ACHIEVING EDGE DEVICE AUTONOMY

Otthi, Pavan

Permalink

https://hdl.handle.net/2142/124809 Copy

Description

Title

ACHIEVING EDGE DEVICE AUTONOMY

Author(s)

Otthi, Pavan

Issue Date

2023-05-01

Keyword(s)

dense neural network, ensemble modeling, filter pruning

Abstract

In the status quo, machine learning systems involving edge devices are often dependent on the connection between edge devices and the cloud. However, when this connection fails, edge devices must learn to operate and predict independently at least until the connection to the cloud server is established once again. We strive to achieve this edge device autonomy through an ensemble modeling strategy where various edge devices are prioritized with specific models in the ensemble. This will be done through taking the large, dense model in the cloud and splitting the model into smaller memory efficient models that are better suited for edge device storage. The first step is to find clusters of similarity between classes based on the output of the large dense neural network. To generate the simplified models in the ensemble, filter reduction algorithms specific to each cluster of classes are used so each model is best suited to predict its cluster. The expectation is that using the filter reduction algorithm allows for greater filter reduction for similar accuracy, allowing for greater model compression. After training the ensemble, an edge device’s current model will be decided based on a model transition algorithm. The results show that when applied to a specified cluster of classes, the filter reduction algorithm reduces the model significantly more than when applied to a general set of classes. Some challenges involved the filter reduction process from the original model. PyTorch’s library set the weights of the filter 0, achieving the goal of filter reduction with the reduction in floating point operations (FLOPS) but not memory. Thus, a new filter reduction method had to be developed to achieve the memory benefit which is the primary concern.

Type of Resource

text

Language

eng

Owning Collections