Extracting the embedded knowledge in class visualizations from artificial neural networks for applications in dataset and model compression and combinatorial optimization

Abstract

Artificial neural networks are efficient learning algorithms, considered universal approxima-tors for solving numerous real-world problems in areas like computer vision, language processing, or reinforcement learning. To approximate any given function, neural networks train a large number of parameters that can go up to the millions or even billions in some cases. The large number of parameters and hidden layers in neural networks makes them hard to interpret, which is why they are often referred to as black boxes. In the quest to make artificial neural networks interpretable in the field of computer vision, feature visualization stands outas one of the most developed and promising research directions. While feature visualizations are a useful tool to gain insights about the underlying function learned by a neural network, they are still considered simply as visual aids that require human interpretation. In this doctoral work, we propose that feature visualizations—class visualizations in particular—are analogous to mental imagery in humans and contain the knowledge that the model extracted from the training data. Therefore, when correctly generated, class visualiza-tions can be considered as a conceptual compression of the data used to train the underlying model, resembling the experience of perceiving the actual training samples just as mental imagery resembles the real experience of perceiving the actual physical event. We present results showing that class visualizations can be considered a conceptual compression of the training data used to train the underlying model and present a methodology that enables the use of class visualizations as training data. To achieve this goal, we show that class visualizations can be used as training data to develop new models from scratch, achieving, in some cases, the same accuracy as the underlying model. Additionally, we explore the nature of class visualizations through different experiments to gain insights on what exactly class visualizations represent and what knowledge is embedded in them. To do so, we com- pare class visualizations to the class average image from the training data and demonstrate how the other classes that a model is trained on affect the shape and the knowledge embedded in a class visualization. We show that class visualizations are equivalent to visualizing the weight matrices of the output neurons in shallow network architectures and demonstrate that class visualizations can be used as pretrained convolutional filters. We experimentally show the potential of class visualizations for extreme model compression purposes. Finally, we present a novel methodology to enable the use of Artificial Neural Networks along with class visualizations for the solution of combinatorial optimization problems, such as the 2D Bin Packing Problem, by training an Artificial Neural Network to score potential solutions to a 2D BPP and then using that network to generate an ’optimal’ (local optima) solution to the problem by extracting a class visualization from the network via backpropagation to the network’s input. Even though we show the use of class visualizations as a tool to solve the bin packing problem, it is important to note that class visualizations have the potential to be used in the same way to solve other types of combinatorial optimization problems. For other types of combinatorial optimization problems, we just need to design a neural network that is capable of scoring solutions to the particular combinatorial optimization problem and extract class visualizations from such a network to generate a candidate solution to the problem.