Abstract
This dissertation delves into the novel application of neuromorphic Spiking Neural Networks (SNNs) for classifying water potability, an important step in acquiring access to safe drinking water and advancing environmental monitoring. The research addresses the growing need for efficient, accurate water quality monitoring systems by exploring the efficacy of SNNs, particularly against the backdrop of traditional machine learning models. The study is anchored on the Leaky Integrate-and-Fire (LIF) neuron model within the snnTorch framework, a strategic choice aimed at mimicking biological neuron processes for enhanced computational efficiency and accuracy.
The methodology involves constructing a fully connected SNN, with synaptic connections and neuron dynamics intricately modeled to simulate biological processes. A key feature of this approach is the input encoding through rate coding, which transforms real-valued water quality data into spike trains. This is complemented by a supervised learning rule utilizing Binary Cross-Entropy with Logits, optimized using the Adam algorithm, to refine the model's predictive accuracy.
The research is driven by several objectives: to evaluate the classification accuracy of SNNs, assess their energy consumption, examine their computational performance, and conduct a comparative analysis with state-of-the-art methods for water quality classification. A key emphasis is placed on balancing energy efficiency and classification accuracy, an essential factor in the effective implementation of these models in real-world applications.
Experimental results revealed moderate accuracy levels across training, validation, and test sets, with the highest accuracy observed in the training set (70.65%) and lower levels in validation (64.73%) and test sets (63.13%). This pattern suggested a potential overfitting to training data. The model's precision generally exceeded its recall, indicating a conservative prediction approach aimed at minimizing false positives. The F1 scores and AUC-ROC scores, though reasonable, underlined the importance of balancing precision and recall.
A pivotal aspect of this study is the energy efficiency of the SNN model, with an estimated consumption of approximately 140.7034×10−6Joules, highlighting its potential in low-energy,
[viii]
high-efficiency computing environments. This energy efficiency, coupled with moderate accuracy and computational speed, underscores the SNN's viability in scenarios where energy conservation is crucial.
Comparatively, the SNN model showcased effective performance but did not surpass the most advanced models like CNNs in accuracy. However, its energy efficiency and computational speed offer a valuable trade-off, particularly beneficial in edge computing scenarios.
In sum, this research significantly advances the field of neuromorphic computing, especially in environmental sensing. It positions SNNs as a competitive alternative to traditional computational models, especially where energy efficiency is prioritized over maximal accuracy. The findings illuminate the potential of SNNs in real-world applications like water quality monitoring in resource-constrained environments, setting the stage for future enhancements in model accuracy, generalizability, and practical deployment.
Keywords: Spiking Neural Networks, Water Potability, Neuromorphic Computing, Machine Learning, Environmental Monitoring, Energy Efficiency.