Valentin Radu

Valentin Radu

Auto-generated profile, based on publicly available information

Valentin Radu

Devotion AI Labs, London, United Kingdom

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
8 May 2023
Verified
Conference Paper
Open access
HSHongrui Shi
Hongrui Shi
VRValentin Radu
Valentin Radu
PYPo Yang
Po Yang
Data privacy and data protection are crucial issues for automatic speech recognition (ASR) system when relying on client generated data for training. The best protection is achieved when training is distributed fashion, close to the client local data, rather than centralising the training. However, distributed training suffers from system heterogeneity, due to clients having unequal computation resources, and data heterogeneity, due to training data being non-independent and identically distributed (non-IID). To tackle these challenges, we introduce FedKAD, a Federated Learning (FL) framework that uses local Knowledge Aggregation over top level feature maps and Knowledge Distillation. We show that our FedKAD achieves better communication efficiency than standard FL methods that use uniform models, due to transferring parameters of smaller size client models, and overall better accuracy than FedMD, an alternative KD-based approach designed for heterogeneous data. Our work enables faster, cheaper and more inclusive participation of clients in heterogeneous distributed training.

20 June 2026
Verified
Conference Paper
Open access
VRValentin Radu
Valentin Radu
MPMaryam Vahdat Pour
Maryam Vahdat Pour
MPMassimo Pegoraro
Massimo Pegoraro
MFMyles Foley
Myles Foley
FFFrank Faricy
Frank Faricy
MGMark Girolami
Mark Girolami

20 June 2026
Verified
Conference Paper
Open access
VRValentin Radu
Valentin Radu
MFMyles Foley
Myles Foley
MPMaryam Vahdat Pour
Maryam Vahdat Pour
MPMassimo Pegoraro
Massimo Pegoraro
FFFrank Faricy
Frank Faricy
MGMark Girolami
Mark Girolami

4 November 2025
Verified
Conference Paper
Open access
HLHaoxuan Li
Haoxuan Li
XWXijia Wei
Xijia Wei
ZWZhiqiang Wei
Zhiqiang Wei
VRValentin Radu
Valentin Radu
CWCheng Wang
Cheng Wang
Wenda LiWenda Li
Wenda Li

5 December 2023
Verified
Conference Paper
Open access
HSHongrui Shi
Hongrui Shi
VRValentin Radu
Valentin Radu
PYPo Yang
Po Yang
With a growing variety of devices capable of generating data and due to their data holding diverse characteristics, hardware heterogeneity and data heterogeneity have become pressing issues in Federated Learning (FL). Many studies aim to address the model shifts either by partitioning the local model training or by retraining classifiers on the server. However, these approaches fail to identify the actual components of the model that are susceptible to model shifts, or may compromise user privacy by uploading local client features to the server. On another track of research, few-shot learning studies show that retraining only the classifier part of a model achieves results just shy of state-of-the-art in metric and meta learning. Bringing the advancements achieved in few-shot learning to the scope of federated learning, we introduce FedFSC, a novel approach that leverages few-shot learning to address both the stragglers and model shift issues. FedFSC is a training approach that extends established FL algorithms, by restricting the training on clients with lower computational resources to only the classifier part of the model, and full model update on the rest of clients. We evaluate FedFSC on three classification tasks in heterogeneous FL settings, and demonstrate its ability to improve the performance of a few popular FL algorithms substantially, while reducing the workload of smaller devices by one order of magnitude.

13 November 2023
Verified
Conference Paper
Open access
ATAnselmo Talotta
Anselmo Talotta
VRValentin Radu
Valentin Radu
LSLorenzo Sorgi
Lorenzo Sorgi
Floorplans are useful for navigating indoor spaces, for resource allocation and for indoor space management among many others. But in the absence of readily available digital floorplans, these are hard to generate. In this work, we enhance the generation of floorplans from point clouds to be more robust to noisy measurements of sensor data. In our approach, we train an object detector to expose room shapes in a density map produced from a 3D point cloud, as well as the position of relevant landmarks, such as doors and windows. We improve the robustness of the room detector by training in two stages, firstly using point clouds extracted from synthetic 3D graphical representations of plausible indoor spaces; and secondly, extending the training of the model in a new domain by using real-world data collected with Tango devices. This two-tier training nudges the model closer to our target domain, of generating floorplans from easily collected point cloud scans in the real-world. Finally, we showcase the capability of our solution when operating with noisy Lidar scans collected from a drone with pose estimation.