Modern deep learning methods have achieve remarkable performance on a variety of computer vision tasks, but frequently require large, well-balanced training datasets to achieve high-quality results. Data-efficient performance is critical for downstream tasks such as automated driving or facial recognition. We propose two methods of data-efficient learning for the tasks of image synthesis and instance segmentation. We first propose a method of high-quality and diverse image generation from finetuning to only 5-100 images. Our method factors a pretrained model into a small but highly expressive weight space for finetuning, which discourages overfitting in a small training set. We validate our method in a challenging few-shot setting of 5-100 images in the target domain. We show that our method has significant visual quality gains compared with existing GAN adaptation methods. Next, we introduce a simple adaptive instance segmentation loss which achieves state-of-the-art results on the LVIS dataset. We demonstrate that the rare categories are heavily suppressed by textit{correct background predictions}, which reduces the probability for all foreground categories with equal weight. Due to the relative infrequency of rare categories, this leads to an imbalance that biases towards predicting more frequent categories. Based on this insight, we develop DropLoss -- a novel adaptive loss to compensate for this imbalance without a trade-off between rare and frequent categories. / Master of Science / Many of the impressive results seen in modern computer vision rely on learning patterns from huge datasets of images, but these datasets may be expensive or difficult to collect. Many applications of computer vision need to learn from a very small number of examples, such as learning to recognize an unusual traffic event and behave safely in a self-driving car. In this thesis we propose two methods of learning from only a few examples. Our first method generates novel, high-quality and diverse images using a model fine-tuned on only 5-100 images. We start with an image generation model that was trained a much larger image set (70K images), and adapts it to a smaller image set (5-100 images). We selectively train only part of the network to encourage diversity and prevent memorization. Our second method focuses on the instance segmentation setting, where the model predicts (1) what objects occur in an image and (2) their exact outline in the image. This setting commonly suffers from long-tail distributions, where some of the known objects occur frequently (e.g. "human" may occur 1000+ times) but most only occur a few times (e.g. "cake" or "parrot" may only occur 10 times). We observed that the "background" label has a disproportionate effect of suppressing the rare object labels. We use this to develop a method to balance suppression from background classes during training.

abstract: Robotic lower limb prostheses provide new opportunities to help transfemoral amputees regain mobility. However, their application is impeded by that the impedance control parameters need to be tuned and optimized manually by prosthetists for each individual user in different task environments. Reinforcement learning (RL) is capable of automatically learning from interacting with the environment. It becomes a natural candidate to replace human prosthetists to customize the control parameters. However, neither traditional RL approaches nor the popular deep RL approaches are readily suitable for learning with limited number of samples and samples with large variations. This dissertation aims to explore new RL based adaptive solutions that are data-efficient for controlling robotic prostheses. This dissertation begins by proposing a new flexible policy iteration (FPI) framework. To improve sample efficiency, FPI can utilize either on-policy or off-policy learning strategy, can learn from either online or offline data, and can even adopt exiting knowledge of an external critic. Approximate convergence to Bellman optimal solutions are guaranteed under mild conditions. Simulation studies validated that FPI was data efficient compared to several established RL methods. Furthermore, a simplified version of FPI was implemented to learn from offline data, and then the learned policy was successfully tested for tuning the control parameters online on a human subject. Next, the dissertation discusses RL control with information transfer (RL-IT), or knowledge-guided RL (KG-RL), which is motivated to benefit from transferring knowledge acquired from one subject to another. To explore its feasibility, knowledge was extracted from data measurements of able-bodied (AB) subjects, and transferred to guide Q-learning control for an amputee in OpenSim simulations. This result again demonstrated that data and time efficiency were improved using previous knowledge. While the present study is new and promising, there are still many open questions to be addressed in future research. To account for human adaption, the learning control objective function may be designed to incorporate human-prosthesis performance feedback such as symmetry, user comfort level and satisfaction, and user energy consumption. To make the RL based control parameter tuning practical in real life, it should be further developed and tested in different use environments, such as from level ground walking to stair ascending or descending, and from walking to running. / Dissertation/Thesis / Doctoral Dissertation Electrical Engineering 2020