I am a doctoral researcher in Computer Science at the University of Cambridge, supervised by Prof. Cecilia Mascolo. My work as a whole enables machine learning models to learn richer semantics of high-dimensional and complex data (signals, audio, images, text or other modalities). I am thankful to be supported by Jesus College Cambridge, the EPSRC, and the ERC.
My research is driven by doing more with less information. The most prominent bottleneck of deep learning today is access to labeled datasets, carefully curated for niche tasks. To this end, I work on data-efficient models that learn generalizable and personalized representations by leveraging the following fundamental paradigms:
- Self-supervision. I haven’t seen any parents teaching physics to an infant. Instead, gravity can only be learned by pure observation/interaction. Similarly, we align the input data in such a way that we pretend there is a part missing and then train a model to fill the gaps. But which is the best alignment?
- Multi-tasking. Humans perform thousands of tasks naturally, but our models are still painfully single-purpose. Bonus: with competing objectives you may discover some surprising properties of the tasks themselves.
- Transfer learning. We never start learning something new from scratch; we build upon basic abstractions instead. Our models can learn to re-use and fine-tune, but what are their limits?
Data-driven models of human behavior encode many complexities of the real world and hence I spend an inordinate amount of time thinking about sparsity, irregular sampling, long-term dependencies, noise, multi-modality, and long-tails. Beyond theory, I collaborate closely with world-class experts from other high-impact areas (health, natural and social sciences) to apply robust concepts from data science and accelerate scientific discovery.
Previously, during my studies I have been fortunate to work at diverse industries including multinational telcos (Telefonica Research), internet startups (Qustodio), retail tech companies (Ocado), and research labs. Further, our ongoing research in audio AI diagnostics (covid-19-sounds.org) has drawn international attention (covered by BBC, The Guardian, Forbes, The Times, El País).
🗞️ Feb 2021 — Our paper "Self-supervised transfer learning of physiological representations from free-living wearable data" has been accepted to ACM CHIL 2021! A new paper out of our COVID project, "Exploring automatic COVID-19 diagnosis via voice and symptoms from crowdsourced data", is accepted to IEEE ICASSP 2021. To give back to the community, we offer our data to the Computational Paralinguistics Challenge at Interspeech 2021.
🗞️ Jan 2021 — Our paper "SelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data" has been accepted to ACM IMWUT, while our opinion piece with Oxford/Turing "Digital Phenotyping and Sensitive Health Data: Implications for Data Governance" will appear in JAMIA.
Publications
2021
Self-supervised transfer learning of physiological representations from free-living wearable data
Dimitris Spathis, Ignacio Perez-Pozuelo, Soren Brage, Nicholas Wareham, Cecilia Mascolo
ACM Conference on Health, Inference, and Learning (CHIL), Virtual Event
Exploring Automatic COVID-19 Diagnosis via voice and symptoms from Crowdsourced Data
Jing Han, Chloë Brown*, Jagmohan Chauhan*, Andreas Grammenos*, Apinan Hasthanasombat*, Dimitris Spathis*, Tong Xia*, Pietro Cicuta, Cecilia Mascolo
International Conference on Acoustics, Speech, & Signal Processing
(ICASSP), Toronto, Canada
SelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data
Chi Ian Tang, Ignacio Perez-Pozuelo*, Dimitris Spathis*, Soren Brage, Nicholas Wareham, Cecilia Mascolo
Proc. on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT/Ubicomp), 5(1)
Digital Phenotyping and Sensitive Health Data: Implications for Data Governance
Ignacio Perez-Pozuelo, Dimitris Spathis, Jordan Gifford-Moore, Jessica Morley, Josh Cowls
Journal of the American Medical Informatics Association
Wearables, smartphones and artificial intelligence for digital phenotyping and health
