Search

The detection of anomalous radioactive sources in environments characterized by a high level of variation in the background radiation is a challenging problem in nuclear security. A variety of natural and artificial sources contribute to background radiation dynamics including variations in the absolute and relative concentrations of naturally occurring radioisotopes in different materials, the wet-deposition of $^{222}$Rn daughters during precipitation, and background suppression due to physical objects in the detector scene called ``clutter." This dissertation presents a new datacentric algorithm for radiation anomaly detection in dynamic background environments. The algorithm is based on a custom deep neural autoencoder architecture called the Autoencoder Radiation Anomaly Detection (ARAD) model. An autoencoder is a type of neural network that compresses data at its input through a series of computational layers into a dimensionally-constrained representation called the latent space. It then uses another set of layers to decompress the latent space back into its original dimensionality. When trained on typical background radiation data, ARAD learns an efficient representation of the components comprising typical gamma ray background radiation spectra. If a gamma ray spectrum containing anomalous radiation signatures is presented to ARAD, it is unable to reconstruct the spectrum at its input which can be used to trigger a detection alarm. ARAD was trained and tested on both simulated and real world gamma ray radiation data. Results indicate that ARAD shows a high level of resilience to background dynamics while also displaying good detection performance on a variety of simulated and real world sources.

Spiking neural networks (SNNs) have recently gained a lot of attention for use in low-power neuromorphic and edge computing. On their own, SNNs are difficult to train, owing to their lack of a differentiable activation function and their inherent tendency towards chaotic behavior. This work takes a strictly neuroscience-inspired approach to designing and training SNNs. We demonstrate that the use of neuromodulated synaptic time dependent plasticity (STDP) can be used to create a variety of different learning paradigms including unsupervised learning, semi-supervised learning, and reinforcement learning. In order to tackle the highly dynamic and potentially chaotic spiking behavior of SNNs both during training and testing, we discuss a variety of neuroscience-inspired hoemeostatic mechanisms for keeping the network's activity in a healthy range. All of these concepts are brought together in the development of a SNN model that is trained and tested on the MNIST handwritten digits dataset. In order to achieve this, we introduce also introduce a custom Python package called Ganglion that can be used to rapidly design and test SNN architectures.