Research Areas

Prof. Utpal Garain at present is exploring Deep Learning methods for Language Engineering, Image and Video Processing. The current research topics of Prof. Garain are listed below:

1.  eXplainable Artificial Intelligence (XAI)
2.  Adversarial attacks on Deep learning systems
3.  Deep learning for Medical research
4.  Natural Language Processing (NLP)
5.  Document Image Analysis (DIA)
6.  Previous research

1. eXplainable Artificial Intelligence (XAI): The goal of this research to evaluate trustworthiness of AI systems for language vision tasks. Some of our recent research works have clearly showed the fragile nature of the otherwise high performance deep neural systems. Therefore, accuracy which has been so far viewed as measure for robustness is questioned and new metrics for measuring robustness is being explored. Scholars are investigating counterfactual analysis to find out reasoning ability of the current deep learning systems.

2. Adversarial attacks on Deep learning systems:In last several years a substantial research effort on our behalf brings out several weaknesses of the current deep neural systems through designing exciting adversarial attacks. Invariance-based attack is one such technique used to fool visual question-answering (VQA). Later, the research is extended for neural machine translation (NMT) systems. Defence strategies against such adversarial attacks have been designed for NMT systems.

3. Deep learning for Medical research: High performing deep neural systems have been designed for segmentation and classification tasks and new state of the art results are achieved for Psoriasis analysis. Detection of Munro’s Microabscess has been a subject of research and a multi-instance Capsule network (MICaps) has been designed. Analysis of peripheral blood smear images is being investigated through deep neural systems. Several novel datasets have bene generated for facilitating medical research. Symbolic AI is being merged with connectionist approach for analysis for ECG signal for cardio-related disease detection.

4. Natural Language Processing (NLP): Language generation is being explored to achieve style-guide generation and get rid of hallucination. In parallel, performance of current QA systems on conversational chats is analysed to understand the reasoning ability of the systems. Striking results are obtained for morphological analysis for under-resource languages and a state of the art neural lemmatizer has been developed and made available for the community (e.g., BenLem). AI and NLP tools are nicely integrated to solve a text-to-diagram conversion problem. In this research, machine is involved to draw the diagram described in a piece of text (e.g. geometric/physics problems).

5. Document Image Analysis (DIA):Research on Indic Script OCR is still continued. Designing of a recurrent neural network based system shows new state of the art results for Bengali handwriting recognition. A finite-state transducer (FST) based post-processing module has been developed for handwriting recognition. A novel method for automatic algorithm selection has been proposed for document image binarization. Research on recognition of handwritten mathematical expressions is continued. The CROHME (competition on recognition of handwritten mathematical expressions) initiative is still on and the sixth edition of CROHME was organized along with ICDAR 2019 (Int. Conf. on Document Analysis and Recognition). CROHME dataset is now endorsed by Technical Committee 11 (TC-11) of IAPR (Int. Assoc. for Pattern Recognition).

6. Previous Research: (i) Natural Language Processing (NLP): Under computational linguistics for Indic languages, two issues were researched: (i) anaphora resolution and (ii) lemmatization. Rule-based, MaxEnt, CRF and RNN-based approaches are being explored. Stanford POS tagger and MaltParser (dependency parser) have been retrained for Bengali language. A new initiative had been taken in the field of BioNLP. A novel probabilistic framework was tried for event extraction for cancer genetics.
(ii) Information Retrieval (IR): This work concerned with retrieval of OCR'd text. Unlike English OCRs, Indic OCRs are not very matured in producing high quality output and therefore, managing good information retrieval (IR) efficiency is a challenge while dealing with low quality OCR'd data. A probabilistic method was developed to model the OCR errors to help the IR engines. The framework was tested on a large dataset of Bengali and Hindi (Devanagari) OCR'd text. Because of this research a new initiative called RISOT (retrieval of Indic script OCR'd text) had been started under FIRE (Forum for Information Retrieval Evaluation) from 2011. IR from OCR'd text is now exploring cross language information retrieval (CLIR) issues where the queries are given in English and documents are in Bengali OCR'd text. A statistical transliteration module has been developed for transliterating out-of-vocabulary words. In another initiative, retrieval of imaged documents was also explored with an emphasis on retrieval of document images from compressed domain.
(iii)Computational Forensics(CF):This research was aimed at developing techniques for quick and easy authentication of security paper documents. Image processing and pattern recognition principles form the basis of this authentication technique. The goal was two-fold: (i) to check security features in a document in question in order to establish its authenticity, and at the same time (ii) analysis of security features to grade them according to their vulnerability against counterfeiting effort in order to help the designers for preparing of such security documents in future. Some research was done for authenticating Indian banknotes. The role of fluorescent pulp for detecting fake banknotes was investigated. A method for determination of ink age in old documents was developed. Handwriting analysis was used for manuscript dating, writer identification, and verification.
(iv)Other Pattern Recognition Tasks: Several others PR tasks like synthetic sample generation, artificial immune system (AIS) for PR tasks, compressed domain document processing, etc. are being studied.