Towards a Cognitive Computational Neuroscience of Auditory Phantom Perceptions

TL;DR

This paper advocates for building biologically plausible computational models to understand tinnitus, emphasizing the importance of hypothesis-driven modeling over purely data-driven approaches, and proposes a new research agenda integrating AI, psychology, and neuroscience.

Contribution

It introduces a novel research framework that combines computational modeling with experimental testing to mechanistically understand tinnitus, moving beyond traditional data-centric methods.

Findings

Modern neural assessment technologies are rich but insufficient alone for mechanistic insights.

Hypothesis testing must be complemented with predictive computational models.

Progressive increase in biological fidelity of models can enable in silico testing of treatments.

Abstract

In order to gain a mechanistic understanding of how tinnitus emerges in the brain, we must build biologically plausible computational models that mimic both tinnitus development and perception, and test the tentative models with brain and behavioral experiments. With a special focus on tinnitus research, we review recent work at the intersection of artificial intelligence, psychology and neuroscience, indicating a new research agenda that follows the idea that experiments will yield theoretical insight only when employed to test brain-computational models. This view challenges the popular belief, that tinnitus research is primarily data limited, and that producing large, multi-modal, and complex datasets, analyzed with advanced data analysis algorithms, will finally lead to fundamental insights into how tinnitus emerges. However, there is converging evidence that although modern technologies allow assessing neural activity in unprecedentedly rich ways in both, animals and humans, empirical testing one verbally defined hypothesis about tinnitus after another, will never lead to a mechanistic understanding. Instead, hypothesis testing needs to be complemented with the construction of computational models that generate verifiable predictions. We argue, that even though, contemporary artificial intelligence and machine learning approaches largely lack biological plausibility, the models to be constructed will have to draw on concepts from these fields, since they have already proven to do well in modeling brain function. Nevertheless, biological fidelity will have to be increased successively, leading to ever better and fine-grained models, allowing at the end for even testing possible treatment strategies in silico, before application in animal or patient studies.