Dr. Thomas Jumperkables Winterbottom

A publication planned for the outcomes of my current KTP, pertinent details of which are currently behind NDA for now, nonetheless publication planned before 2025.
                    

The final project of my PhD thesis continued. I found interesting results using neurolinguistic scores of similarity in a VQA classification context. However, an expanded study considering the effects of word norms in LLM benchmarks of today is currently in the works. Aiming to have a version on arXiv before the 2025.
                    

A project from 2024 that is ongoing in my spare time. I have aligned a very large dataset of air pollution data using WAQI as a base, but crucially aligned (as best i can) with weather data, surrounding landuse statistics , OpenStreetMap data, and elevation. I will be looking to publish a paper on my work here in the coming months. arXiv preprint will be ready soon.
                    

@misc{winterbottom2024powernextframepredictionlearning,
    title={The Power of Next-Frame Prediction for Learning Physical Laws}, 
    author={Thomas Winterbottom and G. Thomas Hudson and Daniel Kluvanec and Dean Slack and Jamie Sterling and Junjie Shentu and Chenghao Xiao and Zheming Zuo and Noura Al Moubayed},
    year={2024},
    eprint={2405.17450},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2405.17450}, 
}
                    

A project from early 2020 that was ongoing in various forms. The original inspiration was essentially "languauge modelling for vision" which was more novel back in 2020. But the project grew to become more of an analytic exploration of what we can quantifiably and qualitatively show that next frame prediction induces. I led the experiments and writing for this paper's work, but there are several works spawning from the original led by my co-authors in the works for the near future.
                    

@article{Winterbottom2022ADL,
    title={A deep learning approach to fight illicit trafficking of antiquities using artefact instance classification},
    author={Thomas Winterbottom and Anna Leone and Noura Al Moubayed},
    journal={Scientific Reports},
    year={2022},
    volume={12}
}
                    

A project from 2022 from a smaller part time postdoc position. The work represents the beginnings of a proof of concept for using CNNs for doing arefact detection. The results were pretty good, with an increased dataset size allowing me to expand the scope of the classification setup I think there are promising implications here.
                    

@inproceedings{mbintvqa,
    title={Bilinear pooling in video-QA: empirical challenges and motivational drift from neurological parallels},
    author={Winterbottom, T. and Xiao, S. and McLean, A. and Al Moubayed, N.},
    booktitle={PeerJ},
    year={2022}
}
                    

My second publication during my PhD. Quite a long one, this broadly has two cores to it. I wanted to take a critical look at bilinear pooling, a method for performing feature fusion in neural networks particularly popular in the nascent multimodal ML field of 2020. Authors were iterating on various methods of bilinear pooling with clever maths to bring the memory cost down and performance up, which is invaluable research. However, I noticed interestingly that some of the parallels to certain neurological theories that some of the earlier bilinear techniques had were fading, and the justification of these news techniques is mostly just improved performance. I wanted to be careful to keep my core message of "please don't overstate results", and "hey there are interesting parallels to some neuroscience here", without drifting too far into non-ML areas I didn't have grounding in. I'm happy with this paper, but wanted to focus more on implementation for the next few things.
                    

@inproceedings{mbintvqa,
    title={On Modality Bias in the TVQA Dataset},
    author={Winterbottom, T. and Xiao, S. and McLean, A. and Al Moubayed, N.},
    booktitle={Proceedings of the British Machine Vision Conference ({BMVC})},
    year={2020}
}
                    

An analysis and quantification of text shortcuts in TVQA, a large and impressive video-QA dataset from around 2019. Broadly, I wanted to emphasise care should be taken in using such shortcuts. Can one say a reliance on these shortcuts is indeed multimodal? I show that the then-state-of-the-art model actually supressed the vision inputs when they were trained together. I also outlined an inclusion-exclusion method of sorts to try and isolate subsets of a given dataset that can be answered with various modality combinations. My first ever publication, during my PhD, and my work quality has naturally improved. I hope readers understand that TVQA is still indeed a very good and valuable dataset. :)
                    

Ask the very kind first author for the code.
                    

@article{Gajbhiye_2020,
    title={Bilinear Fusion of Commonsense Knowledge with Attention-Based NLI Models},
    ISBN={9783030616090},
    ISSN={1611-3349},
    url={http://dx.doi.org/10.1007/978-3-030-61609-0_50},
    DOI={10.1007/978-3-030-61609-0_50},
    journal={Lecture Notes in Computer Science},
    publisher={Springer International Publishing},
    author={Gajbhiye, Amit and Winterbottom, Thomas and Al Moubayed, Noura and Bradley, Steven},
    year={2020},
    pages={633–646}
}
                    

I helped Amit integrate bilinear pooling into his model. Funtimes, he's a great guy.
                    

@phdthesis{winterbottom2023,
  author       = {Thomas Winterbottom},
  title        = {Scalable Methodologies and Analyses for Modality Bias and Feature Exploitation in Language-Vision Multimodal Deep Learning},
  school       = {Durham University},
  year         = {2023},
  type         = {PhD thesis},
  note         = {Unpublished, successfully defended},
  url          = {https://etheses.dur.ac.uk/14891/},
}

                    

Hard to TL;DR an entire PhD. It has 4 contribution chapters. 2 positive results, 2 negative results (but the negative results are IMO sufficiently explored). Was a fantastic time. Basically, its important to me that machine learning research try to keep an eye to being applicable to future dataset and model designs.
                    

Worked on the cutting edge of skin cancer ML with a large multimodal dataset of skin cancer conditions and diagnoses, integrating text into the image classification task, and controlling for various shortcuts and biases.
                    

A short term position where I worked on air pollution data in machine learning. To date, this has culminated in me aligning a very large dataset of air pollution data using WAQI as a base, but crucially aligned (as best i can) with weather data, surrounding landuse statistics , OpenStreetMap data, and elevation. Aligning these data sources has its problems, but I'm proud of the dataset, and once I generate reliable results for a baseline, I will be looking to publish a paper on my work here. This work is ongoing in my spare time.
                    

A short time but full time internship where I worked on a proof of concept for a multimodal classifier for skin conditions. The proof of concept yielded promising initial results, that were later expanded upon during my KTP in 2024.
                    

This one is quite self explanatory. I worked on a proof-of-concept using CNNs for detecting known artefacts using Durham's Oriental Museum data. The results were high for for a proof on concept, and it was nice to get an early look into PDRA life during the end of my PhD. I got a Nature's Scientific Reports paper out of it too.
                    

A bit of demonstration as a PhD as is relatively standard at Durham. I demonstrated for computer systems, and learning python. Demonstration is really good fun and I found myself giving a lot to it, and had a lot of explicit feedback from students that I helped them a lot.
                    

A summer internship after my ungraduate but before my PhD where I focused on topic modelling. It was an interesting time, and the SMTM technique was relatively novel, giving me an early taste of independent research.
                    

I spent a good amount of seasonal and summer time working as a labourer with my step dad. Lots of painting, prep work, concreting, digging, grouting, and a touch of brick work and carpentry. Great work I feel very lucky to have experienced, and has fuelled my DIY instincts.
                    

Fast food, restaurant, cafe, and pot wash experience all rolled into one. Genuinely was a lovely place to work.