About Me

I am a postdoctoral fellow at the University of Cambridge, working in the CaMLSys lab with Prof. Nic Lane. My research focuses on machine unlearning, with applications in privacy and trustworthy AI across diverse domains such as computer vision, language models, and tabular data, in both centralized and federated learning environments. I completed my PhD at the University of Warwick under the supervision of Prof. Peter Triantafillou, where I worked on Learned Database systems. My PhD research provided novel and robust solutions for updating machine learning models in databases, addressing both continual learning of new data and unlearning of old data. I am always keen to collaborate. feel free to reach out!

News

  • February 2025: Our Proposal got into the SPRIN-D's challenge. We won a grant worth €530k! I'll be co-leading the team with Nic! Link
  • January 2025: One paper is accepted at ICLR '25 pdf
  • September 2024: One paper is accepted at NeurIPS '24 pdf
  • July 2024: I started a Postdoc at the University of Cambridge in the CaMLSys's lab
  • June 2024: New Preprint: Machine Unlearning Competition Post Analysis pdf
  • April 2024: I sucessfully defended my PhD!
  • November 2023: One paper is accepted at SIGMOD '24 pdf
  • September 2023: One paper is accepted at NeurIPS '23 pdf
  • September 2023: Our Neurips23 Machine Unlearning Competition is now live on Kaggle!
  • December 2022: One paper is accepted at ACM SIGMOD '23 pdf
  • January 2022: One paper is accepted at CIDR '21 pdf
  • December 2021: One paper is accepted at ACM SIGMOD '21 pdf
  • October 2020: I started a PhD in Machine Learning and Data Systems at the University of Warwick, under the supervision of Prof. Peter Triantafillou

Research

HTML5 Bootstrap Template by colorlib.com

What makes unlearning hard and what to do about it

NeurIPS 2024 | K Zhao *, M Kurmanji, G Bărbulescu, E Triantafillou, P Triantafillou

Machine unlearning is the problem of removing the effect of a subset of training data (the “forget set”) from a trained model e.g. to comply with users’ requests to delete their data, or remove mislabeled, poisoned or otherwise problematic data. With unlearning research still being at its infancy, many fundamental open questions exist: Are there interpretable characteristics of forget sets that substantially affect the difficulty of the problem? How do these characteristics affect different state-of-the-art algorithms? We present the first investigation into these questions. We identify two key factors affecting unlearning difficulty and the performance of unlearning algorithms. Our evaluation on forget sets that isolate these identified factors reveals previously-unknown behaviours of state-of-the-art algorithms that don’t materialize on random forget sets. Based on our insights, we develop a framework coined Refined-Unlearning Meta-algorithm (RUM) that encompasses: (i) refining the forget set into homogenized subsets, according to different characteristics; and (ii) a meta-algorithm that employs existing algorithms to unlearn each subset and finally delivers a model that has unlearned the overall forget set. RUM substantially improves top-performing unlearning algorithms. Overall, we view our work as an important step in deepening our scientific understanding of unlearning and revealing new pathways to improving the state-of-the-art.

HTML5 Bootstrap Template by colorlib.com

Towards Unbounded Machine Unlearning

NeurIPS 2023 | M Kurmanji*, P Triantafillou, Jamie Hayes, E Triantafillou

Deep machine unlearning is the problem of removing the influence of a cohort of data from the weights of a trained deep model. This challenge is enjoying increasing attention due to the widespread use of neural networks in applications involving user data: allowing users to exercise their ‘right to be forgotten’ necessitates an effective unlearning algorithm. However, deleting data from models is also of interest in practice for other applications where individual user privacy is not necessarily a consideration: removing biases, outof-date examples, outliers, or noisy labels, and different such applications come with different desiderata. We propose a new unlearning algorithm (coined SCRUB) and conduct a comprehensive experimental evaluation against several previous state-of-the-art models. The results reveal that SCRUB is consistently a top performer across three different metrics for measuring unlearning quality, reflecting different application scenarios, while not degrading the model’s performance.

