Intelligent Edge Computing
Permanent URI for this collection
Browse
Recent Submissions
Item MergeKD: An Empirical Framework for Combining Knowledge Distillation with Model Fusion Using BERT Model(2025-01-07) Tran, Ngoc Minh; Le, Bang Giang; Ta, Viet CuongBERT-based models have become the mainstream in sentiment classification approaches. However, due to the divergence of the text domains, each domain requires a specific fine-tuned model which is often impractical for scaling. Additionally, the large model's size requires heavy computation resources. In our work, we propose a framework which could address such issues by combining the domain adaptation task with a lightweight model distillation. From each trained model of a specific domain, a merged model is created by fusing all models without the need to finetune on a combined dataset. Consequentially, the resulting model is distilled into a smaller model to lower the required computation. We test our framework on semantic classification with Vietnamese datasets with a pre-trained BERT-based architecture. The results highlight that our merged model achieves the highest average accuracy overall substantially while the distilled model maintains a competitive performance with a 50\%{} reduction in inference time.Item The MAGPIE: Satellite Autonomy for Uncooperative Environments(2025-01-07) Mehlman, Cameron; Kounios, Asterios; Lai, Laurance; Prasad, Adhyan; Kully, Will; Brown, Denis; Hughes, Ryan; Chalamalasetty, Shashank; Distler, Jonathan; Dilone, Rafael; Palomino, Edward; Lewick, Keegan; Calabrese, Matteo; Boschetti, Nicolò; Goel, Mahika; Phillips, Sean; Falco, GregoryAs the space industry continues to grow, satellites are increasingly encountering non-cooperative environments. Such scenarios require edge-based autonomy to react to adversarial spacecraft in the complex 6 degree of freedom (DOF) environment. We present the MAGPIE (Multi Agent Generative Path planning for Intelligent Evasion), an autonomous system designed to be run on the edge specifically for satellites performing in non-cooperative 6DOF environments. In this paper, we describe the edge-based system architecture which entails a sensing suite, on-board computer, and custom software for planning and data fusion. We also discuss the constraints of satellite systems and how they are accounted for in the design of the architecture. In addition, we provide a framework for implementing the system on a quadcopter as a hardware test-bed, and present our results gathered from initial testing.Item An Efficient Method of Lifelong Learning with Differentiable Memory for Edge Computing(2025-01-07) Moon, Hyung-Jun; Cho, Sung-BaeThe proliferation of deep learning leads to great success in various domains such as computer vision, natural language processing, and even edge computing. However, it often comes at the cost of substantial computational power, memory usage, and extensive data requirements, posing challenges for edge computing with limited resources. This paper presents a novel, memory-efficient incremental learning method optimized for edge computing. By harnessing differentiable memory storage and lifelong learning principles, the proposed method facilitates efficient concurrent learning and storage of knowledge, significantly reducing the need for extensive retraining while preserving privacy and enhancing space efficiency. Experimental results on three benchmark datasets demonstrate that the proposed method yields accuracy gains reaching up to 7%p, approximately 47% improvement in space efficiency and 36% improvement in time efficiency against the SOTA methods. It surpasses conventional methods in accuracy, space efficiency, and learning time, making it a robust solution for continual learning in resource-constrained environments.Item An Approach to Time Series Forecasting With Derivative Spike Encoding and Spiking Neural Networks(2025-01-07) Manna, Davide; Di Caterina, Gaetano; Vicente-Sola, Alex; Kirkland, PaulTimely and energy-efficient time series forecasting can play a key role on edge devices, where power requirements can be stringent. Spiking Neural Networks (SNNs) are regarded as a new avenue in which to solve time series problems, but with lower SWaP (Size, Weight, and Power) needs. We propose an SNN pipeline to process and forecast time series, developing a novel data spike-encoding mechanism and two loss functions that optimise the prediction of the upcoming spikes. Our approach encodes a signal into sequences of spikes that approximate its derivative, preparing the data to be processed by the SNN, while our proposed loss functions account for the reconstruction of the output spikes into a meaningful value to promote convergence to top-level solutions. Results show that our solution can effectively learn from the encoded data and the SNN trained with our loss function can outperform the same model trained with SLAYER's default loss.Item Introduction to the Minitrack on Intelligent Edge Computing(2025-01-07) Juric, Radmila; Ronchieri, Elisabetta; Sanfilippo, Filippo; Bihl, TrevorItem Generative Data for Neuromorphic Computing(2025-01-07) Baietto, Anthony; Bihl, TrevorNeuromorphic computing is a next-generation model of computation that leverages biologically-inspired artificial neurons to perform complex tasks. Unlike neurons found within traditional Artificial Neural Networks (ANNs), which output on a continuous range, neuromorphic neurons found within Spiking Neural Networks (SNNs) output discrete spikes over time, similar to the individual firing of biological neurons. This low Size Weight and Power (SWaP) model of computation, when realized on specialized neuromorphic hardware and coupled with low-power neuromorphic sensors, makes neuromorphic computing an ideal candidate for edge computing applications. However, training neuromorphic models is challenging because of the scarcity of quality neuromorphic datasets. In this paper, we present a platform-agnostic approach for creating synthetic neuromorphic datasets with a Conditional Generative Adversarial Network (CGAN). Neuromorphic models trained on the generated datasets perform comparably to those trained on the original IBM DVSGesture dataset. We show that neuromorphic dataset generation produces quality samples which can further aid the development and deployment of neuromorphic computing models.