Active Learning in the auditory domain
A round table with Bert de Vries, Josef Schlittenlacher and Dennis Barbour. Moderator Jan-Willem Wasmann
Audiovisual version on Youtube:
Josef Schlittenlacher is a lecturer now in Manchester. Before that, he was a postdoc in Cambridge and he did a PhD in psychology in Germany. He studied electrical engineering. His background in machine learning audiology started in Cambridge on a grant for Bayesian active learning applied to what Brian Moore and Richard Turner do.
Dennis Barbour is a professor of biomedical engineering at Washington University in St Louis. His educational background is electrical engineering and biomedical engineering, and he also has an MD. We became interested in perceptual training and trying to induce therapeutic changes in brain function. The ability to achieve perceptual training effects that are persistent and transfer in any complex domain depends on the lab and the preparation and it’s not highly reproducible. We started working on models and testing regimes that can bridge those gaps and the amount of data required using classical methods is just prohibitive. So my goal nowadays is to join the audiogram to other relevant tests into a unified or mostly unified testing procedure that integrates the data across all the tests to decide what the optimal sequence of queries is going to be (based on Bayesian active learning).
Bert de Vries works as a professor in electrical engineering signal processing, at the Technical University in Eindhoven and still has a small affiliation with GNresound. His interest is basically automating the design of algorithms using the same strategies as the brain. More generally: his lab at TU Eindhoven focuses on transferring a theory about how brains compute, namely the Free Energy Principle (by Karl Friston), to engineering applications. Ultimately, the aim is to largely automate the hearing aid design process solely by in-situ interactions between a HA patient and an FEP-agent.
What motivated you to start using active learning applied to audiometry?
For Bert, the motivation to start with active learning for determining audiograms was the lack of professional audiological support for fitting hearing aids in remote locations or for OTC hearing aids. Josef adds that our methods in scientific experiments and clinical tests are pretty rigid. Why should I test several times at 30 dB and 35 dB, or 1 and 2 kHz, but never between? The heuristic rules that traditional methods use, make them slower than they need to be. We have the computational power to always choose the best option. In a more general see the three participants see many applications of automating the data selection process for machine learning applications.
What is unique about your approach?
Josef explains how Bayesian Active Learning was introduced to psychophysics by Kontsevich & Tyler in 1999! We basically use the same approach, calculate some probabilities and choose the most informative parameters. It needed their genius to overcome the limitations of computation power at the time. It took twelve more years until Houlsby et al. (2011) introduced Gaussian Processes to this paradigm. Soon after, several labs used it for audiometric tests. We use the approach for several audiometric tests. In addition to the audiogram, we have developed tests for auditory filters, equal-loudness contours, and dead regions. We try to optimize every test by tweaking it further. For example in the audiogram, we can get more information from the user if we don’t just ask for a Yes/No response but let them count tones. At the moment we are working on remote tests – certainly not unique after the lockdowns. What many users don’t understand is the need for calibration in an audiogram. An audiogram on an unknown phone simply doesn’t work. We are working on tests that can characterize a hearing loss when the hardware and audio settings are not known.
Bert adds that in their approach they use a prior distribution for audiograms, that has been trained on a large database of common audiograms. This leads to fewer trials when taking an audiogram.
Dennis believes all three of them have sought to generalize the estimation tools used to apply signal detection theory to individuals. Familiar psychophysical tasks are coupled with advanced machine learning algorithms to result in individual perceptual models. Flexible, nonparametric models are trained with behavioral data from task trials determined in real-time using Bayesian active learning. Usually, the more flexible the model the more data are required to fit it. We have collectively shown in a variety of applications that this combination allows complex models to be learned in practical amounts of time.
Dennis’s focus has been documenting that for fitting more complex perceptual models, the computational task is equivalent to performing a high-dimensional classification. In other words, he seeks to define the boundary between successful and unsuccessful task performance. Tremendous effort in the machine learning field has gone into designing classifiers useful for this task. Dennis has adopted and extensively evaluated a probabilistic classifier, the Gaussian Process (GP), but other classifiers could be useful for this effort, as well.
