Project

DESCRIPTION:

Augmentations in a deep neural network (DNN) are introduced to enhance image classification accuracy while significantly reducing inference time. The architecture comprises a hybrid framework that combines binary classifiers based on Convolutional Neural Networks (CNNs) with refined classifiers employing Vision Transformers (ViTs). Inference time and energy optimization aspects of the architecture for multi-GPU systems will be investigated.

We investigate the augmentation of an input CNN with simple DNN structures to the accuracy of specific low-accuracy classes. The proposed approach includes an algorithm that optimizes the integration of all classes within the cascading structure to improve overall accuracy.  The added components will be designed based on the appearance probabilities of the classes they process. It is proposed that they will be trained by considering the distributions of images they receive from their preceding components. Some added components may utilize Vision Transformers (ViTs).

We will also investigate DNN-based scheduling problems on multi-GPU systems. Each user has selected a DNN for a specific data set that optimizes classification accuracy using the proposed architectural considerations. Each data set must be processed within a specified deadline. The DNN-based tasks must be scheduled on multi-GPU systems to minimize the total energy subject to the task deadlines. Scheduling will benefit from appropriate batch processing for each task.