Ignacio Perez-Pozuelo, Dimitris Spathis, Emma Clifton, Cecilia Mascolo
Digital Health, Chapter 3
2020
Exploring Automatic Diagnosis of COVID-19 from Crowdsourced Respiratory Sound Data
Chloë Brown*, Jagmohan Chauhan*, Andreas Grammenos*, Jing Han*, Apinan Hasthanasombat*, Dimitris Spathis*, Tong Xia*, Pietro Cicuta, Cecilia Mascolo
International Conference on Knowledge Discovery and Data Mining
(KDD), San Diego, USA
Learning Generalizable Physiological Representations from Large-scale Wearable Data
Dimitris Spathis, Ignacio Perez-Pozuelo, Soren Brage, Nicholas Wareham, Cecilia Mascolo
NeurIPS Machine Learning for Mobile Health workshop (ML4MH @ NeurIPS 2020), Vancouver, Canada
Exploring Contrastive Learning for Human Activity Recognition for Healthcare
Chi Ian Tang, Ignacio Perez-Pozuelo, Dimitris Spathis, Cecilia Mascolo
NeurIPS Machine Learning for Mobile Health workshop (ML4MH @ NeurIPS 2020), Vancouver, Canada
2019
Sequence Multi-task Learning to Forecast Mental Wellbeing from Sparse Self-reported Data
Dimitris Spathis, Sandra Servia, Katayoun Farrahi, Cecilia Mascolo, Jason Rentfrow
International Conference on Knowledge Discovery and Data Mining
(KDD), Anchorage, USA
Passive mobile sensing and psychological traits for large scale mood prediction
Dimitris Spathis, Sandra Servia, Katayoun Farrahi, Cecilia Mascolo, Jason Rentfrow
International Conference on Pervasive Computing Technologies for Healthcare (PervasiveHealth), Trento, Italy
Interactive dimensionality reduction
using similarity projections
Dimitris Spathis, Nikolaos Passalis, Anastasios Tefas
Knowledge-Based Systems, 165: 77-91
2018
Fast, Visual and Interactive Semi-supervised Dimensionality Reduction
Dimitris Spathis, Nikolaos Passalis, Anastasios Tefas
ECCV Efficient Feature Representation Learning workshop (CEFRL @ ECCV 2018), Munich, Germany
2017
Diagnosing Asthma and Chronic Obstructive Pulmonary Disease with Machine Learning
Dimitris Spathis, Panayiotis Vlamos
Health Informatics Journal, 25(3): 811–827 (issue published in 2019)
Class-based Prediction Errors to Detect Hate Speech with
Out-of-vocabulary Words
Joan Serra, Ilias Leontiadis, Dimitris Spathis, Gianluca Stringhini, Jeremy Blackburn, Athena Vakali
ACL Abusive Language Online workshop (ALW @ ACL 2017), Vancouver, Canada
2016
A comparison between semi-supervised and supervised text mining techniques on detecting irony in greek political tweets
Basilis Charalampakis, Dimitris Spathis, Elias Kouslis, Katia Kermanidis
Engineering Applications of Artificial Intelligence, 51: 50-57
2015
Detecting Irony on Greek Political Tweets: A Text Mining Approach
Basilis Charalampakis, Dimitris Spathis, Elias Kouslis, Katia Kermanidis
International Conference on Engineering Applications of Neural Networks, Rhodes, Greece
2014
Glocal News: An Attempt to Visualize the Discovery of Localized Top Local News, Globally
Dimitris Spathis, Theofilos Mouratidis, Spyros Sioutas, Athanasios Tsakalidis
International Conference on Conceptual Modeling, Hong Kong, China
Preprints
Improving the Definition of Depressed Mood with Digital Phenotyping
Maxime Taquet, Dimitris Spathis, Jason Rentfrow, Cecilia Mascolo, Guy M Goodwin
SSRN preprint, 3725630, 2020
Detecting sleep in free-living conditions without sleepdiaries: a device-agnostic, wearable heart rate sensing approach
Ignacio Perez-Pozuelo, Marius Posa, Dimitris Spathis, Kate Westgate, Nicholas Wareham, Cecilia Mascolo, Soren Brage, Joao Palotti
medRxiv preprint, 2020
Photo-Quality Evaluation based on Computational Aesthetics: Review of Feature Extraction Techniques
Dimitris Spathis
arXiv preprint, 1612.06259, 2016
Service
I take great joy in participating in the academic community, contributing and learning from peer review. I am serving on the following international conferences and journals:
Program Committee (PC): AAAI 2021, IJCAI 2020, KDD 2020 & 2021
Reviewer: Nature Scientific Reports, ICLR, ICML, AAAI, IJCAI, KDD, CHI, Ubicomp/IMWUT, ICASSP, Expert Systems with Applications, Neurocomputing, WWW/The Web Conference, Engineering Applications of Artificial Intelligence, ICWSM, ICPR, and more.
