Abstract
Deep neural networks have become the default choice for many applications like image and video recognition, segmentation and other image and video related tasks.However, a critical challenge with these models is the lack of explainability.This requirement of generating explainable predictions has motivated the research community to perform various analysis on trained this http URL this study, we analyze the learned feature maps of trained models using MNIST images for achieving more explainable predictions.Our study is focused on deriving a set of primitive elements, here called visual concepts, that can be used to generate any arbitrary sample from the data generating distribution.We derive the primitive elements from the feature maps learned by the model.We illustrate the idea by generating visual concepts from a Variational Autoencoder trained using MNIST images.We augment the training data of MNIST dataset by adding about 60,000 new images generated with visual concepts chosen at random.With this we were able to reduce the reconstruction loss (mean square error) from an initial value of 120 without augmentation to 60 with augmentation.Our approach is a first step towards the final goal of achieving trained deep neural network models whose predictions, features in hidden layers and the learned filters can be well explained.Such a model when deployed in production can easily be modified to adapt to new data, whereas existing deep learning models need a re training or fine tuning. This process again needs a huge number of data samples that are not easy to generate unless the model has good explainability.
Abstract (translated)
URL
https://arxiv.org/abs/2109.00479
