Deep learning for human engineered systems: Weak supervision, interpretability and knowledge embedding

Pattern recognition has its origins in engineering while machine learning developed from computer science. Today, artificial intelligence (AI) is a booming field with many practical applications and active research topics that deals with both pattern recognition and machine learning. We now use software and applications to automate routine labor, understand speech (using Natural Language Processing) or images (extracting hierarchical features and patterns for object detection and pattern recognition), make diagnoses in medicine, even intricate surgical procedures and support basic scientific research. This thesis deals with exploring the application of a specific branch of AI, or a specific tool, Deep Learning (DL) to real world engineering problems which otherwise had been difficult to solve using existing methods till date. Here we focus on different Deep Learning based methods to deal with several such problems. We also explore the inner workings of such models through an explanation stage for each of the applied DL based strategies that gives us a sense of how such typical black box models work, or as we call it, an explanation stage for the DL model. This explanation framework is an important step as previously, Deep Learning based models were thought to be frameworks which produce good results (classification, object detection, object recognition to name a few), but with no explanations or immediately visible causes as to why it achieves the results it does. This made Deep Learning based models hard to trust amongst the scientific community. In this thesis, we aim to achieve just that by deploying such explanation frameworks, which will be discussed later in the subsequent chapters. We use one such explanation framework to develop a surrogate model to predict properties of microstructures as well. Furthermore, we dig deep into the realm of semi-supervised or weakly supervised learning which utilizes minimal data for its training phase and develop a novel framework capable of accurately predicting yields of crops like sorghum. We utilize this framework to learn from available data in an online fashion, annotate new data, significantly reduce manual annotation time and at the same time predict crop yield, a first of its kind within the target application domain. We also propose a new generative modeling approach, Surrogate Invariance Network (S-InvNet), that can efficiently model data spaces with known invariances. We devise an adversarial training algorithm to encode such invariances into the data distribution. We validate our framework by reconstructing two-phase microstructures with desired physical properties. The physical properties are governed by a data-driven surrogate model which acts as the invariance for our case. In essence, we fuse a generative and a classification model with the classification model acting as an invariance checker that enforces the property of the target microstructure. How a Deep Learning model can iteratively solve a non-linear Partial Differential Equation (PDE) based on a given set of initial conditions and using a Physics-governed loss function (instead of traditional loss functions used to optimize Deep Learning models) is presently an area of great research interest. We explore this as well. Previous literature deal with utilizing a single loss function to solve such examples of PDE-governed systems. Moreover, the most recent work proposes to solve only a simple spatially varying PDE (Darcy Flow). We extend the framework to deal with both spatial and time-varying PDEs (Burgers\u27 Equation). Furthermore, we also propose an alternating minimization process for optimizing neural networks as PDE-solvers. This alternatively minimizes two separate loss functions. We note that this process works just as well as the single loss minimization and in certain cases, performs even better. From a neural network standpoint, we use a Convolutional Encoder-Decoder framework as our PDE-surrogate