**PhD Project Topics 2021-2022**

The PhD topics I offer are in the area of neural information retrieval. Below are some PhD topics, however these are just ideas and can be tuned to match both our interests.

If you are a self-funded student considering a PhD in any of the topics below please email me to discuss before applying. Click here for information on applying for Research Degrees at Loughborough University.

I am recruiting sponsored or self-funded PhD students who wish to undertake projects in data science and A.I with focus on information retrieval, including projects within the topics below and here:

Multi-modal neural information retrieval, fusion and summarisation (*NEW*)
The project concerns the development of algorithms and a tool for multi-modal neural information retrieval and summarisation of results obtained from multi-modal user queries. The project will develop methods to:
– retrieve relevant multi-modal information sorted in ranked order of relevance to the user’s query;
– develop algorithms that generate multi-modal summaries comprising of text aligned with images that contain information presented in a meaningful order that best answers the user’s query;
– create Query Reformulation Based on Relevance Feedback functionality that allows users to revise their multi-modal queries by selecting relevant multi-modal results.

Topic 1: Information Retrieval of similar source-code files and fragments using Deep and Transfer Learning. This exciting and current topic concerns the development and training of deep-learning based algorithms to detect similarity between source-code files for information retrieval purposes. The objectives are: 1) train deep learning algorithms to learn the similarity between different code fragments; 2) given a set of comments and corresponding code fragments train an algorithm to retrieve relevant source-code for text-code-text retrieval; 3) extend the functionality to cross programming language text-code-text retrieval using transfer learning; 4) case studies using large data repositories. To apply use this link and specify the topic in the application.

Topic 2: Multi-modal Transfer Learning for Cross-Modal Information Retrieval. Deep learning is a subset of machine learning in artificial intelligence (AI) that is capable of learning from data. Deep learning is receiving a lot of attention due to its ability to achieve unprecedented levels of performance in terms of accuracy and speed, to the point where deep learning algorithms can outperform humans at decision making, and tasks such as classifying images, and real-time detection. Multi-modal data fusion is an important task for many machine learning applications, including human activity recognition, information retrieval, and real-time applications of A.I. For example, datasets can include audio and visual data; image and sensor data; multi-sensor multi-modal data; and text and image data. Multi-modal data can be complex, noisy and imbalanced, particularly when collected from real environments. Therefore, it is a significant challenge to create deep learning models using multi-modal data. Training deep neural networks to learn a very accurate mapping from inputs (unimodal or multimodal) to outputs (e.g. labels also known as classes) requires large amounts of labelled data. Even when these models are trained, they have limited ability to generalise to conditions that are different to the ones used for training the model. Cross-Modal Retrieval is the task of retrieving data across different modalities. such as image-text, video-text, and audio-text. It is a challenging task to generate new representations from different modalities in the shared subspace. The project may focus on:
(1) Investigating bias in transfer learning; metrics for detecting, preventing, and mitigating bias in unimodal and multimodal data
(2) Investigating bias in multi-modal transfer learning for cross-modal data retrieval
(3) Feature extraction and fusion of multi-modal data for information retrieval, query expansion and relevance feedback or question answering
(4) Application of the proposed methods to various multi-modal datasets for cross-modal retrieval tasks.

Topic 3: Using AI to ethically describe images using text. This project involves the development of algorithms for learning image and text embeddings. The idea behind this project is to 1) develop methods towards accessible, explainable and ethical AI; 2) explore the suitability and limitation of text generation algorithms such as GPT-2 and GPT-3; 2) to develop algorithms that can be used to describe images for users with visual disabilities. Deep Learning and Information Retrieval methods will be implemented. The proposed methods and tools will be applied to a real-world application in collaboration with project partners.

Topic 4: Retrieval of video keyframes and textual video summarisation. This project involves the development of algorithms for generating summaries of events using key frames extracted from videos. The project will commence with developing methods for 1) selecting keyframes automatically and semi-automatically using text queries; and then 2) generating textual summaries describing the keyframes. The project will involve the development of methods from the AI (deep learning) and informational retrieval domains. The proposed methods and tools will be applied to a real-world application in collaboration with project partners. Applications on human activity recognition and surveillance.

Topic 5: Ethical Continual/Lifelong Deep Learning for multi-modal, cross-modal information retrieval: Training deep neural networks to learn a very accurate mapping from inputs (such as image data, sensor data, text) to outputs (e.g. labels also known as classes) requires large amounts of labelled data. Even when these models are trained, they have limited ability to generalise to conditions that are different to the ones used for training the model. Projects under this topic concern the development of Continual/Lifelong learning algorithms that can learn continuously and adaptively, to autonomously and incrementally develop new knowledge.

Topic 6: Multimodal data fusion using Deep Learning for information retrieval: Deep learning is a subset of machine learning in artificial intelligence (AI) that is capable of learning from data. Deep learning is receiving a lot of attention due to its ability to achieve unprecedented levels of performance in terms of accuracy and speed, to the point where deep learning algorithms can outperform humans at decision making, and tasks such as classifying images, and real-time detection. Multi-modal data fusion is an important task for many machine learning applications, including human activity recognition, information retrieval, and real-time applications of A.I. For example, datasets can include audio and visual data; image and sensor data; multi-sensor multi-modal data; and text and image data. Multi-modal data can be complex, noisy and imbalanced, particularly when collected from real environments. Therefore, it is a significant challenge to create deep learning models that can classify multi-modal data. Machine learning models designed to classify imbalanced data are biased toward learning the more commonly occurring classes. Such bias occurs naturally since the models better learn classes which contain more records, a similar process that would occur with human learning. The aim of this project is to: (1) devise feature engineering approaches for multi-modal data; (2) identify whether fusing multi-modal data improves results as opposed to using uni-modal datasets in specific machine learning tasks; and (3) to develop computational approaches and methods for fusing imbalanced multi-modal data.

Topic 8: Information Retrieval. I very much welcome ideas in Natural Language Processing and in Information Retrieval. I am interested in using Deep Learning methods for improving relevance feedback; multi-modal information retrieval; topic modelling; and cross-lingual information retrieval; and open to suggestions.