Judges’ Queries and Presenter’s Replies

  • May 20, 2013 | 07:11 p.m.

    What specific hypotheses are you trying to test in these experiments and what conclusions can you draw from your experiments?

  • Icon for: Isaac Carruthers

    Isaac Carruthers

    May 23, 2013 | 01:14 a.m.

    The main hypothesis displayed here was that we could accurately predict the responses of neurons in the auditory cortex based on the amplitude and frequency modulations of the stimuli. Although not shown in this poster for reasons of simplicity, we compared the performance of our predictive model with the most common similar model, which takes sound-spectrum rather than modulations as input. We found that our model significantly outperformed the common model, indicating that frequency and amplitude modulations may yield a better representation of the auditory features coded by auditory cortex.

  • May 21, 2013 | 05:53 p.m.

    What are the practical benefits of being able to predict cell responses to different subsets of vocalization in rats? Will results from the rat’s auditory cortex give insight into the behavior of the auditory cortex in humans? Why have you used rats as opposed to other animals such as cats? Have similar studies been conducted on Humans?

  • Icon for: Isaac Carruthers

    Isaac Carruthers

    May 21, 2013 | 09:48 p.m.

    Due to the invasive nature of neural recordings it is unusual to perform such studies in humans. Some studies will use human subjects who have had electrodes inserted for other reasons, such as for the treatment of epilepsy. However, these studies generally have very little control over the location of the electrodes, and are often unable to isolate individual cells.

    We used rats rather than other non-human animals because rat vocalizations are rich and complex in terms of modulation structure and syntax, and yet are quite simple to parametrize. Rats are also very simple to house and study when compared to other animals with comparable vocalization repertoire, such as primates.

    As to the practical benefits, the ability to identify auditory features that are represented in higher brain areas allows us to better understand the computations performed by neural circuits of the auditory system. This in turn allows us to interact more effectively with the auditory system, whether by developing more efficient audio codecs, by creating more effective hearing aids and cochlear implants, or by inventing new treatments for tinnitus.

  • May 21, 2013 | 08:32 p.m.

    In your plot comparing observed response with Amp-FN LN prediction, is this prediction of the same data to which the model had been fit, or is it out-of-sample prediction? How does the predictive ability of your model compare to that of other statistical models for this type of data?

  • Icon for: Isaac Carruthers

    Isaac Carruthers

    May 23, 2013 | 01:07 a.m.

    The plot shows predictions for portions of the the stimulus not included in the model fitting.

    We compared our model to a popular model based on the audio spectrum of the stimulus (a model which, being based on two-dimensional input rather than one-dimensional, had far more fitting parameters than our own), and found that our model performed significantly better for most neurons tested.

  • Further posting is closed as the competition has ended.

Presentation Discussion

  • Icon for: Joni Falk

    Joni Falk

    May 23, 2013 | 04:11 p.m.

    Hi Isaac, Thanks for this. Found the video very easy to follow. To what degree do you think that what you learn from rats can be applied to humans? I’d be interested to hear more about what you feel are the practical implications of this work, perhaps further down the road.

    Also, and unrelated I thought you might be interested in the work presented at
    http://posterhall.org/igert2013/posters/404 as you both seem to have some overlapping areas of expertise.

  • Icon for: Isaac Carruthers

    Isaac Carruthers

    May 23, 2013 | 10:39 p.m.

    Hi Joni, glad you liked the video.

    We can never say for sure how much of what we learn in the rat will apply directly to humans, but whatever we do learn will give us more ideas for where to look. For instance, engineers designing new hearing-aids and cochlear implants are always looking for new ways to filter incoming sounds. New filtering methods can improve the ability of the listener to pick out elements of complex auditory environments, such as listening to someone’s voice in a noisy room. By identifying features that are useful for identifying these auditory objects in the rat’s brain, we identify new potential features for a hearing-aid to amplify.

    Thanks for linking to Mr. Ocker’s presentation; his work is definitely relevant to my own (and more so to that of other members of my lab, who are working on a rat model of tinnitus).

  • Further posting is closed as the competition has ended.

Icon for: Isaac Carruthers


University of Pennsylvania
Years in Grad School: 3

Identifying the Origins of Auditory Objects: Application of Generalized Linear-Nonlinear Models and Noise-Correlation Analysis to the Neural Representation of Rat Vocalizations in Rat Primary Auditory Cortex

One of the central tasks of the mammalian auditory system is to represent information about acoustic communicative signals, such as vocalizations. However, the neuronal computations underlying vocalization encoding in the central auditory system are poorly understood. To learn how the rat auditory cortex encodes information about con-specific vocalizations, we presented a library of natural and temporally transformed ultra-sonic vocalizations (USVs) to awake rats, while recording neural activity in the primary auditory cortex (A1) using chronically implanted multi-electrode probes. Many neurons reliably and selectively responded to USVs. The response strength to USVs correlated strongly with the response strength to frequency modulated sweeps and the FM rate tuning index, suggesting that related mechanisms generate responses to USVs as to FM sweeps. The response strength further correlated with the neuron’s best frequency, with the strongest responses produced by neurons whose best frequency was in the ultra-sonic frequency range. For responses of each neuron to each stimulus group, we fitted a novel predictive model: a reduced generalized linear-non-linear model (GLNM) that takes the frequency modulation and single-tone amplitude as the only two input parameters. The GLNM accurately predicted neuronal responses to previously unheard USVs, and its prediction accuracy was higher than for an analogous spectrogram-based linear non-linear model. The response strength of neurons and the model prediction accuracy were higher for original, rather than temporally transformed vocalizations. These results indicate that A1 processes original USVs differentially than transformed USVs, indicating preference for temporal statistics of the original vocalizations.