arXiv subjects

Jeong Min Kong

Publications and source records attributed to Jeong Min Kong.

3 recordsLinked to original sources

arXiv2026

Plasticity of Growing and Elastic Neural Networks in Online Continual Learning

Neural networks that can grow or both grow and shrink during learning, referred to as growing neural networks and elastic neural networks, respectively, have recently been explored in offline continual learning with a particular focus on catastrophic forgetting. Driven by the observations that 1) online continual learning closely resembles how animals learn; 2) loss of plasticity---the progressive decline in a learning network's ability to learn---is another crucial challenge facing continual learning; and 3) incremental introduction of randomly initialized hidden units was recently shown to help preserve plasticity, in this paper, we study the plasticity of several foundational growing and elastic networks in online continual learning. Our experiments in supervised learning settings show that adaptive growing networks, which incrementally incorporate new, randomly initialized units to the network while keeping all existing connections adaptive, can maintain high prediction accuracy without losing plasticity despite the continuous increase in the dead hidden unit proportion. Furthermore, we demonstrate that adaptive elastic networks, which in addition to progressively adding new hidden units also prune estimated dead hidden units at the beginning of each new task, can achieve excellent accuracy without loss of plasticity while simultaneously maintaining a near-constant, compact size. Our results suggest that growing and elastic networks, which exhibit the ability to adapt its structure to the relevant learning objectives, can be a promising class of algorithms also for preserving high plasticity in online continual learning.

cs.LG↗

arXiv2025

Satellite Assignment Policy Learning for Coexistence in LEO Networks

Unlike in terrestrial cellular networks, certain frequency bands for low-earth orbit (LEO) satellite systems have thus far been allocated on a non-exclusive basis. In this context, systems that launch their satellites earlier (referred to as primary systems) are given spectrum access priority over those that launch later, known as secondary systems. For a secondary system to function, it is expected to either coordinate with primary systems or ensure that it does not cause excessive interference to primary ground users. Reliably meeting this interference constraint requires real-time knowledge of the receive beams of primary users, which in turn depends on the primary satellite-to-primary user associations. However, in practice, primary systems have thus far not publicly disclosed their satellite assignment policies; therefore, it becomes essential for secondary systems to develop methods to infer such policies. Assuming there is limited historical data indicating which primary satellites have served which primary users, we propose an end-to-end graph structure learning-based algorithm for learning highest elevation primary satellite assignment policies, that, upon deployment, can directly map the primary satellite coordinates into assignment decisions for the primary users. Simulation results show that our method can outperform the best baseline, achieving approximately a 15% improvement in prediction accuracy.

eess.SP↗

arXiv2024

Active Beam Learning for Full-Duplex Wireless Systems

In this paper, we present a novel active beam learning method for in-band full-duplex wireless systems, that aims to design transmit and receive beams which suppress self-interference and maximize the sum spectral efficiency. Rather than rely on explicit estimation of the downlink, uplink, and/or self-interference channels like in most existing work, our method instead actively probes all three channels through measurements of SNR and INR over a fixed number of time slots. Then, once this probing concludes, all collected probing measurements are used to design transmit and receive beams which serve downlink and uplink in a full-duplex fashion. We realize this active beam learning scheme through a network of LSTMs and DNNs, which learns to design each probing beam pair and subsequently extract and record valuable information from each probing measurement such that near-optimal serving beams can be designed following the probing stage. Simulation indicates that our method reliably suppresses self-interference while delivering near-maximal SNR on the downlink and uplink with merely 3-10 probing time slots, while exhibiting robustness to measurement noise and the structure of the self-interference channel.

eess.SP↗