The Inspiration
What began as an academic inquiry into the boundaries of calculator programming capabilities evolved into a comprehensive neural network implementation. This project effectively illustrates that meaningful machine learning applications remain viable even within the context of severe hardware constraints, offering valuable insights for resource-constrained computing environments.
Demonstration Video
Observe the neural network's word correction capabilities in action on the TI-84 Plus Silver Edition.
Pattern Recognition Capabilities
This demonstration illustrates the neural network's sophisticated pattern recognition capabilities. After correctly identifying the word "BEEN", the system proceeds to analyze two additional words, both of which contain characters typed with adjacent keys on the calculator.
The neural network demonstrates remarkable accuracy in determining the correct target words by analyzing the relative distances between letters in the alphabet. When presented with input where letters are nearly equidistant from the target letter (as in the two words entered after "BEEN"), the system can discern that these letters maintain the same relative distance pattern as those in the target word.
This capability enables the neural network to correctly identify the intended words despite character substitutions or transpositions, showcasing its effectiveness as an autocorrection system even within the severe hardware constraints of a calculator.
This initiative was conceived as a proof of concept to demonstrate the viability of executing complex machine learning algorithms on severely constrained hardware. The TI-84 Plus Silver Edition, with its mere 24KB of RAM and utilization of a legacy programming language, presents a compelling case study in computational efficiency.
HERMES OPTIMUS employs a feedforward neural network with a precisely calibrated 4-60-12 architecture and sigmoid activation functions. This configuration enables the system to effectively recognize and autocorrect misspelled or scrambled 4-letter words with remarkable accuracy despite the hardware limitations.
The development process necessitated addressing substantial memory constraints and variable overflow errors. These challenges were resolved through innovative approaches, including external training methodologies and the conversion of neural network weights into calculator-compatible file formats.