Page 16 - Demo

Stochastic gradient descent (SGD) can be used to determine calibration factors
• 39-dimensional parameter space (1-d for each channel) is difficult to optimize
• SGD works to “push” the calibration factors around the parameter space quickly
• Cost function of SGD is the mean-squared deviation of calibrated points from the reconstructed source
• The calibration results can then be scored by deviation from the correct value:
𝐷𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 = 𝑐"*0,2*1 − 𝑐"%#0/(.,%20,#3 𝑐"*0,2*1
• Synthetic results show most deviations as < 5 %, and a correctly calibrated signal!
Presenter: Gabriel Player
16