University of Twente Student Theses
Uncovering the Accuracy-Efficiency Trade-Off in Sparse Neural Network Training
Suidgeest, W.F. (2024) Uncovering the Accuracy-Efficiency Trade-Off in Sparse Neural Network Training.
PDF
3MB |
Abstract: | Recently introduced sparse neural network training methods have been shown to match or even surpass the performance of comparable dense neural networks while increasing their computational efficiency. However, these experiments have been performed in controlled environments with fixed network architectures and hyperparameter configurations, potentially causing skewed results. We conducted experiments on six datasets, using multi-objective hyperparameter optimization in a configurable setting to approximate the accuracy-efficiency trade-off in neural networks with sparse neural network training. Afterwards, we performed a hyperparameter analysis to discover how hyperparameters influence this trade-off. Our results show that the efficiency of neural networks can be heavily improved for a slight decrease in accuracy and that sparse neural network training plays a vital but complex role in this trade-off. |
Item Type: | Essay (Master) |
Clients: | Capgemini, Utrecht, The Netherlands |
Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
Subject: | 54 computer science |
Programme: | Computer Science MSc (60300) |
Link to this item: | https://purl.utwente.nl/essays/104468 |
Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Repository Staff Only: item control page