Here are the most outstanding articles on AI research. It is a carefully curated list of the most recent developments in data science and AI, presented chronologically with a link to a longer article for more information.

Amalgamation of protein sequence, structure and textual information for improving protein-protein interaction identification

A thorough investigation of protein-protein interactions (PPI) is crucial for comprehending metabolism and the regulatory mechanisms of biological entities such as proteins and carbohydrates. Most recent PPI tasks in the BioNLP domain have been conducted only utilizing textual data. This work asserts that including multimodal cues can enhance the automated detection of PPI. Researchers have created two multimodal datasets to facilitate the advancement of multimodal methods for PPI detection. These datasets are expansions and multimodal adaptations of two well-known benchmark PPI collections: BioInfer and HRPD50. In addition to the existing textual modalities, two more modalities, namely 3D protein structure and underlying genetic sequence, are also included for each occurrence. 

In addition, a novel deep multimodal architecture is designed to forecast protein interactions accurately using the created datasets. An extensive experimental investigation demonstrates that the multimodal approach outperforms strong baselines, including unimodal and state-of-the-art methods, on both the created multimodal datasets. 

Researchers:

• Pratik Dutta, Department of Computer Science & Engineering Indian Institute of Technology Patna
• Sriparna Saha, Department of Computer Science & Engineering Indian Institute of Technology Patna

On the Robustness of Face Recognition Algorithms Against Attacks and Bias

Face recognition algorithms have exhibited exceptional accuracy, indicating their appropriateness for practical use. Although these algorithms have improved accuracy, their ability to withstand attacks and prejudice has been questioned. This study outlines various challenges that might significantly impact the effectiveness of a face recognition algorithm, hence compromising its intended functionality. Multiple attacks have been examined, including physical presentation attacks, disguise/makeup, digital adversarial attacks, and morphing/tampering using GANs.

The researchers additionally analyze the impact of bias on facial recognition models and demonstrate that characteristics such as age and gender variations influence the effectiveness of contemporary algorithms. The report also outlines the possible causes for these difficulties and suggests future research avenues to enhance the resilience of face recognition models.

Researchers:

• Richa Singh, IIIT-Delhi, IIT-Jodhpur, India
• Akshay Agarwal, IIIT-Delhi
• Maneet Singh, IIIT-Delhi
• Shruti Nagpal, IIIT-Delhi
• Mayank Vatsa, IIIT-Delhi, IIT-Jodhpur, India

Guided Weak Supervision for Action Recognition with Scarce Data to Assess the Skills of Children with Autism

Testing autistic skills usually involves a physician repeatedly presenting various stimuli and documenting the child's reactions. This work suggests automating reaction measurement by using video capture of the scene after employing Deep Neural models for human action detection from videos. Nevertheless, training neural networks using supervised learning requires substantial quantities of annotated data, which might require more work.

This problem is solved by utilizing the 'similarities' between the action categories in publicly accessible extensive video action datasets and the dataset of interest. The proposed approach is guided by weak supervision, which involves matching each class in the target data to a class in the source data based on maximizing posterior likelihood. Afterwards, the classifier on the target data is trained again by adding samples from the corresponding source classes, coupled with a new loss that promotes separation between different classes. 

The suggested method is tested on two autistic skill assessment datasets, SSBD and a real-world autistic dataset consisting of 37 children of various ages and ethnicities who have been diagnosed with autism. Despite the limited data availability, their proposed strategy enhances the performance of the most advanced multi-class human action identification models.

Researchers

• Prashant Pandey, IIT Delhi
• Prathosh AP, IIT Delhi
• Manu Kohli, IIT Delhi
• Josh Pritchard, Florida Institute of Technology

Image source: Unsplash