My Papers.
Performance Profiling of Federated Learning Across Heterogeneous Mobile Devices
As mobile devices increasingly become more widely used in data analytics and machine learning, Federated Learning (FL) offers a promising decentralised approach that maintains data privacy and reduces data transmission costs. This work analyses FL performance across heterogeneous mobile devices in real deployment. We examine the impacts of device heterogeneity on the training efficiency of FL systems by conducting a series of experiments involving real and emulated smartphones with various computational capabilities and network conditions. Through performance profiling, we identify bottlenecks in mobile settings to overcome the unique challenges mobile environments pose and design an FL system that effectively uses heterogeneous devices. Our experiments make use of a federated CIFAR10 dataset. Our results indicate higher setup times on physical devices than emulated ones, regardless of the network connection type. While data loading from storage is the primary bottleneck, consuming up to 88\% of setup time, with a Broadband connection, downloading the model from the server becomes another major bottleneck with a 4G/LTE connection. Our findings also reveal that the fit function of the training phase takes up to 94\% of training time. These insights can hopefully aid in designing FL systems that adapt effectively to heterogeneous resources in FL environments.
Technologies Used
Python FedKitCitation
'Performance Profiling of Federated Learning Across Heterogeneous Mobile Devices', Esquivel, J. A., Aldous, B., Abdelmoniem, A. M., Alhilal, A. and You, L. (2024)
Comparative Profiling: Insights Into Latent Diffusion Model Training
Generative AI models are at the forefront of advancing creative and analytical tasks, pushing the boundaries of what machines can generate and comprehend. Among these, latent diffusion models represent significant advancements in generating high-fidelity audio and images. This study introduces a systematic approach to study GPU utilisation during the training of these models by leveraging Weights & Biases and the PyTorch Profiler for detailed monitoring and profiling. Our methodology is designed to uncover inefficiencies in GPU resource allocation, pinpointing bottlenecks in the training pipeline. The insights gained aim to guide the development of strategies for enhancing training efficiency, potentially reducing computational costs and accelerating the development cycle of generative AI models. This contribution not only highlights the critical role of resource optimisation in scaling AI technologies but also opens new avenues for research in efficient model training.
Technologies Used
Python PyTorch wandbCitation
'Comparative Profiling: Insights Into Latent Diffusion Model Training', Aldous, B. and Abdelmoniem, A. M. (2024), In Proceedings of the 4th Workshop on Machine Learning and Systems (EuroMLSys '24). Association for Computing Machinery, New York, NY, USA, 176–183. DOI: 10.1145/3642970.3655847
Percussive audio mixing with Wave-U-Nets
A crucial step in the music production process is creating a cohesive combination mixture from the separate recorded components of a piece of music, aided by a variety of audio effects and techniques. Audio mixing is a far more creative and varied task than its subsequential post-production process: mastering, for which there are numerous AI-driven applications on the market. This research takes a look at a promising deep learning method for automatic mixing of drums in dry and wet scenarios; the Wave-U-Net. Modifications are made to the convolutional filters of the current Wave-U-Net architecture to increase the number of trainable parameters of the model, with the idea that this will aid in better understanding patterns in the audio.
Technologies Used
Python PyTorchCitation
'Percussive audio mixing with Wave-U-Nets', Aldous, B. (2022), MSc Thesis, DOI: 10.13140/RG.2.2.30052.04480/1
Data Privacy: Negligence at Marriott International
The 2014 Marriott International data breach is discussed, relating some of the key issues of the case with some important legal frameworks around data privacy, like the UK General Data Protection Regulation and Data Protection Act 2018, the UK Data Ethics Framework, and the ICO Data Sharing Code. For their failure in keeping customer's personal data secure, Marriott International were obligated to pay a fine of £18.4 million by the ICO. This study aims to look at the principles and articles of the above frameworks that Marriott International breached or failed to uphold, and then compare this to the extent of the penalty received to make a conclusion as to whether it is just. Following this, recommended adjustments to policies around data privacy are discussed, so as to prevent similar occurrences from happening in the future.
Citation
'Data Privacy: Negligence at Marriott International', Aldous, B. (2022), Critical Review, DOI: 10.13140/RG.2.2.19595.46881
The early and late-time accelerated expansion of the universe
For decades, since its discovery in the early 20th century, the idea of an expanding universe has been an incredibly active area of research among astronomers and cosmologists. This discovery means that our fundamental idea of how gravity acts fails over cosmological distances. It was previously thought that the universe was static, and so Einstein introduced the cosmological constant term (usually denoted by Λ) into his field equations for general relativity in order to counteract the effect of gravity. Einstein later abandoned this idea after Edwin Hubble's discovery of cosmic expansion in 1929 and so the value of Λ was considered to be zero by the majority of cosmologists until the 1990s. Through observations of distant type Ia supernovae, the accelerated expansion of the universe was discovered in 1998 and has led theorists to search for concepts that explain this phenomenon.
Technologies Used
PythonCitation
'The early and late-time accelerated expansion of the universe', Aldous, B. (2018), MSci Thesis, DOI: 10.13140/RG.2.2.26306.35524
Primordial black holes as possible candidates for dark matter
The capture of dark matter by pre-stellar cores is considered, subsequently the dark matter will be trapped inside the compact remnant that the star becomes. If the dark matter is made up of primordial black holes (PBHs) then these will rapidly destroy the compact remnant and so constraints on the abundance of PBHs are implied by observations of compact remnants. Observational constraints based on black hole evaporation and gravitational lensing, as well as various dynamical constraints, are all considered and applied to the three allowed PBH mass ranges in which they could comprise the dark matter: 1016g - 1017g, 1020g - 1024g and 1M⊙ - 103M⊙. These mass ranges are confronted with monochromatic and extended mass functions.
Citation
'Primordial black holes as possible candidates for dark matter', Aldous, B. (2017), BSc Thesis, DOI: 10.13140/RG.2.2.12884.58244