US 12,468,949 B2
Systems and methods for few-shot transfer learning
Soheil Kolouri, Calabasas, CA (US); Mohammad Rostami, Malibu, CA (US); and Kyungnam Kim, Oak Park, CA (US)
Assigned to HRL LABORATORIES, LLC, Malibu, CA (US)
Filed by HRL LABORATORIES, LLC, Malibu, CA (US)
Filed on Aug. 5, 2019, as Appl. No. 16/532,321.
Claims priority of provisional application 62/752,166, filed on Oct. 29, 2018.
Prior Publication US 2020/0130177 A1, Apr. 30, 2020
Int. Cl. G06N 3/084 (2023.01); B25J 9/16 (2006.01); G06F 18/24 (2023.01); G06N 3/045 (2023.01)
CPC G06N 3/084 (2013.01) [B25J 9/163 (2013.01); G06F 18/24 (2023.01); G06N 3/045 (2023.01)] 22 Claims
OG exemplary drawing
 
1. A method for training a controller to control a robotic system in a target domain, the method comprising:
receiving a neural network of an original controller for controlling the robotic system based on a plurality of origin data samples from an origin domain and corresponding labels in a label space, the neural network of the original controller comprising a plurality of first encoder parameters and a plurality of classifier parameters, the neural network being trained to:
map an input data sample from the origin domain to a feature vector in a feature space in accordance with the first encoder parameters; and
assign a label of the label space to the input data sample based on the feature vector in accordance with the classifier parameters;
computing a plurality of second encoder parameters by iteratively updating a plurality of intermediate encoder parameters based on the first encoder parameters to minimize a dissimilarity, in the feature space, between:
a plurality of origin feature vectors computed from the origin data samples using the first encoder parameters; and
a plurality of target feature vectors computed from a plurality of target data samples from the target domain using the intermediate encoder parameters, the target data samples having a smaller cardinality than the origin data samples,
wherein the dissimilarity is computed based on an unsupervised loss function term and a supervised loss function term in accordance with a sliced Wasserstein distance between:
the origin feature vectors encoded in the feature space using the first encoder parameters; and
the target feature vectors encoded in the feature space using the intermediate encoder parameters, and
wherein the intermediate encoder parameters and the second encoder parameters are different from the first encoder parameters; and
updating the controller with the second encoder parameters to control the robotic system in the target domain.