Postdoctoral Fellow
Department of Computer Science, Toronto Metropolitan University, Toronto
mkn [at] torontomu [dot] ca | miladkhademinori [at] gmail [dot] com
I build the mathematics of continual learning: task-confusion bounds that explain when forgetting must occur, the geometry of pre-trained representations that explains when it need not, and memory policies for continually learning agents.
About
I am a Postdoctoral Fellow in the Department of Computer Science at Toronto Metropolitan University, working with Prof. Guanghui (Richard) Wang on continual and federated learning with pre-trained models. I earned my PhD in Electrical and Computer Engineering from Queen's University in 2024, supervised by Prof. Il-Min Kim, with a thesis on federated and incremental learning, completed within the Huawei-funded doctoral research program.
My research gives continual learning a mathematical foundation. The starting point is a simple observation: models that keep learning after deployment forget, and the field lacked a theory of when forgetting is unavoidable. My task-confusion framework proved that discriminative learners trained in stages are infeasible in general; the margin and gauge results that followed show what property of a pre-trained representation lifts that infeasibility, and by how much, in closed form; and my current work turns the same quantities into provable memory policies for agents that must decide what to keep, what to consolidate into weights, and what to forget.
Every result in this line was produced on a single workstation GPU: the program is deliberately reproducible by any group, without cluster access.
Research
Task-confusion bounds, infeasibility theorems, and exponential-decay margins: when forgetting must happen, when it need not, and the quantities that decide.
Why frozen and lightly adapted backbones evade the classical limits: margins, head gauge, and certificates computable before training.
Learning across clients without shared data: communication trade-offs, non-IID heterogeneity, and federated incremental learning.
What an agent should write to memory, when to consolidate into weights, and what to forget: rent-or-buy ledgers and policies with guarantees.
News
Paper accepted at ICML 2025 (Vancouver): autoencoder-based hybrid replay for class-incremental learning; presented in person.
Travel award from the Faculty of Science, Toronto Metropolitan University, to present at ICML 2025.
Paper accepted at ICLR 2025 (Singapore): federated class-incremental learning with latent exemplars and data-free techniques; presented in person.
Paper accepted at NeurIPS 2024 (Vancouver): the task-confusion framework for class-incremental learning; presented in person.
Completed the PhD at Queen's University (January 2024) and joined Toronto Metropolitan University as a Postdoctoral Fellow (March 2024).
Publications
Autoencoder-based hybrid replay for class-incremental learning
@inproceedings{nori2025autoencoder,
author = {Khademi Nori, Milad and Kim, Il-Min and Wang, Guanghui},
title = {Autoencoder-based hybrid replay for class-incremental learning},
booktitle = {International Conference on Machine Learning (ICML)},
year = {2025}
}Federated class-incremental learning: a hybrid approach using latent exemplars and data-free techniques to address local and global forgetting
@inproceedings{nori2025federated,
author = {Khademi Nori, Milad and Kim, Il-Min and Wang, Guanghui},
title = {Federated class-incremental learning: a hybrid approach using latent exemplars and data-free techniques to address local and global forgetting},
booktitle = {International Conference on Learning Representations (ICLR)},
pages = {84398--84411},
year = {2025}
}Task confusion and catastrophic forgetting in class-incremental learning: a mathematical framework for discriminative and generative modelings
@inproceedings{nori2024task,
author = {Khademi Nori, Milad and Kim, Il-Min},
title = {Task confusion and catastrophic forgetting in class-incremental learning: a mathematical framework for discriminative and generative modelings},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
volume = {37},
pages = {47678--47707},
year = {2024}
}On the effectiveness of activation noise in both training and inference for generative classifiers
Fast federated learning by balancing communication trade-offs
EDMARA2: a hierarchical routing protocol for EH-WSNs
A QGSA cluster head selection approach for hierarchical routing protocol in the EH-WSNs
Education
PhD, Electrical and Computer Engineering
Queen's University, Kingston, Canada
Thesis: Federated Learning and Incremental Learning: Private Non-IID Learning. Supervisor: Prof. Il-Min Kim. Completed as part of the Huawei-funded doctoral research program, with semi-annual reviews at Huawei Canada, Ottawa.
M.Sc., Electrical Engineering, ranked first in the class
Amirkabir University of Technology, Tehran
Thesis: an energy-efficient routing protocol for energy-harvesting wireless sensor networks in the IoT. Supervisor: Prof. Saeed Sharifian.
B.Sc., Electrical Engineering, ranked first in the class
Semnan University, Semnan, Iran
Digital electronic systems. Supervisor: Prof. Naser Eskandarian. Capstone: a Bluetooth-controlled quadcopter.
Experience
Postdoctoral Fellow
Department of Computer Science, Toronto Metropolitan University
Continual and federated learning with pre-trained models, with Prof. Guanghui (Richard) Wang. All published results of the appointment were produced on a single workstation GPU.
Research Assistant
Queen's University, Innovation Park
Exploratory analysis of prostate-cancer datasets with Prof. Sidney Givigi: CT volumes, mass spectrometry, cleaning and preprocessing, volumetric data, landmark registration, four technical reports.
Doctoral Researcher, Huawei-funded program
Queen's University with Huawei Technologies Canada, Ottawa
Non-IID deep learning: federated learning under communication burden and heterogeneity; continual learning, forgetting and task confusion. Semi-annual in-person research presentations.
Instructor, Lab Instructor, Teaching Assistant
Department of Electrical and Computer Engineering, Queen's University
MATLAB programming and mathematics (about 110 students per week) and Python for discrete-time signal processing (about 90 students); laboratories and mentoring.
Vice-President, speaker, public-relations chair
Ingenuity Lab, Queen's University
Public talks on the impacts of artificial intelligence; undergraduate mentoring.
Network Specialist
Huawei Technologies, Tehran
Network equipment; taught internal networking classes; Huawei Certified Network Associate.
Awards
Teaching
Service
Contact
Department of Computer Science, Toronto Metropolitan University, Toronto, Ontario, Canada.
Email: mkn [at] torontomu [dot] ca | miladkhademinori [at] gmail [dot] com