Why I Built a Neural Network for a Calculator

A Proof of Concept

The primary impetus for developing HERMES OPTIMUS stemmed from a fundamental research question: Is it feasible to implement a fully functional neural network on a calculator with merely 24KB of RAM, utilizing a programming language developed in the 1980s?

This research endeavor was not primarily concerned with practical applications but rather with expanding our understanding of computational possibilities within severely constrained hardware environments. In our current technological landscape, where machine learning models typically require gigabytes of memory and specialized GPU hardware, this project deliberately explores the opposite end of the computational spectrum.

The TI-84 Plus Silver Edition represents an accessible computing platform for many students, thereby offering an intriguing medium for demonstrating fundamental machine learning concepts without necessitating specialized hardware resources. This accessibility enhances the project's educational value.

Educational Value

Constructing a neural network from first principles on such constrained hardware necessitates a comprehensive understanding of the core theoretical foundations. The absence of high-level abstraction libraries means that every component of the neural network architecture must be meticulously implemented and optimized manually.

This project forced me to deeply understand:

• Feedforward neural network architecture
• Activation functions and their implementations
• Weight initialization strategies
• Memory optimization techniques
• The trade-offs between accuracy and performance

By stripping away all the abstractions, I gained insights into neural networks that might have been missed when using modern frameworks like TensorFlow or PyTorch.

The Challenge

The intellectual satisfaction derived from resolving complex technical challenges represents a significant motivational factor. The inherent constraints of the TI-84 Plus Silver Edition presented a series of substantive technical obstacles that required innovative solutions:

• Limited RAM (24KB) for storing weights and variables
• Slow processor speed (15 MHz)
• TI-BASIC's limitations as a programming language
• Limited display capabilities for debugging

Overcoming these constraints required creative solutions and workarounds, making the project intellectually stimulating and rewarding when successful.

Nostalgia and Appreciation

For many professionals in the field, the TI-84 Plus calculator series represented an initial introduction to programming concepts. There exists a meaningful historical continuity in revisiting this platform with the accumulated knowledge and technical expertise developed through years of programming experience.

This project is also a way to appreciate how far computing has come. The neural networks we run today on powerful hardware have their roots in these fundamental concepts that can be implemented even on a calculator.

By demonstrating that even a calculator can run a neural network, we gain perspective on the incredible capabilities of modern machine learning systems and the innovation that has occurred in this field.