arXiv subjects

Özgür Soysal

Publications and source records attributed to Özgür Soysal.

2 recordsLinked to original sources

arXiv2026

DROD: Dual-Ratio Ordered Dropout for Federated Continual Learning on Heterogeneous Devices

Federated learning for capacity-heterogeneous clients and federated continual learning are each well studied on their own, but their intersection -- clients holding differently sized models while every client's data stream also drifts over time -- remains underexplored. We study this setting directly, training a ResNet-18 fleet from scratch on Split-CIFAR-100 (10 tasks, class-incremental), and introduce DROD (Dual-Ratio Ordered Dropout), which builds on FjORD's Ordered Dropout by using two independent reduction axes to extract sub-models: a shortcut ratio and a branch ratio. On a shared parameter/FLOP budget ladder, DROD achieves the highest mean accuracy among the heterogeneous-model federated learning strategies we compare. Its advantage over the two strongest baseline geometries persists across every replay-buffer size we test and under ER, DER++, and BiC; under EWC, all three geometries collapse alike.

cs.LG↗

arXiv2026

A New Upper Bound on the Binary Deletion Channel Capacity

We prove that the capacity of the binary deletion channel satisfies $C(d)\le (1-d)/4$ for every $13/20\le d<1$. The proof describes the output from right to left, using a six-bit context to assign a description length. We bound the increase in expected description length minus output entropy when one input bit is added. A relative-entropy identity reduces this bound to finitely many linear inequalities. A potential on input windows of length 26 makes the inequalities telescope, giving a bound for every input word. Deletion composition extends the result from $d=13/20$ to all larger deletion probabilities. We also obtain finite-block bounds on mutual information and decoding error, with explicit $O(\log n/n)$ corrections. The finite certificate is checked using exact integer arithmetic, and the proof is formalized end to end in Lean.

cs.IT↗