This collection features the year's most intriguing research publications in artificial intelligence and data science in India. It is organized chronologically and includes links to in-depth articles and source code, where available. The collection highlights significant AI research articles and provides a hand-picked selection of recent developments in the field.

Adaptive mixing of auxiliary losses in supervised learning

Researchers: 

• Durga Sivasubramanian - Indian Institute of Technology Bombay, Google Research, India.
• Ayush Maheshwari - Indian Institute of Technology Bombay, 
• Pradeep Shenoy - Indian Institute of Science, Bengaluru, 
• Prathosh AP - Google Research, India, 
• Ganesh Ramakrishnan - Indian Institute of Technology Bombay

Auxiliary losses are commonly employed in many supervised learning scenarios to include supplementary information or limitations into the primary supervised learning objective. Knowledge distillation, for example, seeks to replicate the outputs of a competent teacher model. Similarly, in rule-based techniques, weak labelling information is supplied by labelling functions that may be noisy rule-based approximations of the actual labels. 

The researchers address the issue of acquiring the ability to integrate these losses systematically. Their method, AMAL, employs a bi-level optimization criterion on validation data to develop optimal mixing weights at an instance level throughout the training data. The authors present a meta-learning strategy for addressing this bi-level objective and demonstrate its applicability to supervised learning settings. Experiments conducted in several domains of knowledge distillation and rule-denoising demonstrate that AMAL yields significant improvements compared to other competing methods. They systematically analyse their process and provide insights into the mechanisms that lead to improved performance.

Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks Using an Incompetent Teacher

Researchers

• Vikram S Chundawat - Mavvex Labs, India, 
• Ayush K Tarun - Mavvex Labs, India, 
• Murari Mandal - School of Computer Engineering, Kalinga Institute of Industrial Technology Bhubaneswar, 
• Mohan Kankanhalli - School of Computing, National University of Singapore

Machine unlearning has gained significance due to the growing requirement for machine learning (ML) systems to adhere to rising data privacy rules. It enables the removal of specific sets or classes of data from a pre-trained machine-learning model without the need to retrain the model from the beginning. Recently, numerous endeavours have been made to enhance the effectiveness and efficiency of the process of unlearning. 

The authors suggest a new technique for machine unlearning that involves examining the effectiveness of skilled and unskilled teachers within a student-teacher paradigm to promote amnesia. The knowledge imparted by proficient and inept teachers is selectively sent to the learner to create a model devoid of any information about forgotten data. The researchers empirically demonstrate that this strategy exhibits strong generalization and is characterized by rapidity and efficacy. 

In addition, they propose using the zero retrain forgetting (ZRF) measure to assess the effectiveness of any unlearning technique. Contrary to the current unlearning metrics, the ZRF score is independent of the presence of the costly retrained model. This feature is beneficial for analyzing the model that has yet to be learned, even after deployment. The researchers describe the findings of experiments conducted on several deep networks and across diverse application domains to study random subset forgetting and class forgetting.

City-Scale Pollution Aware Traffic Routing by Sampling Max Flows Using MCMC

Researchers

• Shreevignesh Suriyanarayanan - Machine Learning Lab, IIIT Hyderabad, 
• Praveen Paruchuri - Machine Learning Lab, IIIT Hyderabad, 
• Girish Varma - Machine Learning Lab, IIIT Hyderabad

The substantial influx of vehicular traffic is the primary factor contributing to global air pollution in urban areas. Prolonged exposure to extreme pollution can lead to significant health complications. An effective strategy for addressing this issue involves developing a pollution-aware traffic routing policy that achieves a balance between several objectives:

• Preventing excessive pollution in any given region
• Ensuring quick transit times
• Maximizing the utilization of road capabilities

The researchers suggest an innovative sampling-based method for addressing this challenge. Their work presents the initial implementation of a Markov Chain capable of generating integer max flow solutions for a planar graph. The probability of these solutions is guaranteed to depend on the overall transit length. Using the SUMO traffic simulator, the researchers devised a traffic strategy by utilizing several samples and conducting simulated traffic on authentic road maps. When using maps of large cities worldwide, they noticed a significant reduction in areas with severe pollution compared to other methods.

Image source: Unsplash