FedUHD: Unsupervised Federated Learning using In-Memory Hyperdimensional Computing

You Hak Lee; Keming Fan; Xiaofan Yu; Quanling Zhao; Tianqi Zhang; Flavio Ponzina; Tajana Rosing

cs.LGAug 16, 2025

FedUHD: Unsupervised Federated Learning using In-Memory Hyperdimensional Computing

Authors: You Hak Lee, Keming Fan, Xiaofan Yu, Quanling Zhao, Tianqi Zhang, Flavio Ponzina, Tajana Rosing

Organizations: University of California San Diego, USA

Abstract

Unsupervised federated learning (UFL) enables privacy-preserving distributed training without data labeling, yet practical deployment remains challenging due to non-IID data, high computational and communication costs at edge devices, and sensitivity to communication noise. We propose FedUHD, the first UFL framework based on Hyperdimensional Computing (HDC). On the client side, FedUHD employs kNN-based cluster hypervector removal to mitigate non-IID effects by filtering detrimental local outliers. On the server side, cluster-aware HDC aggregation leverages cluster-level statistics to stabilize learning across heterogeneous clients. To further improve efficiency, we design a compute-in-memory (CIM) accelerator based on a novel phase-change memory (PCM) device, integrated with lightweight ASIC digital modules to execute the client-side HDC pipeline within the accelerator. The intrinsic robustness of HDC to low precision and device variations enables efficient mapping onto analog PCM crossbars, exploiting massive parallelism while minimizing data movement. Experimental results show that FedUHD achieves comparable accuracy to state-of-the-art neural network-based UFL methods across all datasets. On HAR and CIFAR10/100, FedUHD delivers an average 2,239x speedup and 1,542x higher energy efficiency on GPU. In addition, FedUHD reduces communication cost by up to 176x on HAR and CIFAR10/100 and demonstrates greater robustness than Orchestra under communication noise. Compared to GPU implementation of FedUHD, the proposed PCM-based accelerator provides an additional 4.07x speedup and three orders of magnitude higher energy efficiency on average. Furthermore, the results demonstrate the benefit of PCM over RRAM as a CIM substrate.

Figures & tables

Fig. 1 : Visualization of weight divergence induced by non-IID data [ 1 ] . After

T

rounds, non-IID data in federated learning causes substantial divergence from centralized learning (

W_{T}^{(f)}

vs.

W_{T}^{(c)}

Fig. 2 : Overview of the HDC training and inference flow.

Fig. 3 : Superlattice PCM device [ 17 ] and PCM-based CIM block.

Fig. 5 : Visualization of the kNN-based cluster HV removal process.

Fig. 6 : FedUHD accelerator architecture overview. The left panel illustrates a CIM block MAC operation: data HV segments are stored in the 2T2R PCM array, and 3-bit input elements are applied through the source line (SL) as

V_{\mathrm{ref}}\pm V_{\mathrm{pulse}}

, producing bitline currents that are digitized by row-multiplexed ADCs.

Fig. 7 : Block diagram of the CIM block and digital ASIC modules. DRU and NIF implement kNN-based cluster HV removal, while CAU and CUE support k-means local clustering.

Fig. 8 : Timing diagram of pipelined CIM and digital ASIC operations. Top: kNN-based cluster HV removal. Bottom: k-means local clustering.

FedUHD accelerator
Total area	35.12 $\mathrm{mm}^{2}$
Array size	128 $\times$ 128 3-bit PCM cells
CIM blocks	1280 (16 $\times$ 80) arrays
Digital ASIC modules	S&A, CAU, DRU, NIF, CUE
Computation Latency and Energy
tREAD=20ns, tWRITE=960ns, tMAC=320ns

TABLE II : System specifications of the FedUHD accelerator. READ/WRITE latency and energy are reported per selected row, while MAC latency and energy are reported per array operation.

