analysis of spectro-temporal receptive fields in an auditory neural network madhav nandipati

Analysis of spectro-Analysis of spectro-temporal receptive temporal receptive fields in an auditory fields in an auditory

neural networkneural networkMadhav NandipatiMadhav Nandipati

PurposePurpose Compare computer model to Compare computer model to

biological systembiological system Evaluate utility of receptive fieldsEvaluate utility of receptive fields

BackgroundBackground STRF describes properties in STRF describes properties in

frequency and time domainsfrequency and time domains Have been studied in many animals Have been studied in many animals

but not adequately in a computer but not adequately in a computer modelmodel

Neural networkNeural network

Training setTraining set Different neurons trained on simple Different neurons trained on simple

soundssounds

0 Time (ms) 200 0 Time (ms) 2000 Time (ms) 200

Freq

uenc

y (k

Hz)

5

Freq

uenc

y (k

Hz)

5

Freq

uenc

y (k

Hz)

5

Upsweeps Pure tones Downsweeps

Training procedureTraining procedure Oja’s rule – modified Hebbian Oja’s rule – modified Hebbian

learninglearning Update rule:Update rule:

Some particulars:Some particulars: Training set of 276 stimuliTraining set of 276 stimuli 600 iterations through each stimulus600 iterations through each stimulus

Results – weightsResults – weights

210 Frequency (Hz) 14501320 22601920 3075

2650 38503460 44004270 5510

Generation of STRFsGeneration of STRFs Create ripple stimuli by equation:Create ripple stimuli by equation:

Generation of STRFs Generation of STRFs cont.cont.

Feed ripple into model and get Feed ripple into model and get outputoutput

Fit curve to outputFit curve to output

Generation of STRFs Generation of STRFs cont.cont.

Magnitude and phase shift become Magnitude and phase shift become one pixel of transfer function:one pixel of transfer function:

Generation of STRFs Generation of STRFs cont.cont.

2D inverse Fourier transform to 2D inverse Fourier transform to obtain STRFobtain STRF

Pad matrix with zeros to get Pad matrix with zeros to get smoother imagesmoother image

Use similar color schemes as Use similar color schemes as literatureliterature

Results - STRFsResults - STRFs

Comparison to tuning Comparison to tuning curvescurves

Depicts response intensity to Depicts response intensity to different freqs.different freqs.

Peak of curve = best frequency (BF)Peak of curve = best frequency (BF)

Comparison of BFsComparison of BFs Peak of tuning curve compared to Peak of tuning curve compared to

max value of STRFmax value of STRF

Temporal componentTemporal component Line of best fit around max areaLine of best fit around max area Different slopes means different Different slopes means different

temporal characteristicstemporal characteristics

Temporal component Temporal component cont.cont.

Discussion & ConclusionDiscussion & Conclusion Real STRFs look similar to Real STRFs look similar to

computational STRFscomputational STRFs Means underlying processing is also Means underlying processing is also

similarsimilar