Socially assistive robots (SARs) can act as assistants and caregivers, interacting and communicating with people through touch gestures. There has been ongoing research on using them as companion robots for children with autism as therapy assistants and playmates. Building touch-perception systems for social robots has been a challenge. The sensors must be designed to ensure comfortable and natural user interaction while recording high-quality data. The sensor must be able to detect touch gestures. Accurate touch gesture classification is challenging as different users perform the same touch gesture in their own unique way. This study aims to build and evaluate a skin-like sensor by replicating a recent paper introducing a novel silicone-based sensor design. A touch gesture classification is performed using deep-learning models to classify touch gestures accurately. This study focuses on 8 gestures: Fistbump, Hitting, Holding, Poking, Squeezing, Stroking, Tapping, and Tickling. They were chosen based on previous research where specialists determined which gestures were essential to detect while interacting with children with autism. In this work, a user study data collection was conducted with 20 adult subjects, using the skin-like sensor to record gesture data and a load cell underneath to record the force. Three different types of input were used for the touch gesture classification: skin-like sensor & load cell data, only skin-like sensor data, and only load cell data. A Convolutional Neural Network-Long Short Term Memory (CNN-LSTM) neural network architecture was developed for inputs with skin-like sensor data, and an LSTM network for only load cell data. This work achieved an average accuracy of 94% with skin-like sensor & load cell data, 95% for only skin-like sensor data, and 45% for only load cell data after a stratified 10-fold validation. This work also performed subject-dependent splitting and achieved accuracies of 69% skin-like sensor & load cell data, 66% for only skin-like sensor data, and 31% for only load cell data, respectively.

Generative models are deep neural network-based models trained to learn the underlying distribution of a dataset. Once trained, these models can be used to sample novel data points from this distribution. Their impressive capabilities have been manifested in various generative tasks, encompassing areas like image-to-image translation, style transfer, image editing, and more. One notable application of generative models is data augmentation, aimed at expanding and diversifying the training dataset to augment the performance of deep learning models for a downstream task. Generative models can be used to create new samples similar to the original data but with different variations and properties that are difficult to capture with traditional data augmentation techniques. However, the quality, diversity, and controllability of the shape and structure of the generated samples from these models are often directly proportional to the size and diversity of the training dataset. A more extensive and diverse training dataset allows the generative model to capture overall structures present in the data and generate more diverse and realistic-looking samples. In this dissertation, I present innovative methods designed to enhance the robustness and controllability of generative models, drawing upon physics-based, probabilistic, and geometric techniques. These methods help improve the generalization and controllability of the generative model without necessarily relying on large training datasets. I enhance the robustness of generative models by integrating classical geometric moments for shape awareness and minimizing trainable parameters. Additionally, I employ non-parametric priors for the generative model's latent space through basic probability and optimization methods to improve the fidelity of interpolated images. I adopt a hybrid approach to address domain-specific challenges with limited data and controllability, combining physics-based rendering with generative models for more realistic results. These approaches are particularly relevant in industrial settings, where the training datasets are small and class imbalance is common. Through extensive experiments on various datasets, I demonstrate the effectiveness of the proposed methods over conventional approaches.

This thesis encompasses a comprehensive research effort dedicated to overcoming the critical bottlenecks that hinder the current generation of neural networks, thereby significantly advancing their reliability and performance. Deep neural networks, with their millions of parameters, suffer from over-parameterization and lack of constraints, leading to limited generalization capabilities. In other words, the complex architecture and millions of parameters present challenges in finding the right balance between capturing useful patterns and avoiding noise in the data. To address these issues, this thesis explores novel solutions based on knowledge distillation, enabling the learning of robust representations. Leveraging the capabilities of large-scale networks, effective learning strategies are developed. Moreover, the limitations of dependency on external networks in the distillation process, which often require large-scale models, are effectively overcome by proposing a self-distillation strategy. The proposed approach empowers the model to generate high-level knowledge within a single network, pushing the boundaries of knowledge distillation. The effectiveness of the proposed method is not only demonstrated across diverse applications, including image classification, object detection, and semantic segmentation but also explored in practical considerations such as handling data scarcity and assessing the transferability of the model to other learning tasks. Another major obstacle hindering the development of reliable and robust models lies in their black-box nature, impeding clear insights into the contributions toward the final predictions and yielding uninterpretable feature representations. To address this challenge, this thesis introduces techniques that incorporate simple yet powerful deep constraints rooted in Riemannian geometry. These constraints confer geometric qualities upon the latent representation, thereby fostering a more interpretable and insightful representation. In addition to its primary focus on general tasks like image classification and activity recognition, this strategy offers significant benefits in real-world applications where data scarcity is prevalent. Moreover, its robustness in feature removal showcases its potential for edge applications. By successfully tackling these challenges, this research contributes to advancing the field of machine learning and provides a foundation for building more reliable and robust systems across various application domains.