In this project, we address the problem of re-training a deep neural network for a new class with limited training data (n to n+1 class learning) using a novel concept of Deep part embeddings (DPEs). DPEs are sub-networks of neuron activation extracted from a trained network identifying a visual and distinguishable element of a class. We identify visual elements that intuitively constitute a new class and extract the corresponding DPEs from the network pre-trained for the class sharing the identified visual element.
Finally, we assemble them into a new network and re-train the model on limited samples of the new class and a subset of data from `n’ classes to achieve high accuracy on the new class without significantly affecting the accuracy of n classes. We studied and generated results for DPE integration under two configurations- (i) sequential, when DPEs are sourced from different CNN architectures and (ii) shared; when DPEs are sourced from the same CNN architecture.
Abhinav Jain
Machine Learning Engineer
My research interests include computer vision, machine learning and deep reinforcement learning.
Publications
Learning Convolutional Neural Networks with Deep Part Embeddings
