 Today, I would like to introduce the features of machine learning audiometry and some of its advantages over conventional audiometry. Before introducing the audiogram itself, however, I will first review some basic principles of perceptual models and how they are typically estimated. Here we have a basic psychometric function in green, mapping stimulus properties to task performance. In this case, task performance is probability of success. This psychometric function cannot be observed only estimated. Additionally, the values used to fit the model, shown as black circles, cannot be observed and themselves must be estimated. Each circle therefore represents a proportion of success for a particular value of each stimulus property. All of these probabilistic estimates combine to require large amounts of data for estimating the entire psychometric function. This is referred to as the two-parameter model. Alpha is the threshold, beta is the psychometric spread or slope, and the entire function is generally referred to as psi. Three and four-parameter models are sometimes used depending on the application. The result of this type of modeling is fully predictive models capable of providing a probability of any particular outcome, mapping any particular stimulus property to success. Empirical studies have shown that 200 to 500 trials are required to obtain reasonable estimates of these models. This requirement is prohibitive for almost all clinical or even research applications. Fair case methods are a simpler form of threshold estimation. If a subject completes any particular task successfully, the next task becomes easier. If however, a subject fails to complete the task, the next task becomes easier. I don't think that's right. Two, one, go. Today, I would like to introduce the features of machine learning audiometry and some of its advantages over conventional audiometry. Before introducing the audiogram test itself, however, I will first review some basic principles of perceptual models and how they are typically estimated. Here we have a basic psychometric function in green, mapping stimulus property along the horizontal axis to task performance along the vertical axis. In this case, task performance is probability of success. This psychometric function cannot be observed only estimated. Additionally, the values used to fit the model, shown as black circles, also cannot be observed and must be estimated. Each circle represents the proportion of success for a particular value of each stimulus property. All of these probabilistic estimates combine to require large amounts of data for estimating the psychometric function. This form, the psychometric function psi, is referred to as the two-parameter model. Alpha is the threshold, and beta is the psychometric spread or slope of this function. Three and four-parameter models are sometimes used depending on the application. The results of this type of modeling are fully predictive models capable of providing a probability of any particular outcome as a function of any particular stimulus property. Empirical studies have shown that 200 to 500 trials are required to obtain reasonable estimates of these models. This requirement is unfortunately prohibitive for almost all clinical or even research applications. Staircase methods are a simpler form of threshold-only estimation that are quite efficient. If a subject completes any particular task in the sequence successfully, the next task becomes harder. If a subject fails to complete any particular task, however, the next task becomes easier. This simple one-step memory sequence provides a robust estimation procedure. A staircase can be thought of as a one-parameter model. It delivers only the threshold, however. The threshold is not a fully predictive model. In other words, not a generative or probabilistic model, complete model. It only reveals the probability of correct task performance at one input value. The threshold spread or slope are not estimated, but assumptions about those values are used to make the staircase more efficient. The audiogram has two independent variables, unlike the one independent variable psychometric functions we've just seen. These variables are intensity or sound level and frequency. The intensity variable follows all the rules of psychometric function estimation with stimulus properties that we just saw. Frequency however does not. The most common approach to estimating audiograms is not to build more complex models that can accommodate those two types of input variables. Instead, multiple staircase procedures are performed at different frequencies in sequence. The result is not a fully predictive model, but a series of discrete thresholds. While relatively efficient and robust, a major limitation of this procedure is that information cannot be shared effectively. As a result, it is challenging to improve the current methodology further. Here is an example of a conventional audiogram conducted by an audiologist. Note that audiograms are usually depicted with inverted vertical axes. We can see this audiogram take shape with each tone delivery. The series of staircases is obviously suboptimal. Even after we have collected a few responses at 1 kHz, for example, we continue to sample there even though it's clear we know more about the audiogram structure at 1 kHz than at any other frequency. This estimation method provides a relatively incomplete picture of this ear's hearing. Sampling only at a few discrete frequencies. It does not yield a fully predictive model. And possible extensions of this method are limited. The machine learning audiogram, or MLag, is a truly multidimensional estimation method. It uses Gaussian process classification in a Bayesian active learning framework to optimize where to sample over the full input domain, meaning tone intensity and tone frequency. You can see the difference between data acquisition of this model estimation procedure and the last one as so. Despite varying both sound frequency and level simultaneously, threshold estimates converge to values close to those of conventional audiometry in the purple line. This is a Bayesian method, but we have used an uninformative prior belief here to provide a lower bound on efficiency while still allowing estimator accuracy to be quantified. Here we see a few frames of the movie that you just viewed. MLag returns a fully predictive model in less time than conventional audiometry returns a few thresholds only. We can visualize this here as a continuous frequency audiogram that forms along with psychometric spread estimates. Those are shown at 1 kHz in the insets of these panels. This method provides more information than conventional audiometry and takes less time on its own, but the real advantage of this procedure is the new capabilities that MLag makes available for the first time. We will discuss one of several audiometric extensions that we have developed for the rest of the talk. Thresholds with very different thresholds in their two ears can experience cross hearing where tones probing threshold in one ear can be a bone conduction rise above detection threshold for the other ear. In this example, we have audiograms of two ears of one person, and the triangular area indicates tones at risk of cross hearing because their intensity is high relative to the contralateral ear, the contralateral ear's thresholds. Masking noise delivered in the contralateral ear ensures that loud tones do not cross over through bone conduction to activate that ear and then result in a false positive. Conventionally, correct masking is a cumbersome and time-consuming procedure to implement correctly. Dynamically masked MLag, however, learns the right amount of masking in real time. That's what's depicted here. The results of dynamically masked MLag provide similar threshold estimates between conventional and machine learning procedures at considerable time savings. In this study were individuals with a range of hearing profiles, administered both conventional audiometry and dynamically masked MLag. Small differences in thresholds indicate similarity of estimates between the two methods. Dynamically masked MLags also require no longer to complete than masked MLags and in some individuals substantially less time than the conventional audiogram requires to execute. Most individuals in the population have fairly symmetric hearing across both ears, even if they have hearing loss. The machine learning audiogram so quickly converges to threshold that the masking noise is almost always below the contralateral detection threshold for a large proportion of the population. As a result, these individuals do not even know they are experiencing a masked audiogram. For example, symmetric normal hearing individuals and symmetric hearing loss individuals, very few of the contralateral masking noise stimuli are detected by these individuals. Individuals with asymmetric hearing detect the masking noise but have been shown to be able to disregard it in order to indicate detection of the probe tones. In this case, every audiogram has become a masked audiogram with no problems achieving accurate estimates in normal hearing individuals and those with symmetric hearing loss not normally requiring masking and also is able to perform accurate estimates in individuals who do require masking. The end result is an audiogram estimation procedure that can deliver hearing thresholds in both ears at once, regardless of the amount of hearing asymmetry in less time than is required for a conventional audiogram. Here, the algorithm optimizes over tone frequency, intensity, and ear simultaneously. The result is that the estimates for both ears can be obtained in about the same time as is required to estimate one ear. MLAG in all its forms can be delivered remotely or in a clinic with clinician supervision. And with that, I would like to thank the large number of collaborators and trainees who have contributed to this work and also to our funding sources. Thank you very much and I'm happy to take questions.