 Hello, everyone. My name is Paul Mitchell. I'm a PhD student working under Laurel Carney at the University of Rochester. And I'm here to talk about our computational model of fast-spectro-temporal chirps sensitivity in the inferior colliculus. So our lab deals with the inferior colliculus, or IC, the primary nucleus of the auditory mid-grain. The IC receives diverse input from most brains to nuclei, as well as descending E for an input. And these diverse inputs make the IC sensitive to a range of complex sound features, including pitch, binaural cues, amplitude modulation, or AM, and frig modulation, or FM. This is particularly interesting for speech coding. As you can see, if you zoom in on the spectrogram for the E vowel, speech is rich with these complex features, including fast FM. So this is relevant because a recent recording suggests that IC neurons are highly sensitive to the direction and velocity of fast FM chirps. In our lab, we make recordings, extra cellular recordings with rapid ICC neurons responding to different sound stimuli. And one of these stimuli are the shorter harmonic complexes. So shorter stimuli are unique harmonic complexes with flat magnitude spectra and a specific space spectrum, such that there is a linear frequency chirp that goes from the highest harmonic to the fundamental, or vice versa, all contained within a pitch period. That makes these chirps much faster than slower chirps that are more commonly studied, such as formant transitions. Flipping the phase function changes the chirp direction and changing the fundamental results in a change of chirp velocity. So rabbit neurons are sensitive and often selective for chirp direction and velocity. The dot rasters in the bottom show responses from two separate neurons, where the x-axis here is time. The y-axis within a row are differing repetitions of the same stimulus, and the different rows are different directions of shorter chirp. So neuron one here is selective for the negative direction of shorter chirp, where neuron two is selective for the positive direction. This selectivity is even more striking if you consider that differing directions of shorter chirp are simply time-reversed versions of one another. So we've observed the sensitivity to fast chirps in rabbits and gerbils, which suggests that this may be a property of the general mammalian IC that may extend to humans as well. So we know that chirps and especially vowels contain lots of these, sorry, we know that speech and especially vowels contain lots of chirping features that these neurons may be equally sensitive to. So the question we want to ask is, how does this novel chirp sensitivity factor into speech coding in the IC? So to answer this question, we need a model. So our modeling strategy goes back to some foundational work done by Siebert in the 1960s, who was the first to represent auditory nerve rates as non-homogeneous plus-on processes. This importantly allows the estimation of psychophysical thresholds. Later, chirps and first extended these calculations to Brainstem nuclei to show that you could treat the IC neuron as a coincidence detector of multiple excitatory and inhibitory inputs and as long as those inputs were all non-homogeneous plus-on processes, your output would also be a non-homogeneous plus-on process. So using the strategy, but in the IC, we can propose models with many different configurations of excitatory and inhibitory inputs from multiple frequencies. So to explain why that multi-frequency aspect is so important, we need to examine some of the proposed mechanisms for chirps like to be in the IC. So the most common mechanism is rooted in an asymmetrical inhibition across frequency. So in this figure, you can see that conceptually, if you have an inhibitory side region and a frequency sweep, a sweep in one direction would result in a staggered excitation inhibition and a spike, whereas a sweep in the other direction would result in a coinciding excitation inhibition and no spike. This asymmetry could be achieved with flanking inhibitions, which are stronger on one side or flanking inhibitions, which vary in relative timing to the excitation. Another interesting mechanism which could be used is the cochlear traveling wave latency where lower frequencies arrive at the IC later than higher frequencies. You can imagine a cell that responds when there is a sweep that results in a multi-frequency inputs lining up and compensating for this latency and not responding otherwise. So based on these proposed mechanisms, our model takes input from three frequencies, a excitatory center frequency and two flanking inhibitory frequencies. So I'm going to walk through the model in detail now. We utilize rate inputs from the Geilani et al. 2014 Auditory Nerve Model. The IC stage of the model has many adjustable parameters, which are the excitatory frequency, the two inhibitory frequencies, one low and one high, the strength constants for the low and the high inhibitions, inhibitory delays relative to the excitation. This can be negative meaning that it arrives before the excitation. And lastly, a coincidence detection time window in Delta. So we perform parameter optimization using real IC neuron recordings as the example data. And where the objective function is the mean squared error of the average rate responses between the example data and the model. So this model configuration allows us to fit parameters to many kinds of chirp sensitivity that we observe in physiology. And it's also intended to preserve known IC properties such as frequency response maps and AM tuning, modulates transfer functions. So I'm going to show some preliminary results. So using this model, we can propose mechanisms for chirp sensitivity. On the left, you can see more dot rasters. This is a neuron that is selected for negative Schroeder, but it turns off at higher velocities. So using this neuron as an example data, the model can guess at what some of these parameters would be and different iterations will result in differing guesses. So you can see bolded here, some of these model parameters are very different between these two final model responses, but they result in pretty similar sensitivities. Some iterations will be better at doing some aspects of the physiology than others. For instance, this iteration here retains the ghost of a response at the 50 Hertz positive Schroeder condition, whereas the other one is better at turning off at these higher velocities. So we can simulate cells which are selected to positive Schroeder chirps as well. Here we see a neuron that is selected to positive Schroeder chirps at lower fundamentals, but then at higher fundamentals, it just sort of responds to everything. So the model seems to be able to mimic the selectivity fairly well at lower velocities, but isn't very good at capturing the higher F knots where the cell just responds equally well. This is a work in progress and we need to continue to explore the parameter space and possibly add new inputs to the model, just explore different model configurations. Ultimately, we want a model that can accurately match all of it given neurons responses are very stimulating. When we record in the rabbit, we present a wide variety of characterizing stimuli. So we want the model to reproduce the same frequency response maps and modulation transfer functions that we see in the physiology. So looking at the model's best guess for some of these responses, we have the actual neural response and the model's output. You can see some of the key features are there. For instance, the response map has the same character as a frequency. The modulation transfer functions have the same best modulation frequency, but work remains to be done. For instance, you can see that the spontaneous rates don't match. There is an issue with the modulation transfer functions where the model wants to just suppress everything at higher levels. This is a work in progress. So where would we like the direction of this model to go? Ultimately, like I mentioned before, we want to be able to fit the model to multiple stimulus types, but also simultaneous stimulus types. So this includes Schroder-Turbs response maps, modulation transfer functions, additional stimuli that we do present in the physiology. To do that, we'll likely need to introduce additional complexities or stages to the model, just continue to explore the parameter space. And finally, you really want to ask the question, is chirp sensitivity critical to IC speech responses? This will involve assessing the ability of our chirp model to replicate IC neuron bowel responses and comparing these results to a model without chirp sensitivity. So thank you for your time and I'd be happy to take any questions now.