This paper introduces FEDDE, a general and efficient framework that addresses data redundancy across clients to facilitate effective federated learning (FL). At its core, FEDDE adopts a hierarchical deduplication architecture where clients first perform local, centralized deduplication and then send minimal records that are only meaningful for redundancy detection to the server for global deduplication. To enable flexible trade-offs between FL training efficiency and the accuracy of the training outcomes, FEDDE proposes two-round approximate deduplication protocols. A set of system optimizations is further applied to reduce deduplication overhead.

Federated learning (FL) has emerged as a popular paradigm for distributed machine learning over decentralized data. Data generated by FL clients is prone to noises. While the impact of data noise on centralized learning (CL) is well understood, there is lack of a systematic study for FL. We fill this gap by presenting an empirical investigation to provide a deeper understanding regarding the impact of data noise on FL. Our study is enabled by NoiseMaker, an open-source and extensible toolkit for the injection of controlled data noises across five diverse data modalities. Our experimental evaluation results reveal that FL is significantly more vulnerable to data noise compared to CL.

This paper establishes that General Relativity and Quantum Mechanics are necessary logical consequences of the Axiom of Finite Information. We introduce a new fundamental constant, i, representing the Information Maximum Transfer Speed, and posit that i > c, where c is the speed of light in a vacuum. By substituting i into the relativistic framework, we demonstrate that the finite nature of i is the primary mechanism preventing infinite information density and logical singularities. Furthermore, we prove that a ‘Theory of Everything’ is precluded by the computational cost of self-reference, and propose the observation of Computational Redshift as a definitive empirical test for the gap between c and i.

We propose a conceptual framework to resolve the dichotomy of the Millennium Prize Problems by categorizing mathematical systems based on their capacity for logical simulation. We distinguish between Class I (Structural) problems (e.g., Poincaré, Hodge, Yang-Mills), which rely on symmetries, conservation laws, and coercivity estimates that constrain degrees of freedom effectively, and Class II (Simulational) problems (e.g., P vs NP, Navier-Stokes), which theoretically possess the fidelity to simulate Universal Turing Machines. While not a formal proof of independence, we argue that Class II problems face obstructions isomorphic to the Halting Problem, inhibiting standard analytic techniques. We posit that the ‘intractability’ of these problems arises because they inhabit a complexity class where asymptotic behavior is determined by generalized computation rather than geometric structure.