Utilising reinforcement learning to develop strategies for driving auditory neural implants

Lee, Geoffrey W; Zambetta, Fabio; Li, Xiaodong; Paolini, Antonio G.

doi:10.1088/1741-2560/13/4/046027

Cited by 1 publication

(1 citation statement)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A possible option would be to train a so-called Actor-Critic reinforcement learning algorithm, in which an ' Agent' learns how to control the 'Environment' (Lee et al, 2016;Haarnoja et al, 2018;Sutton and Barto, 2018). In essence, the Environment is a computational model of the CI-user's auditory system, which contains all the knowledge constructed from the electrophysiology, anatomy, psychophysics, and direct stimulation results from all users.…”

Section: Machine Learningmentioning

confidence: 99%

Towards personalized and optimized fitting of cochlear implants

Opstal

Noordanus

2023

Front. Neurosci.

View full text Add to dashboard Cite

A cochlear implant (CI) is a neurotechnological device that restores total sensorineural hearing loss. It contains a sophisticated speech processor that analyzes and transforms the acoustic input. It distributes its time-enveloped spectral content to the auditory nerve as electrical pulsed stimulation trains of selected frequency channels on a multi-contact electrode that is surgically inserted in the cochlear duct. This remarkable brain interface enables the deaf to regain hearing and understand speech. However, tuning of the large (>50) number of parameters of the speech processor, so-called “device fitting,” is a tedious and complex process, which is mainly carried out in the clinic through ‘one-size-fits-all’ procedures. Current fitting typically relies on limited and often subjective data that must be collected in limited time. Despite the success of the CI as a hearing-restoration device, variability in speech-recognition scores among users is still very large, and mostly unexplained. The major factors that underly this variability incorporate three levels: (i) variability in auditory-system malfunction of CI-users, (ii) variability in the selectivity of electrode-to-auditory nerve (EL-AN) activation, and (iii) lack of objective perceptual measures to optimize the fitting. We argue that variability in speech recognition can only be alleviated by using objective patient-specific data for an individualized fitting procedure, which incorporates knowledge from all three levels. In this paper, we propose a series of experiments, aimed at collecting a large amount of objective (i.e., quantitative, reproducible, and reliable) data that characterize the three processing levels of the user’s auditory system. Machine-learning algorithms that process these data will eventually enable the clinician to derive reliable and personalized characteristics of the user’s auditory system, the quality of EL-AN signal transfer, and predictions of the perceptual effects of changes in the current fitting.

show abstract

Section: Machine Learningmentioning

confidence: 99%