Invited Talks
Learning Generalizable Physiological Representations from Large-scale Wearable Data
- 📍 NeurIPS'20 ML4MH workshop (poster), Vancouver, Canada (remote) — December 12, 2020
- 📍 Barclays AI Eagle Lab, Cambridge, UK — March 12, 2020
- 📍 KDD'19, Anchorage, USA — August 4, 2019
- 📍 PervasiveHealth'19, Trento, Italy — May 20, 2019
- 📍 MRC Epidemiology Unit, University of Cambridge, UK — March 5, 2019
- 📍 Ocado, Barcelona, Spain / Hatfield, UK (remote) — July 10, 2019
- 📍 Facebook, PhD Open House (poster), London, UK — October 25, 2018
- 📍 MobiUK Symposium, Cambridge, UK — September 13, 2018
- 📍 ACL'17 Abusive Language workshop (poster), Vancouver, Canada — August 4, 2017
- 📍 EANN'15, Rhodes, Greece — September 25, 2015
Mentoring
I find it incredibly stimulating working with ambitious students in research projects. Some recent examples:
- Chi Ian Tang — Semi-supervised learning in human activity recognition (paper accepted)
- Benjamin Searle — Predicting compulsive behaviours with smartwatches (paper submitted)
- Kevalee Shah — Contrastive learning in unlabeled mobile health data (ongoing)
- Chuen Low — Pure attention models for timeseries (ongoing)
I have also taught small groups of undergrad students and evaluated their code assignments in lab sessions for the following courses:
Playground
Community pop references
Communitypoprefs.com is a data visualization project, where we present every pop-culture reference during the 5 seasons of the TV sitcom Community #sixseasonsandamovie.
Map out your music taste on Spotify
Visualizing my favourite songs on Spotify with statistics, non-linear dimensionality reduction and one-class learning. Published in Cuepoint Magazine .
How do popular book authors use language differently?
Text mining on Game of Thrones, Harry Potter, Hunger Games and Lord of the Rings books. Featured in Medium's Editor Picks .
Share your kid’s pictures online without worrying
Mobile app with face recognition, age estimation, & emotion recognition to blur underage people in pictures or replace their face with emotion-based emoji. Developed during HackZurich 2018.
Discover top local news globally
Glocalne.ws was a mashup of Google News and Google Maps. Unfortunately now defunct due to API discontinuance.
Composing music and text with Recurrent Neural Networks
Training deep neural networks on massive amounts of musical notation (Irish folk) and literature (Shakespeare's oeuvre) and letting them create their own art. Essay in Greek but you can still see/listen to the results.
Personal
Non-academic things about me: I love music, both playing and listening. I am mostly into art rock and indie folk, with the occasional exception of some well-crafted pop. Although I am an accordionist by training, over the last few years I've been playing mostly piano and ukulele. In a previous life, I performed in classical orchestras, cover bands and critically acclaimed indie bands like The Children of the Oldness (aka Kore Ydro). You can listen to the album I participated, "Consortium in Amato", here.
I also enjoy street photography and in particular playing with light—photography comes from Greek φως (light) and γραφή (writing), or drawing with light. A subset of my pictures is on Flickr while I had my 15" of fame when one of my landscapes was featured in the Huffington Post.
Lastly, and perhaps most importantly, I'm always on the lookout for ways to move items from the "non-academic list" to the "academic list"—let me know if you'd like to help!