US 12,032,523 B2
Compressed sensing using neural networks
Yan Wu, St. Albans (GB); Timothy Paul Lillicrap, London (GB); and Mihaela Rosca, London (GB)
Assigned to DeepMind Technologies Limited, London (GB)
Filed by DeepMind Technologies Limited, London (GB)
Filed on Mar. 13, 2020, as Appl. No. 16/818,895.
Claims priority of provisional application 62/817,979, filed on Mar. 13, 2019.
Prior Publication US 2020/0293497 A1, Sep. 17, 2020
Int. Cl. G06F 16/174 (2019.01); G06N 3/045 (2023.01); G06N 3/08 (2023.01)
CPC G06F 16/1744 (2019.01) [G06N 3/045 (2023.01); G06N 3/08 (2013.01)] 25 Claims
OG exemplary drawing
 
1. A method of reconstructing a compressed data item, the method comprising:
receiving an input measurement of an input data item, wherein the input measurement of the input data item is a representation of at least one characteristic of the input data item and wherein the input measurement has a lower dimensionality than the input data item;
initializing a latent representation, wherein the latent representation has a lower dimensionality than the input data item;
for each of one or more optimization steps:
processing the latent representation using a generator neural network having a plurality of generator parameters, wherein the generator neural network is configured to process the latent representation in accordance with current values of the generator parameters to generate a candidate reconstructed data item,
processing the candidate reconstructed data item using a measurement neural network having a plurality of measurement parameters, wherein the measurement neural network is configured to process the candidate reconstructed data item in accordance with current values of the measurement parameters to generate a measurement of the candidate reconstructed data item that has the same dimensionality as the input measurement and has a lower dimensionality than the candidate reconstructed data item, and wherein the measurement neural network that generates the measurement of the candidate reconstructed data item has been trained jointly with the generator neural network that generates the candidate reconstructed data item, the joint training comprising training the measurement neural network on a loss that comprises a measurement loss function that measures, for each of one or more training data items, a difference between (i) a norm of a measurement, generated by the measurement neural network, of a difference between the training data item and a reconstruction of the training data item and (ii) a norm of a difference between the training data item and the reconstruction of the training data item; and
updating the latent representation to reduce an error between the measurement of the candidate reconstructed data item generated by the measurement neural network and the input measurement of the input data item, wherein updating the latent representation to reduce the error comprises performing a gradient descent step with respect to the latent representation, and wherein performing the gradient descent step comprises computing a gradient with respect to the latent representation of the error between the measurement of the candidate reconstructed data item generated by the measurement neural network and the input measurement of the input data item; and
processing the latent representation after the one or more optimization steps using the generator neural network and in accordance with the current values of the generator parameters to generate a reconstruction of the input data item.