[3Rin4] Interactive 1

[3Rin4-23] Approximate function estimation with GAN

Estimate the true value distribution of the whole space that constantly fluctuates from the observation data of small number of sensors

〇akira noda¹ (1.Technology Research Laboratory , Shimadzu)

Keywords:Generative Adversarial Network(GAN), Compressed Sensing, Neural network

In compressed sensing with GAN, you have to prepare dense observation data before learning for recovering the true value from a small number of observation sensor values.However, in the advanced measurement field,sometimes you can not get the dense true value.Therefore, we propose a GAN-like algorithm(fGAN) that obtains a function that indicates the true value of each state, assuming that the observation target has only a finite state.fGAN uses the massive number of sensor data but only few of the data are measured at same time.
Simulation experiments show that the true value function can be approximated when the dimension of the state variable is sufficiently small.This obtained function can be used for compressed sensing which estimate a distribution of true values in the whole space from one set of a small number of observation sensor values.