Presentation Type
Poster Presentation
Mentor/Supervising Professor Name
Alférez, Harvey
Abstract (Description of Research)
Artificial Neural Networks (ANNs) require substantial memory and computational resources, limiting their deployment on resource-constrained devices. Our contribution is a compression method using Mean Compression (MC) to reduce ANN size while preserving functionality and accuracy. MC consolidates connections with similar edge weights into meta-nodes with averaged values. Unlike traditional pruning that only removes connections among neurons, MC restructures networks by recomputing weights and creating meta-nodes. Additionally, unlike fixed pruning thresholds, MC uses flexible weight range patterns. Applied to multilayer perceptron (MLP), ANNs are made more accessible for deployment on constrained devices as proven in several experiments. Specifically, across five classification UCI datasets, there was a mean improvement of accuracy of 7.23% when implementing MC as opposed to pruning, with a similar reduction of storage in terms of Megabytes when compared to pruning.
Included in
Towards Smaller Artificial Neural Network Using Mean Compression*
Artificial Neural Networks (ANNs) require substantial memory and computational resources, limiting their deployment on resource-constrained devices. Our contribution is a compression method using Mean Compression (MC) to reduce ANN size while preserving functionality and accuracy. MC consolidates connections with similar edge weights into meta-nodes with averaged values. Unlike traditional pruning that only removes connections among neurons, MC restructures networks by recomputing weights and creating meta-nodes. Additionally, unlike fixed pruning thresholds, MC uses flexible weight range patterns. Applied to multilayer perceptron (MLP), ANNs are made more accessible for deployment on constrained devices as proven in several experiments. Specifically, across five classification UCI datasets, there was a mean improvement of accuracy of 7.23% when implementing MC as opposed to pruning, with a similar reduction of storage in terms of Megabytes when compared to pruning.