This dissertation investigates two complementary ideas in the literature on machine learning and robotics--those of embodiment and intrinsic motivation--to address a unified framework for skill learning and knowledge acquisition. "Embodied" systems make use of structure derived directly from sensory and motor configurations for learning behavior. Intrinsically motivated systems learn by searching for native, hedonic value through interaction with the world. Psychological theories of intrinsic motivation suggest that there exist internal drives favoring open-ended cognitive development and exploration. I argue that intrinsically motivated, embodied systems can learn generalizable skills, acquire control knowledge, and form an epistemological understanding of the world in terms of behavioral affordances. I propose that the development of behavior results from the assembly of an agent's sensory and motor resources into state and action spaces that can be explored autonomously. I introduce an intrinsic reward function that can lead to the open-ended learning of hierarchical behavior. This behavior is factored into declarative "recipes" for patterned activity and common sense procedural strategies for implementing them in a variety of run-time contexts. These skills form a categorical basis for the robot to interpret and model its world in terms of the behavior it accords. Experiments conducted on a bimanual robot illustrate a progression of cumulative manipulation behavior addressing manual and visual skills. Such accumulation of skill over the long-term by a single robot is a novel contribution that has yet to be demonstrated in the literature.

Using deep learning to synthetically generate music is a research domain that has gained more attention from the public in the past few years. A subproblem of music generation is music extension, or the task of taking existing music and extending it. This work proposes the Continuer Pipeline, a novel technique that uses deep learning to take music and extend it in 5 second increments. It does this by treating the musical generation process as an image generation problem; we utilize latent diffusion models (LDMs) to generate spectrograms, which are image representations of music. The Continuer Pipeline is able to receive a waveform as an input, and its output will be what the pipeline predicts the next five seconds might sound like. We trained the Continuer Pipeline using the expansive diffusion model functionality provided by the HuggingFace platform, and our dataset consisted of 256x256 spectrogram images representing 5-second snippets of various hip-hop songs from Spotify. The musical waveforms generated by the Continuer Pipeline are currently at a much lower quality compared to human-generated music, but we affirm that the Continuer Pipeline still has many uses in its current state, and we describe many avenues for future improvement to this technology.

Given concern over shark attacks in coastal regions, the recent use of unmanned aerial vehicles (UAVs), or drones, has increased to ensure the safety of beachgoers. However, much of city officials' process remains manual, with drone operation and review of footage still playing a significant role. In pursuit of a more automated solution, researchers have turned to the usage of neural networks to perform detection of sharks and other marine life. For on-device solutions, this has historically required assembling individual hardware components to form an embedded system to utilize the machine learning model. This means that the camera, neural processing unit, and central processing unit are purchased and assembled separately, requiring specific drivers and involves a lengthy setup process. Addressing these issues, we look to the usage of smartphones as a novel integrated solution for shark detection. This paper looks at using an iPhone 14 Pro as the driving force for a YOLOv5 based model, and comparing our results to previous literature in shark-based object detection. We find that our system outperforms previous methods at both higher throughput and increased accuracy.

In the past 10 years, the growth of machine learning technology has been significant, largely due to the availability of large datasets for training. However, gathering a sufficient amount of data on a central server can be challenging. Additionally, with the rise of mobile networking and the large amounts of data generated by IoT devices, privacy and security issues have become a concern, resulting in government regulations such as GDPR, HIPAA, CCPA, and ADPPA. Under these circumstances, traditional centralized machine learning methods face a problem in that sensitive data must be kept locally for privacy reasons, making it difficult to achieve the desired learning outcomes. Federated learning (FL) offers a solution to this by allowing for a global shared model to be trained by exchanging locally computed optimums instead of sharing the actual data. Despite its success as a natural solution for IoT machine learning implementation, Federated learning (FL) still faces challenges with regards to security and performance. These include high communication costs between IoT devices and the central server, the potential for sensitive information leakage and reduced model precision due to the aggregation process in the distributed IoT network, and performance concerns caused by the heterogeneity of data and devices in the network. In this dissertation, I present practical and effective techniques with strong theoretical supports to address these challenges. To optimize communication resources, I introduce a new multi-server FL framework called MS-FedAvg. To enhance security, I propose a robust defense algorithm called LoMar. To address data heterogeneity, I present FedLGA, and for device heterogeneity, I propose FedSAM.

With artificial intelligence (AI) becoming ubiquitous in a broad range of application domains, the opacity of deep learning models remains an obstacle to adaptation within safety-critical systems. Explainable AI (XAI) aims to build trust in AI systems by revealing important inner mechanisms of what has been treated as a black box by human users. This thesis specifically aims to improve the transparency and trustworthiness of deep learning algorithms by combining attribution methods with image segmentation methods. This thesis has the potential to improve the trust and acceptance of AI systems, leading to more responsible and ethical AI applications. An exploratory algorithm called ESAX is introduced and shows how performance greater than other top attribution methods on PIC testing can be achieved in some cases. These results lay a foundation for future work in segmentation attribution.

