Business, political, and other social structures create strong motivation to understand the attitudes, motivations, feelings, and emotions of a population of interest. Social media is a rich source of self-disclosed information by individuals from all walks of life about virtually every domain of the human experience, but the vast quantity of data is impossible to effectively analyze without advanced natural language processing algorithms. This research creates a transfer learning based emotion classification model for Indonesian language Twitter data. Transfer learning consists of two steps: pre-training and fine tuning. Three variations of Indonesian Bidirectional Encoder Representations from Transformers (IndoBERT) are tested with hyperparameters tuned via designed experiment. The top IndoBERT model, tested on an open source corpus of 4,401 labeled Indonesian Tweets, outperforms all known prior studies with an F1 score of approximately 0.791. Additionally, this research explores the relationship between training set size and model validity for fine tuning of the transfer learning models; datasets ranging from 100 to 3900 observations are trained and then validated on five unique test sets. Results indicate that as few as 1000 observations can obtain results comparable to using the full training corpus.

In recent years, real-valued neural networks have demonstrated promising, and often striking, results across a broad range of domains. This has driven a surge of applications utilizing high-dimensional datasets. While many techniques exist to alleviate issues of high-dimensionality, they all induce a cost in terms of network size or computational runtime. This work examines the use of quaternions, a form of hypercomplex numbers, in neural networks. The constructed networks demonstrate the ability of quaternions to encode high-dimensional data in an efficient neural network structure, showing that hypercomplex neural networks reduce the number of total trainable parameters compared to their real-valued equivalents. Finally, this work introduces a novel training algorithm using a meta-heuristic approach that bypasses the need for analytic quaternion loss or activation functions. This algorithm allows for a broader range of activation functions over current quaternion networks and presents a proof-of-concept for future work.

Object detection algorithms have reached nearly superhuman levels within the last decade; however, these algorithms require large diverse training data sets to ensure their operational performance matches performance demonstrated during testing. The collection and human labeling of such data sets can be expensive and, in some cases, such as Intelligence, Surveillance and Reconnaissance of rare events it may not even be feasible. This research proposes a novel method for creating additional variability within the training data set by utilizing multiple models of generative adversarial networks producing both high- and low-quality synthetic images of vehicles and inserting those images alongside images of real vehicles into real backgrounds. This research demonstrates a 17.90% increase in mean absolute percentage error, on average, compared to the YOLOv4-Tiny Model trained on the original non-augmented training set as well as a 14.44% average improvement in the average intersection over union rate. In addition, our research adds to a small, but growing, body of literature indicating that the inclusion of low-quality images into training data sets is beneficial to the performance of computer vision models.