Further, he has developed the concept of Bayesian active model selection, in which models of arbitrary complexity can be pre-defined and new data from individuals used to optimally accumulate evidence in favor of one model or another for that person. This method allows theoretical, population and individual knowledge to be combined to directly address a target question.
What are potential applications for researchers?
All three Machine Learning Audiometry papers (Barbour et al., 2019; Cox & de Vries, 2021; Schlittenlacher et al., 2018) approach audiometry from a Bayesian reasoning viewpoint. The Bayesian approach is not only practically useful but also fundamentally very sound. The hearing research community would benefit from studying Bayesian modeling.
There are more tests that could benefit from Bayesian active learning. Take for example the auditory-filter tests. They took hours if you wanted to do them on several frequencies. In the Bayesian active-learning version that is reduced to thirty minutes for a range of three octaves. That means other audiology researchers can include it in their test batteries and get more information when they for example want to know more about cochlear synaptopathy or the link between hearing loss and dementia.In general, Bayesian Active Learning has huge potential for research. Any experiment can be sped up.
The most significant new capability for researchers is to bring individualized behavioral modeling into the realm of feasibility because the data requirements to train such models will be reduced over current methods. More complex perceptual constructs can be hypothesized, formulated and tested. In particular, constructs that manifest variability across heterogeneous populations would benefit from individual modeling, as opposed to the more common population modeling.
What are potential applications (benefits) for clinicians?
Bert replies that in as far as the (Bayesian audiometry) papers would be realized in products, these products would be more efficient in collecting and processing trial data, relative to non-Bayesian-based products.
Josef thinks that clinicians do not need to monitor the testing closely and can focus on interpreting the results. Clinicians need to spend less time getting more information from various tests that the patients can do on their own. That leads to better hearing-aid fittings at a lower cost.
This aligns with Dennis’s opinion that practical clinical tests must be conducted within reasonable amounts of time. That limitation naturally constrains the complexity of the diagnostic hypotheses that can be evaluated with behavioral data. By making more complex models testable in a reasonable time, clinicians’ ability to more accurately diagnose patients and designate appropriate treatment paths will be improved.
Quotes from the interview
Dennis: ‘No Bayesianists are born, they are all converted’ (origin unknown)
Josef: ‘The audiogram is the ideal testbed for Bayesian active learning.’
Bert’s favorite quote: ‘Everything is the way it is because it got that way’ (D’Arcy Wentworth Thompson, 1860–1948) The later quote reflects on the idea that everything evolved to where it is now. It’s not a quote from the Free Energy Principle but it has everything to do with it. The hearing system evolved to where it is now. To design proper hearing aid algorithms, we should not focus on the best algorithm but rather on an adaptation process that converges to better algorithms than before.
Further reading and exploring
Audiogram estimation using Bayesian active learning:
Barbour, D. L., Howard, R. T., Song, X. D., Metzger, N., Sukesan, K. A., DiLorenzo, J. C., Snyder, B. R. D., Chen, J. Y., Degen, E. A., Buchbinder, J. M., & Heisey, K. L. (2019). Online Machine Learning Audiometry. Ear and Hearing, 40(4), 918–926. https://doi.org/10.1097/AUD.0000000000000669
Cox, M., & de Vries, B. (2021). Bayesian Pure-Tone Audiometry Through Active Learning Under Informed Priors. Frontiers in Digital Health, 3, 107. https://doi.org/10.3389/fdgth.2021.723348
Houlsby, N., Huszár, F., Ghahramani, Z., & Lengyel, M. (2011). Bayesian active learning for classification and preference learning. ArXiv Preprint ArXiv:1112.5745.
Kontsevich, L. L., & Tyler, C. W. (1999). Bayesian adaptive estimation of psychometric slope and threshold. Vision Res, 39(16), 2729–2737.
Schlittenlacher, J., Turner, R. E., & Moore, B. C. J. (2018). Audiogram estimation using Bayesian active learning. The Journal of the Acoustical Society of America, 144(1), 421–430. https://doi.org/10.1121/1.5047436
Wasmann, J.-W., Pragt, L., Eikelboom, R., & Swanepoel, D. W. (2021). Digital approaches to automated and machine learning assessments of hearing: A scoping review. Journal of Medical Internet Research. https://doi.org/10.2196/32581