Ternary Neural Networks for Efficient Biometric Data Analysis

Advancements in artificial intelligence, especially deep neural networks (DNNs), have driven significant progress in fields like computer vision and natural language processing. However, the high computational demands of these models present challenges for deployment on power-constrained and memory-limited devices, such as wearable ones. This thesis addresses these challenges in the context of gait analysis for biometric applications on resource-constrained wearable devices. While smartphones and smartwatches offer convenient platforms for capturing gait data via IMU sensors, their limited computational resources hinder real-time DNN deployment. To overcome this, we introduce a Ternary Neural Network (TNN) framework that combines quantization and pruning to achieve high sparsity, setting most weights to zero to reduce memory and energy usage while maintaining model accuracy. Our approach dynamically adjusts quantization during training, achieving sparsity rates above 90% with entropy levels below 1 bit per symbol, making the model highly compressible and effective simultaneously. We evaluate the model on two biometric tasks: identification, where the model differentiates individuals based on gait patterns, and authentication, which verifies identity by comparing gait against a reference. Results indicate that TNNs retain strong discriminative power, enabling efficient and accurate gait-based biometric recognition.