Prishita Ray bio photo

Prishita Ray

Deep Learning Researcher and Open Source Developer

Email Google Scholar LinkedIn Github Stackoverflow ResearchGate ORCID

Publications

Workshops

2023

Improving Environment Robustness of Deep Reinforcement Learning Approaches for Autonomous Racing Using Bayesian Optimization-based Curriculum Learning
Learning Robot Super Autonomy Workshop @IROS 2023, Oct 2023
[Paper] [Code] [Poster] [Slides] [BibTeX]

Abstract

Deep reinforcement learning (RL) approaches have been broadly applied to a large number of robotics tasks, such as robot manipulation and autonomous driving. However, an open problem in deep RL is learning policies that are robust to variations in the environment, which is an important condition for such systems to be deployed into real-world, unstructured settings. Curriculum learning is one approach that has been applied to improve generalization performance in both supervised and reinforcement learning domains, but selecting the appropriate curriculum to achieve robustness can be a user-intensive process. In our work, we show that performing probabilistic inference of the underlying curriculum-reward function using Bayesian Optimization can be a promising technique for finding a robust curriculum. We demonstrate that a curriculum found with Bayesian optimization can outperform a vanilla deep RL agent and a hand-engineered curriculum in the domain of autonomous racing with obstacle avoidance.

2022

Charon: a FrameNet Annotation Tool for Multimodal Corpora
Proceedings of the 16th Linguistic Annotation Workshop (LAW-XVI) @LREC 2022, Jun 2022
*Acknowledged for GSoC 2020 contribution
[Paper] [BibTeX]

Abstract

Multimodality refers to the property of any communication phenomenon where two or more modes – defined as experientially recognized resources for meaning-making shaped by society and culture – are brought into play (Jewitt and Kress, 2003; Kress, 2010; Bateman et al., 2017). This paper approaches the expansion of FrameNet annotation into the multimodal domain, as proposed in Belcavello et al. (2020), by presenting Charon: a semi-automatic, human-in-the-loop tool for annotating static and dynamic images for semantic frames. Charon was developed to meet the following key requirements: (i) compatibility with existing FrameNet software; (ii) annotation of image with FrameNet categories; (iii) linkage of image and textual annotations.

Conference Proceedings

2023

An Intelligent RL-based Scheduler to Control Flooding in a Renewable Energy powered Automatic Water Dam control system
IEEE International Conference on Artificial Intelligence and Green Energy (ICAIGE), Oct 2023
[Paper] [Code] [Slides] [BibTeX]

Abstract

The use of non-renewable energy to operate large-scale systems is proving detrimental to the environment through pollution as well as leading to its depletion. As an alternative, the focus of industries, both private and public, is shifting towards the profitable use of renewable energy sources where it is available in abundance for such purposes. Water dams have been built above rivers to tap this natural resource to supply water to the nearby population. However, in the rainy season, excessive flooding wreaks havoc in the surrounding areas, leading to crop failures, damage to households and dam infrastructure. To counter this problem, this paper proposes an intelligent scheduler for an efficient renewable-energy powered automatic water dam control system which is modelled as a non-convex optimization problem. The scheduler uses the Soft Actor Critic with Emphasized Recent Experience and Prioritized Experience Replay Scheduler (SEPS), as the automatic Reinforcement Learning (RL) based scheduling algorithm to manage flooding in the control system through irrigation, hydroelectricity generation and reservoir storage such that it can sufficiently satisfy the needs of domestic households while also ensuring its efficient working. Simulated sensor readings record the current state of the environment and the proposed scheduler can make decisions based on these environmental variables with less human oversight and great accuracy. Experimental results clearly indicate that the SEPS scheduler outperforms its RL counterpart Proximal Policy Optimization Scheduler (PPOS) and the Natural Evolution Strategies Scheduler (NESS) in terms of average rewards obtained for managing floods efficiently and using renewable energy to run the control system.

2021

A New Combined Model with Reduced Label Dependency for Malware Classification
3rd International Conference on Integrated Intelligent Computing Communication & Security (ICIIC), Aug 2021
[Paper] [Code] [BibTeX]

Abstract

With the technological advancements in recent times, security threats caused by malware are increasing with no bounds. The first step performed by security analysts for the detection and mitigation of malware is its classification. This paper aims to classify network intrusion malware using new-age machine learning techniques with Reduced label dependency and identifies the most effective combination of feature selection and classification technique for this purpose. The proposed model, L2 Regularized Autoencoder Enabled Ladder Networks Classifier (RAELN-Classifier), is developed based on a combinatory analysis of various feature selection techniques like FSFC, variants of autoencoders and semi-supervised classification techniques such as ladder networks. The model is trained and tested over UNSW-NB15 and benchmark NSL-KDD datasets for accurate real time model performance evaluation using overall accuracy as well as per-class accuracy and was found to result in higher accuracy compared to similar baseline and state-of-the-art models.

2020

Stochastic Game Frameworks for Efficient Energy Management in Microgrid Networks
Innovative Smart Grids Technology (ISGT Europe), IEEE PES, Netherlands, Nov 2020
[Paper] [Code] [Slides] [BibTeX]

Abstract

We consider the problem of energy management in microgrid networks. A microgrid is capable of generating power from a renewable resource and is responsible for handling the demands of its dedicated customers. Owing to the variable nature of renewable generation and the demands of the customers, it becomes imperative that each microgrid optimally manages its energy. This involves intelligently scheduling the demands at the customer side, selling (when there is a surplus) and buying (when there is a deficit) the power from its neighboring microgrids depending on its current and future needs. In this work, we formulate the problems of demand and battery scheduling, energy trading and dynamic pricing (where we allow the microgrids to decide the price of the transaction depending on their current configuration of demand and renewable energy) in the framework of stochastic games. Subsequently, we propose a novel approach that makes use of independent learners Deep Q-learning algorithm to solve this problem.

Book Chapters

2021

Contemporary Developments and Technologies in Deep Learning based IoT
Deep Learning for Internet of Things Infrastructure, CRC Press, Taylor and Francis
[Paper] [BibTeX]

Abstract

Deep Learning and the Internet of Things are two of the most popular technologies today. Any complex application with any kind of input can be modeled using deep neural network architecture. Internet of Things (IoT) has enabled devices to connect with each other and share resources over the Internet. Multimodal input, be it in the form of text, images, video, audio, etc. obtained via sensors on the IoT devices can be processed. Moreover, to reduce computation load, edge computing and cloud-based deployment of applications are being enforced widely. Owing to hardware resource constraints of IoT infrastructure, proper design becomes important to ensure good performance. Security also becomes a key factor when data is shared using the cloud. Therefore, a discussion on current trends, technologies and challenges to be addressed when creating a DL based IoT application is presented below.