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

Analysis of spectrotemporal receptive fields in
an auditory neural network
Madhav Nandipati
Purpose


Compare computer model to biological system
Evaluate utility of receptive fields
Background


STRF describes properties in frequency and
time domains
Have been studied in many animals but not
adequately in a computer model
Neural network
Training set
0
Time (ms)
Upsweeps
200
Frequency (kHz) 5
Frequency (kHz) 5
Different neurons trained on simple sounds
Frequency (kHz) 5

0
Time (ms)
Pure tones
200
0
Time (ms)
Downsweeps
200
Training procedure

Oja’s rule – modified Hebbian learning
Update rule:

Some particulars:

Training set of 276 stimuli
 600 iterations through each stimulus

Results – weights
210 Frequency (Hz) 1450
2650
3850
1320
2260
3460
4400
1920
4270
3075
5510
Generation of STRFs

Create ripple stimuli by equation:
Generation of STRFs cont.


Feed ripple into model and get output
Fit curve to output
Generation of STRFs cont.

Magnitude and phase shift become one pixel of
transfer function:
Generation of STRFs cont.



2D inverse Fourier transform to obtain STRF
Pad matrix with zeros to get smoother image
Use similar color schemes as literature
Results - STRFs
Comparison to tuning curves


Depicts response intensity to different freqs.
Peak of curve = best frequency (BF)
Comparison of BFs

Peak of tuning curve compared to max value of
STRF
Temporal component


Line of best fit around max area
Different slopes means different temporal
characteristics
Temporal component cont.
Discussion & Conclusion


Real STRFs look similar to computational
STRFs
Means underlying processing is also similar