Fourier spectra of most changed kernels from re-trained DNN (IMAGE)
Caption
Training cutting-edge deep neural networks requires a great deal of data, and the burden for re-training, with current methods, is still significant. After training and re-training a deep learning network to perform different tasks involving complex physics, Rice University researchers used Fourier analysis to compare all 40,000 kernels from the two iterations and found more than 99% were similar. This illustration shows the Fourier spectra of the four kernels that most differed before (left) and after (right) re-training. The findings demonstrate the method’s potential for identifying more efficient paths for re-training that require significantly less data.
Credit
Image courtesy of P. Hassanzadeh/Rice University
Usage Restrictions
Must credit: P. Hassanzadeh/Rice University
License
Original content