Earing

Repository for ear models and associated numerical experiments for research purposes.

Fully Object-Oriented Programming-oriented, most important objects (such as an ear, ear components, experiment or speech recognition run) are based on RunObject.m. They inherit from this abstract class, meaning that their main function is to be run. Their name property is used to concatenate in a single string all their parameters. For example, the name of an ear contains the names of its outer middle, of its auditory nerve, ...

---

Cochlear Models

Sumner2003

• Implementation of 'A nonlinear filter-bank model of the guinea-pig cochlear nerve: Rate responses' (Christian Sumner and Towel P. O'Mard, 2003).
• The sound arrives to the Outer Middle Ear, which moves the stapes. Their velocity affects the liquid in the cochlear, thus, in the Basilar Membrane, the little hair on Inner Hair Cells, the Ihc Cilia, will move back-and-fro, allowing the soma to react and to send ions to their synapses: the An-Ihc Synapses will then send neurotransmitters to the Auditory Nerve.
• All these elements are individually modeled, and their model's files are those highlightened words (found in matlab/models/ear/components/(OuterMiddleEar|BasilarMembrane|...).m).
• File matlab/models/ear/EarSumner2002.m shows how each output is passed on to the next processing stage.
• Note that the spiking behaviour is different from the original paper, because the spike generation mechanism was replaced by a real inhomogeneous Poisson Process. The new method, a proper Poisson process, reduces the spiking activity.

---

Getting Started

Requirements

• Signal Processing toolbox for upsampling waveforms.

Download

Download the repo

git clone [email protected]:allevity/earing.git
cd earing/

Mex files

To speed-up the big calculations, MEX-files are used to compute matrices. They may be available for your system in mex/, otherwise you need to compile them:

cd mex/
mex MAP_AN_forLoop_mex.c MAP_applyRefractoriness_mex.c averageChannels.c \
    spikes2ISI.c MAP_AN_generatePoissonSpikeTrains.c MAP_finalForLoop_mex.c \
    rateSpikeTrain.c subsampleSpikeTrains.c
cd ../

---

 Run the model

The method EarSumner2002.run receives either the path to an mp3 file or the array read from it. You may change the sound level of the ear before running.

ear = EarSumner2002();
ear.db = 79.5;  % to change sound level (default value 80dB)

% Your mp3 file should be small: the temporal scale necessary to run this
% biomechanical model is so fine that matrices easily become huge.
ear.run('/path/to/sound_file.mp3'); % runs the OME, the BM, ...

figure;
ear.an.plot();  % imagesc of the firing probability at each time frame and fiber

The object ear then contains all important variables calculated during the run. Executing ear.clean() removes these variables to reduce the object's size.

If ear.synapse.n_fibers_per_type_per_channel>0 (defaults to 1), spikes are also generated and saved in ear.an.spikes_sparse as a sparse matrix of 0's and 1's. ear.synapse.n_fibers_per_type_per_channel fibers of each type are generated for each ear.bm.best_frequency.

To change the parameters, a structure needs to be provided at ear's initialisation. A useful example that changes a good section of the parameters:

gmax_ca = 7.2e-9;  % gmax_ca and ca_thresh together define the fiber type
ca_thresh = 4.48e-11;
best_frequencies = [250, 500, 1000, 3000, 6000, 12000, 24000];
externalResonanceFilters = [ 0 1 4000 25000; 0 1 4000 25000; 0 1  700 30000;   0 1  700 30000;    0 1  700 30000];
params = ...
 struct(...
    'model', @EarSumner2002, ...
    'db', db, ...  
    'ome', struct(...
        'externalResonanceFilters', externalResonanceFilters), ...
    'bm',  struct(...
        'drnl', struct(...
            'lin', struct(), ...
            'gt', struct(...
                'nl', struct(...
                    'lin', struct())),...
            'lp', struct(...
                'nl', struct(...
                    'lin', struct()))), ...
        'best_frequencies', best_frequencies), ...
    'cilia',  struct(),...
    'synapse',  struct(...
        'n_fibers_per_type_per_channel', 1, ...
        'presynapse', struct(...
            'gmaxca', gmax_ca, ...
            'ca_thresh', ca_thresh)),...
    'an',  struct());
ear = EarSumner2002(params);

The spike trains are submitted to a random refractory period (fix term + random term). If ear.an.refractoriness (defaults to false) was set to true before the run, then another matrix of probabilities of firing is generated and saved in ear.an.prob_firing_refractory. This matrix shows the law of the Poisson process generating the spike trains, which is to say, this each row represents the time histogram of an infinity of spike trains after the random refractory period be applied.

---

Tests

Only tested on MacOS High Sierra 10.13.6

Visualise model output

To run the model on a provided mp3 file, execute the below. The resulting plot (saved in tests/data/plots/) shows the various matrices derived as the information crosses the OME, the BM, ... down to spike generation. The blue lines attempt to link the plots with their biophysical origination, with a zoom on the Organ of Corti (inside the cochlea).

cd tests/code/
visualisation_EarSumner2002()

[1][2]

Compare model versions

• Prerequisite: The function test_EarSumner2002_comparison_MAP_AN_only assumes that MAP_AN_only.m and MAPparamsGP.m are in Matlab's path.

To compare this implementation of the model with the former one, execute the below. Important variables will be checked 2-by-2 and a warning appears if there is a mismatch.

cd tests/code/
test_EarSumner2002_comparison_MAP_AN_only()

Plot from original paper

To reproduce a plot from the paper showing the spiking rate of LSR/MSR/HSR fibers of different best frequencies (250Hz to 24kHz) when reacting to sinusoids at this BF for sound intensity varying between 0 and 100dB, execute the below. Additionally, this compares the effect of previous the filter cascade (blue) with the current one (red).

cd tests/code/
test_EarSumner2002()

---

Experiments:

Speech recognition on neural data

• This part of the code expects HTK http://htk.eng.cam.ac.uk/ to be installed, and various variables to be initialised.
• Set by launching matlab/experiments/speech_recognition/experiment_launcher.m

References:

Plots:

• [1] By Lars Chittka; Axel Brockmann - Perception Space—The Final Frontier, A PLoS Biology Vol. 3, No. 4, e137 doi:10.1371/journal.pbio.0030137 (Fig. 1A/Large version), vectorised by Inductiveload, CC BY 2.5, https://commons.wikimedia.org/w/index.php?curid=5957984
• [2] Henry Gray (1918) Anatomy of the Human Body, https://en.wikipedia.org/wiki/File:Gray931.png

External code sources:

• Hash function (https://raw.githubusercontent.com/FNNDSC/matlab/master/misc/hash.m)
• HTK functions: Mark Hasegawa-Johnson