Andrej Jovanović

Research

Read research overview

At present, my machine learning interests are centred around distributed, privacy-preserving, efficient and collaborative machine learning. In particular, I have a great affinity towards paradigms such as federated learning and neural network compression, and the interplay this has with optimisation theory. Generally speaking, the motivating factor for this is that I am circumspect towards the promise of centralised machine learning paradigms, which require enormous amounts of data and computational resources to achieve their state-of-the-art performance. Instead, I believe that there is much benefit to be derived from investigating alternate paradigms that can scale while promoting privacy inherently, supported by sound mathematical foundations.

If you find any of my research interesting, or you would like to collaborate on any ideas, please do not hesitate to reach out.

search papers: order by:

year	title	authors	tags	paper	misc
2026	LoRDO: Distributed Low-Rank Optimization with Infrequent Communication	jovanović, iacob, safaryan, modoranu, sani, shen, qiu, alistarh, lane	🌳📈	arXiv
2025	MT-DAO: Multi-Timescale Distributed Adaptive Optimizers with Local Updates	iacob, jovanović, safaryan, kurmanji, sani, horváth, shen, qiu, lane	🌳📈	ICLR 2026
2025	DES-LOC: Desynced Low Communication Adaptive Optimizers for Training Foundation Models	iacob, sani, safaryan, giampouras, horváth, jovanović, kurmanji, aleksandrov, shen, qiu, lane	🌳📈	ICLR 2026
2025	Position: It's Time to Act on the Risk of Efficient Personalized Text Generation	iofinova, jovanović, alistarh,	🔖	arxiv
2025	Panza: Design and Analysis of a Fully-Local Personalized Text Writing Assistant	nicolicioiu, iofinova, jovanović, kurtic, nikdan, panferov, markov, shavit, alistarh	💻🗣️🌀	CPAL 2026 and ICLR 2025 Workshop on Foundation Models in the Wild	bonus video
2024	Second-Order Optimisation and Imbalanced Class Distribution in Emotional Analysis	jovanović	🌀📈	MPhil Thesis @ Cambridge
2023	Rumour Detection in the Wild: A Browser Extension for Twitter	jovanović & ross	💻🗣️🌀	NLP-OSS @ EMNLP2023	ACL
2023	EDAC: Efficient Deployment of Audio Classification Models For COVID-19 Detection	jovanović^, mihaly^ & donaldson^*	🌳🌀	arxiv preprint	More info... Awarded second place IBM prize in MLP course awards at UoE.

October 2025 - Present

the university of cambridge

(phd) computer science specialising in federated (pre)training of large language models with, advised by Nic Lane in the CaMLSys group.
October 2023 - August 2024

the university of cambridge

(mphil) machine learning and machine intelligence specialising in speech and language processing. Thesis exploring Second-order Optimisation and Imbalanced Class Distribution in Emotion Analysis under the supervision of Brian (Guangzhi) Sun.
Sept 2019 - May 2023

the university of edinburgh

bsc(hons) artificial intelligence and computer science. thesis completed under the supervision of björn ross

May 2025 - Oct 2025

research assistant | camlsys & flower labs

Conducting research focusing on federated learning applied to the (pre)training of large scale models.
Oct 2024 - March 2025

research assistant | institute of science and technology austria Alistarh Group

Working as a research assistant under the supervision of Dan Alistarh in the Distributed Algorithms and Systems (DAS) group focusing on compressed federated (pre)training of foundation models.
Summer 2023 - Summer 2023

app. ai research intern | jp morgan chase & co. (Applied Innovation of Artificial Intelligence (AI²))

Developed unsupervised learning methods, with an associated parallel data processing pipeline, for anomaly detection which increased previous performance by 30%.
Summer 2022 - Summer 2022

ml eng intern | jp morgan chase & co. (Applied Innovation of Artificial Intelligence (AI²))

Pursued personal research in Natural Language Processing, building a search engine, powered by dense passage retrieval, with q&a capabilities. Assisted researchers in delivering proof of concept projects, onboarding them to CI/CD platforms.