Undertaken research into low-data training for neural networks and few-shot learning. In order to enhance the performance of such systems, I designed a novel Generative Adversarial Network training scheme which involved using an image conditional GAN to meta-learn to generate images that come from the same cardinality as the input image, yet are different enough to be considered a different sample. By leveraging the input noise one could use the DAGAN to generate multiple plausible variations of the input image, thus augmenting the data using augmentations that are learned from the data. Undertaking this work required a strong understanding of deep neural network classifiers and few shot learners and the problems that may be causing them to fail in low data regimes. Furthermore, it required a deep understanding of GANs and deep generative models as well as the ability to build very complex graphs in tensorflow and fine tune both the hyperparameters and the efficiency (using a Multi-GPU system) under the constraint of a few computational resources.

A deep learning driven system with 3 pipelines was implemented which could identify moving objects using advanced background subtraction based algorithms and then classify said objects with accuracy exceeding 90% over several categories (e.g. Cat, Dog, Car, Human). Then another stage in the system would make a decision based on the users preferences (e.g. if used for home security the activation classes would be humans and vehicles). The system was efficient and robust enough to be implemented for an Android device and allows the device to send SMS messages and emails containing pictures of potential issues. This required a good understanding of deep neural network classifiers and transfer learning. Furthermore, it required fluency in Java, Android SDK, Caffe and Unix. Additionally it required the understanding of embedding device's constraints and optimization of performance on Android devices.

Got 86% in my final year dissertation research project by researching the methods / implementing simulations / designing / building apparatus / proving technique experimentally and finally designing a professional grade sensing system capable of converting any sensor into a self-monitoring sensor with the abilities to validate it's own sensor elements / pinpoint faulty element / compensate for fault using machine learning and finally predicting statistically faults before they occur so that the sensor may be replaced before it fails. This project required a good understanding of sensor electronics and electromagnetic theory as well as fluency in C++ and embedded system performance optimization. Furthermore, it required design, experimentation and research on complex electronic systems. Finally, it required good project management skills as not only were the original targets achieved but multiple extra objectives were achieved enough to actually produce a professional grade PCB printed circuit that had the system implemented on it, ready to be used on any resistive sensors.