An Approach Towards Deep Neural Network Model Activation Maximization using Information Retrieval Techniques
Advisor Information
Mahadevan Subramaniam
Location
MBSC Omaha Room 304 - G
Presentation Type
Oral Presentation
Start Date
4-3-2022 12:30 PM
End Date
4-3-2022 1:45 PM
Abstract
Deep learning is prevailing in a wide range of domains including Engineering, Medicine, surveillance with exceptional results. However, there have been several research works conducted in recent years to improve the explainability of Deep learning models due to their default black-box nature. In the light of this recent breakthrough, we propose a novel approach to enhance the predictive performance of deep neural classification models by understanding the inner neuron activations. These inner neuron activations represent visual features of input data and their corresponding class categories. The key idea is to mine and rank neuron activations of a trained DNN model using information retrieval techniques and identify unique features for a class over other classes and shared features for different classes. With this information, we propose an augmentation technique to improve training data by isolating the visual features unique to a class to maximize the class activations. For experiment purposes of this approach, we evaluated the approach using benchmarking datasets such as Mnist and CIFAR-10
Scheduling Link
1
An Approach Towards Deep Neural Network Model Activation Maximization using Information Retrieval Techniques
MBSC Omaha Room 304 - G
Deep learning is prevailing in a wide range of domains including Engineering, Medicine, surveillance with exceptional results. However, there have been several research works conducted in recent years to improve the explainability of Deep learning models due to their default black-box nature. In the light of this recent breakthrough, we propose a novel approach to enhance the predictive performance of deep neural classification models by understanding the inner neuron activations. These inner neuron activations represent visual features of input data and their corresponding class categories. The key idea is to mine and rank neuron activations of a trained DNN model using information retrieval techniques and identify unique features for a class over other classes and shared features for different classes. With this information, we propose an augmentation technique to improve training data by isolating the visual features unique to a class to maximize the class activations. For experiment purposes of this approach, we evaluated the approach using benchmarking datasets such as Mnist and CIFAR-10