Daily fantasy basketball presents interesting problems to researchers due to the extensive amounts of data that needs to be explored when trying to predict player performance. A large amount of this data can be noisy due to the variance within the sport of basketball. Because of this, a high degree of skill is required to consistently win in daily fantasy basketball contests. On any given day, users are challenged to predict how players will perform and create a lineup of the eight best players under fixed salary and positional requirements. In this thesis, we present a tool to assist daily fantasy basketball players with these tasks. We explore the use of several machine learning techniques to predict player performance and develop multiple approaches to approximate optimal lineups. We then compare each different heuristic and lineup creation combination, and show that our best combinations perform much better than random lineups. Although creating provably optimal lineups is computationally infeasible, by focusing on players in the Pareto front between performance and cost we can reduce the search space and compute near optimal lineups. Additionally, our greedy and evolutionary lineup search methods offer similar performance at a much smaller computational cost. Our analysis indicates that due to how player salaries are structured, it is generally preferred to construct a lineup consisting of a few stars and filling out the rest of the roster with average to mediocre players than to construct a lineup where all players are expected to perform about the same. Through these findings we hope that our research can serve as a future baseline towards developing an automated or semi-automated tool to optimize daily fantasy basketball.

Predicting the success of an early-stage startup has always been a major effort for investors and venture funds. Statistically, there are about 305 million total startups created in a year, but less than 10% of them succeed to become profitable businesses. Accurately identifying the signs of startup growth is the work of countless investors, and in recent years, research has turned to machine learning in hopes of improving the accuracy and speed of startup success prediction. To learn about a startup, investors have to navigate many different internet sources and often rely on personal intuition to determine the startup’s potential and likelihood of success. This thesis explores whether online data about a company, particularly general company data, previous funding events, published news articles, internet presence, and social media activity can be used to identify fast-growing startups. Data collected from Crunchbase, the Google Search API, and Twitter was used to predict whether a company will raise a round of funding within a fixed time horizon. A total of ten machine learning models were evaluated and the CatBoost ensemble method achieved the best performance with precision, recall, and F1 scores of 0.663, 0.827, and 0.736 respectively for predicting funding within 3 years. The same ensem- ble method achieved F1 scores of 0.528, 0.683, 0.736, 0.763, and 0.777 at predicting funding 1-5 years into the future. The final objective was to predict whether a startup that had already raised an angel or seed round would raise another investment within a one-year horizon. The CatBoost model with a 0.75 cutoff achieved precision and F0.1 scores of 0.790 and 0.774, beating the results of previous work in this field.

While deep neural networks have shown impressive performance in computer vision tasks, natural language processing, and other domains, the sizes and inference times of these models can often prevent them from being used on resource-constrained systems. Furthermore, as these networks grow larger in size and complexity, it can become even harder to understand the learned representations of the input data that these networks form through training. These issues of growing network size, increasing complexity and runtime, and ambiguity in the understanding of internal representations serve as guiding points for this work. In this thesis, we create a neural network that is capable of predicting up to three path waypoints given an input image. This network will be used in conjunction with other networks to help guide an autonomous robotic vehicle. Since this neural network will be deployed to an embedded system, it is important that our network is efficient. As such, we use a network compression technique known as L1 norm pruning to reduce the size of the network and speed up the inference time, while retaining similar loss. Furthermore, we investigate the effects that pruning has on the internal learned representations of models by comparing unpruned and pruned network layers using projection weighted canonical correlation analysis (PWCCA). Our results show that for deep convolutional neural networks (CNN), PWCCA similarity scores between early convolutional layers start low and then gradually increase towards the final layers of the network, with some peaks in the intermediate layers. We also show that for our deep CNN, linear layers at the end of the network also exhibit very high similarity, serving to guide the dissimilar representations from intermediate convolutional layers to a common representation that yields similar network performance between unpruned and pruned networks.

Due to the difficulty of obtaining formal proofs, there is increasing interest in partially or completely automating proof search in interactive theorem provers. Despite being a theorem prover with an active community and plentiful corpus of 170,000+ theorems, no deep learning system currently exists to help automate theorem proving in ACL2. We have developed a machine learning system that generates recommendations to automatically complete proofs. We show that our system benefits from the copy mechanism introduced in the context of program repair. We make our system directly accessible from within ACL2 and use this interface to evaluate our system in a realistic theorem proving environment.

Digital Democracy is a CalMatters and California Polytechnic State University initia-tive to promote transparency in state government by increasing access to the Califor-nia legislature. While Digital Democracy is made up of many resources, one founda-tional step of the project is obtaining accurate, timely transcripts of California Senateand Assembly hearings. The information extracted from these transcripts providescrucial data for subsequent steps in the pipeline. In the context of Digital Democracy,upleveling is when humans verify, correct, and annotate the transcript results afterthe legislative hearings have been automatically transcribed. The upleveling processis done with the assistance of a software application called the Transcription Tool.The human upleveling process is the most costly and time-consuming step of the Dig-ital Democracy pipeline. In this thesis, we hypothesize that we can make significantreductions to the time needed for upleveling by using Natural Language Processing(NLP) systems and techniques. The main contribution of this thesis is engineeringa new automatic transcription pipeline. Specifically, this thesis integrates a new au-tomatic speech recognition service, a new speaker diarization model, additional textpost-processing changes, and a new process for speaker identification. To evaluate the system’s improvements, we measure the accuracy and speed of the newly integrated features and record editor upleveling time both before and after the additions.