Trainable Acceleration of Classification Operations via Commonly Available Technology
Whitepaper Video Demonstration
TACOCAT is a hardware-based machine learning accelerator built from ordinary components. Its neural network can be trained to recognize any set of three handwritten characters. When tested with five different three-letter subsets of the alphabet, prediction accuracy test results ranged from 86 to 97%
Many useful software-based neural network models are too large and slow to be implemented on mobile devices. Hardware-based neural network accelerators may soon allow these models to run on phones and tablets without sending data to the cloud.
We were interested in building an experimental hardware-based neural network, but all of the research designs that we found used hardware that is not yet commercially available, such as memristor crossbar arrays.
We wondered if it would be possible to build a demonstration-sized network that would be able to classify datasets containing a small number of different characters.
After investigating the project’s feasibility, we set out to design a demonstration-sized network that can recognize a set of three handwritten characters using commonly-available discrete components.
Classifies any set of three different handwritten characters
Prediction accuracy ranges from 86 to 97%
Built from common CMOS components, without any custom VLSI
The members of Group 31 would like to give special thanks to Dr. Chung Yong Chan, who was instrumental in the group’s success from start to finish. Dr. Chan generously provided support, guidance, and a voice of reason throughout the entire duration of the project, and we will remain grateful for his invaluable mentorship.
