Reference Number: ARL-SRP-00093

Project Description

Spiking Neural Networks (SNN) are the third generation of neural networks and are one potential solution to improving the energy efficiency future AI systems and to putting advanced learning algorithms on low power edge devices.  Due to how they operate it is more appropriate to train SNN models on datasets with dynamically changing data rather than static data, but they have primarily been trained on static datasets.  MNIST and EMNIST are handwritten number and character recognition datasets, respectively that have been central to the early development of machine learning algorithms, but they are static datasets displaying only the final result of a character or number after it has already been fully written.  While stylus handwriting datasets are dynamic as they record the pen motions used to write characters and numbers.  The aim of this project will be to create a synthesized stylus MNIST and EMNIST dataset to explore the testing and training of character classification algorithms. 

Existing stylus datasets can be used to make an initial approximation of the stylus EMNIST dataset, but additionally stylus samples may need to be synthesized using generative models such as (Kotani et al. ECCV 2020) or modern GenAI methods such as diffusion or flow matching methods.  The quality of the reconstruction can be evaluated against the original datasets.  The use of SNN models for the classification task is encouraged, but not required. 

Required skill set: Willingness to learn new things.  Experience with Python or C++/C#
Requested Skill Sets – Experience with machine learning, computer vision, deep neural networks, GenAI, optical character recognition, or basic computational neuroscience is beneficial

Technical Skills

• Artificial Intelligence/Machine Learning
• Computer Science
• Data Science
• Electrical Engineering
• Natural Language Processing
• Software Development