Tsetlin Machine integer number recognitionFollow project
|1||STMicroelectronics Discovery MCU Development Kit STM32F746G-DISCO||882-0278|
Tsetlin machine is a new form of machine learning algorithm which moves away from traditional arithmetic design in favour of a faster and more power-efficient logic-based design. It has been shown to be up to 6 times more power-efficient than traditional neural network designs. Tsetlin machine uses many learning automaton to determine which inputs to a problem are useful in determining the solution and which are redundant.
The project involved implementing the algorithm on an embedded system and recognising handwritten characters in real-time. In a software implementation accuracies of over 98% were achieved. Different methods of increasing power efficiency and reducing memory footprint were also tested by using different methods of pre-processing. This resulted in a 4-time reduction in memory footprint and an almost equivalent increase in power efficiency and processing time while keeping an almost equivalent accuracy. Below is an example of what the Tsetlin machine believes is a zero. The colour of the square represents the importance of a pixel in determining the answer, with dark blue being the most important, and dark red being the least important.
Using the STM board linked in the parts list integer recognition was implemented. You can see the correct classification of a three.
This implementation of the algorithm is just the beginning; it can be easily adapted to other pieces of data. Using a breast cancer dataset the algorithm achieved over 97% accuracy and the improved interpretability from using a logic-based design makes the algorithm very useful in high stakes domains such as healthcare.