HTML5 Bootstrap Template by colorlib.com

Machine Unlearning in Learned Databases

SIGMOD 2024 | M Kurmanji*, E Triantafillou, P Triantafillou

Deep Machine learning models based on neural networks (NNs) are enjoying ever-increasing attention in the Database (DB) community, both in research and practice. However, an important issue has been largely overlooked, namely the challenge of dealing with the inherent, highly dynamic nature of DBs, where data updates are fundamental, highly-frequent operations (unlike, for instance, in ML classification tasks). Although some recent research has addressed the issues of maintaining updated NN models in the presence of new data insertions, the effects of data deletions (a.k.a., "machine unlearning") remain a blind spot. With this work, for the first time to our knowledge, we pose and answer the following key questions: What is the effect of unlearning algorithms on NN-based DB models? How do these effects translate to effects on key downstream DB tasks, such as cardinality/selectivity estimation (SE), approximate query processing (AQP), data generation (DG), and upstream tasks like data classification (DC)? What metrics should we use to assess the impact and efficacy of unlearning algorithms in learned DBs? Is the problem of (and solutions for) machine unlearning in DBs different from that of machine learning in DBs in the face of data insertions? Is the problem of (and solutions for) machine unlearning for DBs different from unlearning in the ML literature? what are the overhead and efficiency of unlearning algorithms (versus the naive solution of retraining from scratch)? What is the sensitivity of unlearning on batching delete operations (in order to reduce model updating overheads)? If we have a suitable unlearning algorithm (forgetting old knowledge), can we combine it with an algorithm handling data insertions (new knowledge) en route to solving the general adaptability/updatability requirement in learned DBs in the face of both data inserts and deletes? We answer these questions using a comprehensive set of experiments, various unlearning algorithms, a variety of downstream DB tasks (such as SE, AQP, and DG), and an upstream task (DC), each with different NNs, and using a variety of metrics (model-internal, and downstream-task specific) on a variety of real datasets, making this also a first key step towards a benchmark for learned DB unlearning.

HTML5 Bootstrap Template by colorlib.com

Detect, Distill and Update: Learned DB Systems Facing Out of Distribution Data

SIGMOD 2023 | M Kurmanji*, P Triantafillou

Machine Learning (ML) is changing DBs as many DB components are being replaced by ML models. One open problem in this setting is how to update such ML models in the presence of data updates. We start this investigation focusing on data insertions (dominating updates in analytical DBs). We study how to update neural network (NN) models when new data follows a different distribution (a.k.a. it is "out-of-distribution" -- OOD), rendering previously-trained NNs inaccurate. A requirement in our problem setting is that learned DB components should ensure high accuracy for tasks on old and new data (e.g., for approximate query processing (AQP), cardinality estimation (CE), synthetic data generation (DG), etc.). This paper proposes a novel updatability framework (DDUp). DDUp can provide updatability for different learned DB system components, even based on different NNs, without the high costs to retrain the NNs from scratch. DDUp entails two components: First, a novel, efficient, and principled statistical-testing approach to detect OOD data. Second, a novel model updating approach, grounded on the principles of transfer learning with knowledge distillation, to update learned models efficiently, while still ensuring high accuracy. We develop and showcase DDUp's applicability for three different learned DB components, AQP, CE, and DG, each employing a different type of NN. Detailed experimental evaluation using real and benchmark datasets for AQP, CE, and DG detail DDUp's performance advantages.

HTML5 Bootstrap Template by colorlib.com

PGMJoins: Random Join Sampling with Graphical Models

SIGMOD 2021 | AM Shanghooshabad*, M Kurmanji, Q Ma, M Shekelyan, M Almasi and P Triantafillou

Modern databases face formidable challenges when called to join (several) massive tables. Joins (especially when entailing many-to-many joins) are very time- and resource-consuming, join results can be too big to keep in memory, and performing analytics/learning tasks over them costs dearly in terms of time, resources, and money (in the cloud). Moreover, although random sampling is a promising idea to mitigate the above problems, the current state of the art leaves lots of room for improvements. With this paper we contribute a principled solution, coined PGMJoins. PGMJoins adapts Probabilistic Graphical Models to deriving provably random samples of the join result for (n-way) key joins, many-to-many joins, and cyclic and acyclic joins. PGMJoins contributes optimizations both for deriving the structure of the graph and for PGM inference. It also contributes a novel Sum-Product Message Passing Algorithm (SP-MPA) to make a uniform sample of the joint distribution (join result) efficiently and a novel way to deal with cyclic joins. Despite the use of PGMs, the learned joint distribution is not approximated, and the uniform samples are drawn from the true distribution. Our experimentation using queries and datasets from TPC-H, JOB, TPC-DS, and Twitter shows PGMJoins to outperform the state of the art (by 2X-28X).

