TensorGen

This paper presents a comparative study of three gradient computation algorithms-Finite Differentiation, Symbolic Differentiation, and Reverse Automatic Differentiation (Reverse AD)-within the TensorGen deep learning framework.TensorGen simplifies the creation, training, and experimentation of feed-forward neural networks (FFNs), automating model workflows.We analyze the efficiency, accuracy, and scalability of each method, with Reverse AD demonstrating both accuracy and efficiency for large-scale neural networks, while Finite Differentiation and Symbolic Differentiation present trade-offs in accuracy and complexity.Our findings identify the most suitable algorithm for different deep learning tasks, offering insights to optimize training processes and workflows in TensorGen, contributing to improved performance and resource management.