Machine Learning Special Interest Group Research at IISc is housed in Department of Computer Science and Automation which provides publications in machine learning and related fields. MLSIG conducts research on a variety of aspects of machine learning and related fields, ranging from theoretical foundations to new algorithms as well as several applications. The following are some examples of recent research topics explored in the group:

• Support Vector Machines and Kernel based Learning Methods
• Structured Prediction
• Ranking
• Clustering
• Graphical Models
• Robust Learning Methods for Uncertain Data
• Statistical Learning Theory, Statistical Consistency of Learning Algorithms
• Machine Learning in Bioinformatics
• Machine Learning in Computer Vision
• Machine Learning in Text Mining

Machine Learning and Learning Theory Group current research directions include designing and analyzing algorithms for problems such as ranking and various types of structured prediction tasks, understanding statistical consistency properties for such problems, exploring new issues in machine learning such as those related to privacy, and selected applications of machine learning in computational biology and medicine

Machine and Language learning lab primary research goals is to extract, orgazine, and make readily available the knowledge trapped inside such unstructured text data on a large scale. It’s research spans the areas of Machine Learning and Natural Language Processing.

Machine learning models are susceptible to adversarial perturbations: small changes to input that can cause large changes in output. It is also demonstrated that there exist input-agnostic perturbations, called universal adversarial perturbations, which can change the inference of target model on most of the data samples.
Existing methods to craft universal perturbations are -

• Task specific
• Require samples from the training data distribution
• Perform complex optimizations, because of the data dependence, fooling ability of the crafted perturbations is proportional to the available training data.
• Generalizable and data-free approach for crafting universal adversarial perturbations.
• The objective is Independent of the underlying task & achieves fooling via corrupting the extracted features at multiple layers.
• Proposed objective is generalizable to craft image-agnostic perturbations across multiple vision tasks such as object recognition, semantic segmentation, and depth estimation.
• In the practical setting of black-box attack scenario (when the attacker does not have access to the target model and its training data), objective outperforms the data dependent objectives to fool the learned models.
• Objective remarkably boosts the fooling ability of the crafted perturbations by further exploiting simple priors related to the data distribution.
• Significant fooling rates achieved by the objective to emphasize that the current deep learning models are now at an increased risk, since the objective generalizes across multiple tasks without the requirement of training data for crafting the perturbations.

Increase in the number of sources of data, research in cross-modal matching is becoming an increasingly important area. Several applications like match- ing text with image, matching near infra-red images with visible images example: For matching face images captured during night-time or low-light conditions to standard visible light images in the database), matching sketch images with pictures for forensic applications.

• Novel algorithms are under development for this problem, which is extremely challenging, due to significant differences between data from different modalities.
• Semantic preserving hashing technique has been developed in specific for cross-modal retrieval algorithms, which can work seamlessly for single and multi-label data, as well as in paired and unpaired scenarios.
• The algorithm obtained state-of-the-art performance in different applications.

Understanding the representational power of Restricted Boltzmann Machines (RBMs) with multiple layers is an ill-understood problem and is an area of active research. Motivated from the approach of Inherent Structure formalism (Stillinger & Weber, 1982), extensively used in analyzing Spin Glasses, a novel measure has been proposed called Inherent Structure Capacity (ISC), which characterizes the representation capacity of a fixed architecture RBM by the expected number of modes of distributions emanating from the RBM with parameters drawn from a prior distribution.

• Lean RBMs are nothing but multi-layer RBMs where each layer can have at-most O(n) units with the number of visible units being n.
• For every single layer RBM with Omega(n^{2+r}), r >= 0, hidden units there exists a two-layered lean RBM with Theta(n^2) parameters with the same ISC, establishing that 2 layer RBMs can achieve the same representational power as single-layer RBMs but using far fewer number of parameters.

Computation outcomes in economic systems, which are efficient and fair at the same time. This initiative is headed by Siddharth Barman.

• Incompatible properties of efficiency and fairness can be achieved together.
• Fairness is a fundamental consideration in many resource-allocation problems, results can potentially influence allocation policies in practical settings.

Automatically synthesizing programs and analyzing their behavior is considered a holy grail of Artificial Intelligence: Neural network-based solutions have been developed towards automating tasks. Deep Neural Networks were employed to identify different vehicles in videos obtained from traffic junctions
This initiative is co-headed by Aditya Kanade and Shirish Shevade.

Investigating the use of AI/ML techniques for communications related applications: Techniques to solve hard problems by training a neural network. Key Investigations under progress - Limited training size & Model generalizability
This initiative is headed by Dr. Chandra R. Murthy.

Established by the Indian Institute of Science (IISc), ARTPARK is a not for profit in Bengaluru. It is also supported by AI Foundry in a public-private model. With seed funding of $22mn from the Department of Science under the National Mission on Inter-disciplinary Cyber-Physical Systems (NM-ICPS), ARTPARK will bring about collaborative consortium of partners from industry, academia and government bodies. This will lead to cutting-edge innovations in terms of new technologies, standards, products, services and intellectual properties, said the DST

Develop an AI-based framework that could aid in the real-time environment-aware high-fidelity simulation of a windfarm using Deep Learning, Model Order Reduction, High-performance computing.

Challenges

Solution 

One of the major functions of the neural network model is that it can reduce the cost function and eliminate the loss, and we will be able to get the solution. As the initial process changes, we face many challenges when we train the model. We will encounter many scientific challenges when we use machine learning for scientific applications. The primary objective of using machine learning in scientific computing is to eliminate the high computational cost. Therefore, they focus only on transferring the crucial information to build the model. This process is called federated learning. IISc intends to develop an AI-based framework that can help in the real-time environment-aware high-fidelity simulation of a wind farm using Deep Learning, Model Order Reduction, and High-performance computing. However, one of the challenges faced during the process is that these models need to be retrained for changes in simulation parameters which will result in downtime of the framework. Hence there is a need to generate a robust “Digital Shadow” framework which could reliably perform simulation on real-time data with good accuracy and considerable downtime. 

Problem statement

• To develop more complex NN models to predict the stabilization parameter on a wider range of perturbed problems.
• To make use of existing mathematical methods such as L2-Error minimisation within the training of Neural Networks

Particle Deposition prediction in the human air pathway using ML.

Research Statement

• To model aerosol deposition on human air pathways using Computational models
• To develop ML models which could predict particle deposition within human air pathways for varying input scenarios

The use of supervised ANNs to improve multi-dimensional extension of current state-of-the-art discontinuous Galerkin solvers which are being used in industry and academic community.With ANNs as black box for inner-element shock capturing, we can improve the robustness and accuracy of current DG methods.