TinyEye: Image Classification on a $4 Microcontroller — Zero Floating-Point, Zero GPU

Live demo of CIFAR-10 image classification running entirely on a Raspberry Pi Pico ($4, 256KB RAM, ARM Cortex-M0+) using integer-only arithmetic. No floating-point unit. No GPU. No cloud.

The core innovation is CIA (Constructive Integer Attention) — a new mathematical framework that replaces floating-point matrix operations in neural networks with pure integer computation derived from Egyptian fraction decomposition theory.

In the demo, we also briefly show CIFAR-10 sample images to the Pico and show real-time classification results on screen. If possible, I also briefly show TinyLLM — a language model generating grammatical English sentences on the same $4 chip — to demonstrate this is a general-purpose inference architecture, not a single-task trick.

Technical walkthrough covers:

• The CIA architecture that eliminates floating-point entirely (not quantization — the model is natively integer)
• Memory layout tricks to fit a working neural network in 256KB

**This demo will be presented by two people: Yuichi Suzuki (CEO, BothSides Technology) and Yoshifumi Nagano (team member). Yuichi is recovering from upper arm surgery and being discharged from the hospital on February 19th. He can present and explain the technical details but needs a second person to physically handle the demo hardware. We would appreciate two presenter spots.