Trainee Collaboration Projects

Proponent: Niklas Smedemark-Margulies

Description: Individuals with severe speech and motor impairments can benefit from augmentative and alternative communication systems. BCI designs which measure sequential stimulus-evoked brain responses through EEG can be effective in inferring human intent in such settings. Matrix Speller and RSVP Keyboard are well-established examples of such designs. In these systems, we begin from a language model prior, and gradually accumulate evidence from the user’s responses. We can use recursive Bayesian updates to maintain an estimate of the posterior distribution of the user’s intended symbol. Baseline methods for this problem use a generative modeling approach to assess the likelihood of possible symbols given EEG evidence. The drawback of this approach is the difficulty of modeling high-dimensional EEG data, which necessitates decomposing the data into short windows of time and assuming independence between adjacent time windows. Therefore, we propose to perform this recursive Bayesian update by using a discriminative neural network applied to entire sequences of EEG data. In this way, we avoid making any assumptions of independence between consecutive time points, and make inferences from the full original data dimension.

Research Questions: We seek to evaluate whether a discriminative modeling approach can outperform a generative modeling approach for inferring user intent from EEG in a recursive Bayesian inference setting.

Needed Skills: Clinician, Neural Engineer

Proponent: Basak Celik

Description: BCIs based on EEG can be used to infer user intent for various applications, including expressive language generation through letter selection. Typical users of such assistive technology interfaces may have conditions that lead to interference signals generated by muscles or external devices, which superimpose as artifacts on their brain activity captured by EEG. Such artifacts may significantly reduce the accuracy of intent inference from EEG in these systems. Therefore, it is important to correctly assess the quality of the EEG signals, and remove artifacts where possible. The issue of noise and artifacts is traditionally approached by completely separate stages for filtering and inference. Such a static filtering approach may be considered as a point-estimate of the ideal denoising procedure. We propose to replace this point estimate approach with a probabilistic model of the denoised data, such that we can sample the distribution and take a Bayesian approach to subsequent intent inference tasks. There is a large space of model architectures and training objectives to be explored as the denoising model, including Variational Autoencoders and Invertible Neural Networks.

Research Questions: Can a Bayesian approach to artifact and interference filtering improve subsequent intent inference, as compared to a point estimation approach to filtering?

Needed Skills: Machine Learning, Neural Engineer

Proponent: Nibras Abo Alzahab

Description: This project aims to investigate the auditory-stimuli EEG-Biometric by comparing the effect of ear-conducting auditory stimuli and bone-conducting auditory stimuli. The project will be conducted as follows: 1- preparing the experimental setup 2- collecting ear-conducting auditory stimuli data and bone-conducting auditory stimuli. 3- Building a machine learning model (or deep learning) to classify the data according to the subject (Target). 4- Compare the performance of the two systems This project will open the door for future innovations that are based on advanced biometrics and require authentication.

Research Questions: Is bone-conducting auditory stimuli as effective as ear-conducting auditory-stimuli in EEG-Biometric systems?

Needed Skills: Machine Learning, Neural Engineer