My current research explores quantization and approximate floating-point (AFP) techniques as methods to optimize machine learning models for real-world deployment. Quantization reduces numerical precision which is converting floating-point values into lower-bit formats such as INT8, FP16, INT4, or even INT2, to improve memory efficiency, computational speed, and power consumption. This process is essential for running ML models on edge devices, mobile AI platforms, and cloud inference systems. Different approaches, including Post-Training Quantization (fast but may reduce accuracy), Quantization-Aware Training (higher accuracy retention), and Dynamic Quantization (quantizing activations at runtime), each balance trade-offs between performance and precision. For example, INT8 models can use up to 75% less memory than FP32 and achieve 4x–10x speedups on AI accelerators. Similarly, AFP methods approximate floating-point representations to reduce complexity while preserving accuracy, enabling more efficient computations. Together, quantization and AFP provide powerful strategies for making AI models smaller, faster, and more practical for deployment across diverse hardware systems.

For this project, I will be running simulations and collecting data using a combination of tools, including Tableau, R, and PyTorch. Tableau will allow me to visualize and interpret survey data effectively, R will be used for statistical analysis and data modeling, and PyTorch will support the development and testing of machine learning models. Together, these platforms provide a strong framework for managing data, performing computations, and generating insights that support the goals of my research.