HTML5 Bootstrap Template by colorlib.com

Learned approximate query processing: Make it light, accurate and fast

CIDR 2021 | Q Ma*, AM Shanghooshabad, M Almasi, M Kurmanji, P Triantafillou

The advent of learning algorithms has revealed many opportunities for improving Data Systems’ functionality and performance. Approximate Query Processing (AQP) is one such area where machine learning (ML) models have been used to improve query execution efficiency and accuracy, outperforming the traditional samplingbased approaches. Based on our group’s experience in the ML-for-DBs area,[3–7, 29, 37–39], we contribute a novel AQP engine, coined DBEst++, which extends our previous effort (DBEst,[29]) and sets the state of the art in terms of accuracy and query execution efficiency. The DBEst++ salient design objective is to derive lightweight ML models for the task, allowing a plethora of ML models to coexist, covering a very large fraction of the expected analytical query workload without requiring very large memory footprints. The DBEst++ salient architectural feature rests on a novel blending of word embedding models with neural networks tasked with regression-based predictions for density estimation and aggregation-attribute values. We present design features and motivations/rationale behind DBEst++ and discuss how all the ML models are brought together. We also present how DBEst++ can deal with challenging scenarios, including how to deal with highcardinality categorical attributes and how to ensure high accuracy under data updates. We provide a detailed experimental evaluation using the TPC-DS and Flights datasets against state of the art learned and sampling-based AQP engines, showcasing DBEst++’s gains in terms of accuracy, response-times, and memory space overheads.

HTML5 Bootstrap Template by colorlib.com

A comparison of 2D and 3D convolutional neural networks for hand gesture recognition from RGB-D data

ICEE 2019 | M Kurmanji*, F Ghaderi

Hand gesture recognition from videos has more challenges compared with still images due to the difficulty of representing temporal features and longer training times especially in real-time applications. In this paper, we investigated the potential of 2D over 3D CNN's for representing temporal features and classification of hand gestures in videos. We mapped the frame sequence of hand gestures to a chronological tiled pattern in order to capture the dynamics of the hand movement in a single frame. Then, using 2D CNN's, we generated feature vectors containing both special and temporal features. Additionally, we proposed a new approach for fusing data and predictions through a two-stream architecture to exploit depth information. The effects of different types of augmentation techniques is also investigated. Our results confirm that appropriate usage of 2D CNNs outperforms a 3D CNN implementation in terms of recall, accuracy and time in this task

Education

Uni: University of Warwick

Thesis: Adaptibility of the Machine Learning Based Data Systems

Uni: Tarbiat Modares University

Thesis: Hand Gesture Recognition Using Deep Convolutional Neural Networks

Uni: Isfahan University of Technology

Thesis: FPGA implementation of CDMA signal modulation

Experience

Teaching Assistant

University of Warwick 2020-present

Teaching Assistant for Machine Learning, Relational Databases, and Advanced Databases courses. Proficient in guiding students through complex AI and data science concepts while providing collaborative seminar and lab environments. Facilitated discussions and problem-solving sessions, fostering practical application of AI/ML techniques.

Big Data Engineer

at ICT Research Institute2019-2020

Collaborated within a team to orchestrate end-to-end data workflows, transforming diverse data sources into Hadoop clusters. Evaluated migration strategies and designed Unix/Linux ETL pipelines for processing Big Data. Enhanced data warehousing and business intelligence solutions.

Data and Machine Learning Engineer

Refah Retail Chain Stores Co.2017-2019

Pioneered the creation of an AI-powered retail suite at Refah, implementing cutting-edge techniques such as customer counting, crowded zone detection, and customer behavior analysis. Leveraged machine learning to categorize products and classify social media comments, providing comprehensive insights for store operations and customer interactions

Deep Learning Engineer

Sensifai 2017-2018

Contributed to the development of a sophisticated audio analysis system, utilizing transfer learning for acoustic scene detection, CNN-based mood classification, and weakly-labeled human speech recognition. Engineered a parallelized YouTube dataset crawler and preprocessing pipeline