Fig. 9 : Comparison of convergence with NN-based UFL baselines. The x-axis denotes communication round and the y-axis denotes clustering accuracy (ACC). Dashed horizontal lines indicate empirical centralized full-data references obtained using the same

J

D

, and 3-bit centroid quantization as FedUHD.

Fig. 10 : Per-client performance comparison between FedUHD and Orchestra [ 4 ] on GPU and CIM. Speedup and energy efficiency are shown on a log scale.

Fig. 11 : Comparison of robustness with NN-based UFL [ 4 ] on the CIFAR10 dataset with 10 clients.

Fig. 12 : Final ACC under different neighborhood sizes $k_{n}$ and cluster counts $J$ on CIFAR10 with 10 clients.

Fig. 13 : FedUHD algorithm analysis: (a) effect of quantization (b) relationship between HD dimension and ACC

Fig. 14 : FedUHD ablation study: (a) effect of the kNN-based cluster HV removal (b) effect of the cluster-aware HDC aggregation

Fig. 15 : Per-client comparison of FedUHD on PCM-based CIM relative to GPU and RRAM-based CIM. Energy efficiency is shown on a log scale.

Fig. 16 : FedUHD accelerator analysis: (a) area breakdown (b) energy breakdown.

Explore similar work

Fed-BAC: Federated Bandit-Guided Additive Clustering in Hierarchical Federated Learning
May 12, 2026Satwat Bashir, Tasos Dagiuklas, Muddesar IqbalHeterogeneous Federated Learning Federated Learning
Computer Science and Informatics · London South Bank University · London, UK +1
Accurate and Resource-Efficient Federated Continual Learning
Jun 9, 2026Jebacyril Arockiaraj, Dhruv Parikh, Jayashree Adivarahan +2Federated Learning Replay-Based Continual Learning
University of Southern California · Los Angeles, USA · DEVCOM Army Research Office +1
FEAST: Federated Shared-Space Training for Resource-Heterogeneous Clients
Aug 10, 2026Bostan Khan, Masoud DaneshtalabBalanced Learning Taco
Mälardalen University, Västerås, Sweden

May 12, 2026, cs.LGJ/K move · Enter open · S save

Fed-BAC: Federated Bandit-Guided Additive Clustering in Hierarchical Federated Learning

Satwat Bashir, Tasos Dagiuklas, Muddesar Iqbal

Computer Science and Informatics · London South Bank University · London, UK+1

Heterogeneous Federated Learning Federated Learning Fedavg Clustering Balanced Learning Edge Devices+1

Open paper AskPDF arXiv

Hierarchical federated learning (HFL) leverages edge servers for partial aggregation in edge computing. Yet existing FL methods lack mechanisms for jointly optimizing cluster assignment and client selection under data heterogeneity. This paper proposes Fed-BAC, which integrates additive cluster personalization with a two-level bandit framework: contextual bandits at the cloud learn server-to-cluster assignments, while Thompson Sampling at each edge server identifies high-contributing clients. The additive decomposition enables the sharing of knowledge between groups through a globally aggregated network, while cluster-specific networks capture distribution variations. Across three classification benchmarks (CIFAR-10, SVHN, Fashion-MNIST) under moderate (

α= 0.5

) and severe (

α= 0.1

) Dirichlet non-IID partitioning, Fed-BAC achieves distributed accuracy gains of up to +35.5pp over HierFAVG and +8.4pp over IFCA, while requiring only 80% client participation, converging 1.5 to 4.8

\times

faster depending on dataset and accuracy target, and improving cross-server fairness. These gains are further validated at 5

\times

deployment scale on CIFAR-10. The advantage of Fed-BAC increases with heterogeneity severity, confirming that additive cluster personalization becomes increasingly valuable as data distributions diverge.

Abstract

Explore similar work

Fed-BAC: Federated Bandit-Guided Additive Clustering in Hierarchical Federated Learning

Accurate and Resource-Efficient Federated Continual Learning

FEAST: Federated Shared-Space Training for Resource-Heterogeneous Clients