publications

Conference and workshop papers

2026

1. Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models

   Amir Rezaei Balef, Mykhailo Koshil, and Katharina Eggensperger

   In International Conference on Machine Learning (ICML), 2026
2. Do Tabular Foundation Models Learn Rules or Memorize Exemplars?

   Amir Rezaei Balef, Mykhailo Koshil, Behzad Nourani-Koliji, and 1 more author

   In 2nd ICML Workshop on Foundation Models for Structured Data, 2026

2025

1. Put CASH on Bandits: A Max K-Armed Problem for Automated Machine Learning

   Amir Rezaei Balef, Claire Vernade, and Katharina Eggensperger

   In Advances in Neural Information Processing Systems (NeurIPS), 2025
2. In-Context Decision Making for Optimizing Complex AutoML Pipelines

   Amir Rezaei Balef, and Katharina Eggensperger

   In European Conference on Artificial Intelligence (ECAI), 2025
3. Posterior Sampling using Prior-Data Fitted Networks for Optimizing Complex AutoML Pipelines

   Amir Rezaei Balef, and Katharina Eggensperger

   In European Workshop on Reinforcement Learning (EWRL), 2025
4. Towards Understanding Layer Contributions in Tabular In-Context Learning Models

   Amir Rezaei Balef, Mykhailo Koshil, and Katharina Eggensperger

   In Workshop on AI for Tabular Data at EurIPS, 2025

2024

1. Towards Bandit-based Optimization for Automated Machine Learning

   Amir Rezaei Balef, Claire Vernade, and Katharina Eggensperger

   In Workshop on Practical ML for Low-Resource Settings at ICLR, 2024

   HTML

2023

1. Adaptive Energy-Efficient Waveform Design For Joint Communication and Sensing using Multiobjective Multiarmed Bandits

   Amir Rezaei Balef, Setareh Maghsudi, and Slawomir Stanczak

   In WSA & SCC 2023; 26th International ITG Workshop on Smart Antennas and 13th Conference on Systems, Communications, and Coding, Feb 2023

   Abs HTML

   We study the transmission power adaptation problem for waveform design in a JCAS system. The challenge arises due to the information shortage about crucial factors such as channel quality and their time-varying nature. First, we formalize two performance metrics of such a system, namely, the probability of detection (PD) and the data information rate (DIR) of fading channels. We then optimize the transmission energy efficiency given constraints on the minimum values for the metrics mentioned above. Under uncertainty about the channel quality and dynamicity in a network, the problem is solvable as a piecewise-stationary multi-objective multi-armed bandit. We develop a sliding window Pareto generic upper confidence bound (UCB)-based algorithm to solve this problem. We establish an upper bound for the expected regret of our proposed algorithm. Numerical experiments validate the effectiveness of the designed waveforms and show that our policy outperforms state-of-the-art methods.

2. Piecewise-Stationary Combinatorial Semi-Bandit with Causally Related Rewards

   Behzad Nourani-Koliji, Steven Bilaj, Amir Rezaei Balef, and 1 more author

   In European Conference on Artificial Intelligence (ECAI), Feb 2023

   HTML

Journal articles

2024

1. AMLTK: A Modular AutoML Toolkit in Python

   Edward Bergman, Matthias Feurer, Aron Bahram, and 6 more authors

   Journal of Open Source Software, Feb 2024

   HTML

2023

1. Piecewise-Stationary Multi-Objective Multi-Armed Bandit With Application to Joint Communications and Sensing

   Amir Rezaei Balef, and Setareh Maghsudi

   IEEE Wireless Communications Letters, May 2023

   Abs HTML

   We study a multi-objective multi-armed bandit problem in a dynamic environment. The problem portrays a decision-maker that sequentially selects an arm from a given set. If selected, each action produces a reward vector, where every element follows a piecewise-stationary Bernoulli distribution. The agent aims at choosing an arm among the Pareto optimal set of arms to minimize its regret. We propose a Pareto generic upper confidence bound (UCB)-based algorithm with change detection to solve this problem. By developing the essential inequalities for multi-dimensional spaces, we establish that our proposal guarantees a regret bound in the order of γ_T\log (T/γ_T) when the number of breakpoints γ_T is known. Without this assumption, the regret bound of our algorithm is γ_T\log (T) . Finally, we formulate an energy-efficient waveform design problem in an integrated communication and sensing system as a toy example. Numerical experiments on the toy example and synthetic and real-world datasets demonstrate the efficiency of our policy compared to the current methods.

2022

1. DONE: Distributed Approximate Newton-type Method for Federated Edge Learning

   Canh T. Dinh, Nguyen H. Tran, Tuan Dung Nguyen, and 4 more authors

   IEEE Transactions on Parallel and Distributed Systems, Nov 2022

   Abs HTML

   There is growing interest in applying distributed machine learning to edge computing, forming federated edge learning. Federated edge learning faces non-i.i.d. and heterogeneous data, and the communication between edge workers, possibly through distant locations and with unstable wireless networks, is more costly than their local computational overhead. In this work, we propose \sf DONEDONE, a distributed approximate Newton-type algorithm with fast convergence rate for communication-efficient federated edge learning. First, with strongly convex and smooth loss functions, \sf DONEDONE approximates the Newton direction in a distributed manner using the classical Richardson iteration on each edge worker. Second, we prove that \sf DONEDONE has linear-quadratic convergence and analyze its communication complexities. Finally, the experimental results with non-i.i.d. and heterogeneous data show that \sf DONEDONE attains a comparable performance to Newton’s method. Notably, \sf DONEDONE requires fewer communication iterations compared to distributed gradient descent and outperforms DANE, FEDL, and GIANT, state-of-the-art approaches, in the case of non-quadratic loss functions.

2021

1. Accelerating Federated Edge Learning

   Tuan Dung Nguyen, Amir R. Balef, Canh T. Dinh, and 4 more authors

   IEEE Communications Letters, Oct 2021

   Abs HTML

   Transferring large models in federated learning (FL) networks is often hindered by clients’ limited bandwidth. We propose \textsf FedAA , an FL algorithm which achieves fast convergence by exploiting the regularized Anderson acceleration (AA) on the global level. First, we demonstrate that FL can benefit from acceleration methods in numerical analysis. Second, \textsf FedAA improves the convergence rate for quadratic losses and improves the empirical performance for smooth and strongly convex objectives, compared to FedAvg, an FL algorithm using gradient descent (GD) local updates. Experimental results demonstrate that employing AA can significantly improve the performance of FedAvg, even when the objective is non-convex.