithms straightforwardly. The methodology described in this paper enables us to circumvent the significant expenses linked to invalid configurations, leading to a stable, practical, optimal, and efficient inference algorithm. This algorithm produces high-quality results for pixel-wise object segmentation and stereo-matching tasks in their experiments.

Image source: Unsplash