US 12,229,685 B2
Model suitability coefficients based on generative adversarial networks and activation maps
Hariharan Ravishankar, Bangalore (IN); Rahul Venkataramani, Bangalore (IN); Prasad Sudhakara Murthy, Bangalore (IN); and Annangi P. Pavan Kumar, Bangalore (IN)
Assigned to GE PRECISION HEALTHCARE LLC, Waukesha, WI (US)
Filed by GE Precision Healthcare LLC, Milwaukee, WI (US)
Filed on Jan. 22, 2021, as Appl. No. 17/155,997.
Prior Publication US 2022/0237467 A1, Jul. 28, 2022
Int. Cl. G06N 3/088 (2023.01); G06N 3/045 (2023.01)
CPC G06N 3/088 (2013.01) [G06N 3/045 (2023.01)] 3 Claims
OG exemplary drawing
 
1. A system, comprising:
a processor that executes computer-executable instructions stored in a computer-readable memory, which causes the processor to:
access a deep learning model that is trained on a training dataset;
compute a model suitability coefficient that indicates whether the deep learning model is suitable for deployment on a target dataset, by:
training a generative adversarial network (GAN) to learn a distribution of hidden activation maps that the deep learning model produces for a set of first samples from the training dataset;
executing the deep learning model on a set of second samples from the target dataset, thereby causing the deep learning model to produce a set of target hidden activation maps;
causing a generator of the GAN to generate a set of synthetic hidden activation maps from the learned distribution of the deep learning model; and
iteratively perturbing inputs of the generator until distances between the set of synthetic hidden activation maps and the set of target hidden activation maps are minimized, wherein the model suitability coefficient is an aggregation of the minimized distances; and
in response to a determination that the model suitability coefficient does not satisfy a threshold value, transmitting an electronic notification indicating that the deep learning model warrants fine-tuning before being executed on a remainder of the target dataset.