Machine Learning on the Edge
Research into Internet of things (IoT) began January 21st, 2021, as part of the subaward of Kansas NASA Space Grant 1, regarding machine learning on the edge. The research is multifaceted and pertains microcontrollers capable of running neural networks, and deep learning through the assistance of real-time operating systems (RTOS). The study included a multistep process which involved an understanding of a microcontroller which is capable of RTOS and neural network functionality, such as the MSP432 microcontroller. An understanding of the advantages and limitations of microcontroller regarding on the edge applications is also necessary. The premise of edge machine learning pertains to the necessity of being able to train neural networks using powerful computing. After a neural network is trained, the neural network may be implemented on smaller, less powerful devices. This is beneficial where high latency and restricted bandwidth may impede access to cloud computing. Using the software Excel, PyCharm, and Google Colab, the research has progressed into a deeper understanding of forward propagation, weights, biases, multi-hidden layer networks, and activation functions. Small feedforward neural networks have been developed in Excel in order to better understand the mathematics of training neural networks. This information can later be used in larger neural networks. The research is currently moving into a deeper understanding of backpropagation and training of neural networks which includes an understanding of gradient descent and cost functions. These processes will likely include the application of developed libraries.
Kincheloe, Brandon, "Machine Learning on the Edge" (2021). Video Presentations. 10.