papers/hanzo-dso/sections/11-related-work.tex

\section{Related Work}

\label{sec:related}

\paragraph{Federated Learning.}
FedAvg~\cite{mcmahan2017communication} aggregates local model updates via
averaging. FedProx~\cite{li2020federated} adds a proximal term for
heterogeneous settings. These methods aggregate gradients/parameters, unlike
DSO which aggregates semantic priors.

\paragraph{Byzantine-Robust Aggregation.}
Krum~\cite{blanchard2017machine} selects the update closest to others.
Bulyan~\cite{mhamdi2018hidden} combines Krum with trimmed mean.
\cite{yin2018byzantine} analyze coordinate-wise median and trimmed mean.
Our weighted median extends these ideas to semantic prior aggregation.

\paragraph{Decentralized Machine Learning.}
Swarm Learning~\cite{warnat2021swarm} coordinates peer-to-peer training
without a central server. Gossip-based SGD~\cite{koloskova2020unified}
distributes gradient computation. DSO differs by operating on compressed
semantic priors rather than model parameters.

\paragraph{Blockchain for ML.}
FLock~\cite{flock2024} uses blockchain for federated learning coordination.
Bittensor~\cite{bittensor2023} creates a decentralized marketplace for ML
models. DSO uses blockchain only for integrity and incentive management,
keeping compute and storage off-chain.

\paragraph{Parameter-Efficient Federated Learning.}
FedLoRA~\cite{zhang2024fedlora} and FedPara~\cite{hyeon2022fedpara}
federate LoRA or low-rank parameter updates. These still require gradient
computation; DSO is entirely gradient-free.

\paragraph{Experience Sharing in Multi-Agent RL.}
\cite{christianos2021scaling} and \cite{baker2020emergent} study experience
sharing in multi-agent reinforcement learning. DSO adapts these ideas to the
LLM setting with compressed semantic priors instead of raw trajectories.

% ============================================================================
% 12. CONCLUSION
% ============================================================================