Process to learn new image classes without labels

1. A system for learning object labels for control of an autonomous platform, the system comprising:

one or more processors and a non-transitory computer-readable medium having executable instructions encoded thereon such that when executed, the one or more processors

based on a similarity measure computed between a set of prior trained source models, creating a source similarity graph of similar source models;

performing pseudo-task optimization to identify a pseudo-task for each source model in the source similarity graph, wherein performing pseudo-task optimization comprises:

identifying a pseudo-task for a target data set X T ϵ as a function of {f i : → } i and {α i } i ,

where {α i } i denotes similarities of the target data set to the set of prior trained source models {X i } i such that Σ i α i =1, and

where {f i : → } i denotes corresponding pseudo-tasks for the set of prior trained source models;

combining the pseudo-tasks identified;

using the combined pseudo-tasks and a set of unlabeled target data, training a target network;

using the trained target network, extracting salient components from input image data, resulting in a plurality of source image components having corresponding labels for each source model;

mapping the plurality of source image components and corresponding labels onto sets of abstract attributes, resulting in semantically meaningful clusters of abstract attributes;

aligning, in feature space, new image components from a previously unseen input image with feature space representations of the plurality of source image components;

mapping the new image components onto corresponding abstract attributes;

generating a new target network from the mapping;

using the new target network, assigning an object label to an object in the previously unseen input image; and

controlling the autonomous platform based on the assigned object label.

2. The system as set forth in claim 1 , wherein controlling the autonomous platform further comprises:

generating an executable control script appropriate for the object label; and

causing the autonomous platform to execute the control script and perform an action corresponding to the control script.

3. The system as set forth in claim 1 , wherein the salient components are extracted using unsupervised data decomposition.

4. The system as set forth in claim 1 , wherein the autonomous platform is a vehicle, and wherein the one or more processors further perform an operation of causing the vehicle to perform a driving operation in accordance with the assigned object label.

5. A computer implemented method for learning object labels for control of an autonomous platform, the method comprising an act of:

causing one or more processors to execute instructions encoded on a non-transitory computer-readable medium, such that upon execution, the one or more processors perform operations of:

based on a similarity measure computed between a set of prior trained source models, creating a source similarity graph of similar source models;

performing pseudo-task optimization to identify a pseudo-task for each source model in the source similarity graph, wherein performing pseudo-task optimization comprises:

identifying a pseudo-task for a target data set X T ϵ as a function of {f i : → } i and {α i } i ,

where {α i } i denotes similarities of the target data set to the set of prior trained source models {X i } i such that Σ i α i =1, and

where {f i : → } i denotes corresponding pseudo-tasks for the set of prior trained source models;

combining the pseudo-tasks identified;

using the combined pseudo-tasks and a set of unlabeled target data, training a target network;

using the trained target network, extracting salient components from input image data, resulting in a plurality of source image components having corresponding labels for each source model;

mapping the plurality of source image components and corresponding labels onto sets of abstract attributes, resulting in semantically meaningful clusters of abstract attributes;

aligning, in feature space, new image components from a previously unseen input image with feature space representations of the plurality of source image components;

mapping the new image components onto corresponding abstract attributes;

generating a new target network from the mapping;

using the new target network, assigning an object label to an object in the previously unseen input image; and

controlling the autonomous platform based on the assigned object label.

6. The method as set forth in claim 5 , wherein controlling the autonomous platform further comprises:

generating an executable control script appropriate for the object label; and

causing the autonomous platform to execute the control script and perform an action corresponding to the control script.

7. The method as set forth in claim 5 , wherein the salient components are extracted using unsupervised data decomposition.

8. The method as set forth in claim 5 , wherein the autonomous platform is a vehicle, and wherein the one or more processors further perform an operation of causing the vehicle to perform a driving operation in accordance with the assigned object label.

9. A computer program product for learning object labels for control of an autonomous platform, the computer program product comprising:

computer-readable instructions stored on a non-transitory computer-readable medium that are executable by a computer having one or more processors for causing the processor to perform operations of:

based on a similarity measure computed between a set of prior trained source models, creating a source similarity graph of similar source models;

performing pseudo-task optimization to identify a pseudo-task for each source model in the source similarity graph, wherein performing pseudo-task optimization comprises:

identifying a pseudo-task for a target data set X T ϵ as a function of {f i : → } i and {α i } i ,

where {α i } i denotes similarities of the target data set to the set of prior trained source models {X i } i such that Σ i α i =1, and

where {f i : → } i denotes corresponding pseudo-tasks for the set of prior trained source models;

combining the pseudo-tasks identified;

using the combined pseudo-tasks and a set of unlabeled target data, training a target network;

using the trained target network, extracting salient components from input image data, resulting in a plurality of source image components having corresponding labels for each source model;

mapping the plurality of source image components and corresponding labels onto sets of abstract attributes, resulting in semantically meaningful clusters of abstract attributes;

aligning, in feature space, new image components from a previously unseen input image with feature space representations of the plurality of source image components;

mapping the new image components onto corresponding abstract attributes;

generating a new target network from the mapping;

using the new target network, assigning an object label to an object in the previously unseen input image; and

controlling the autonomous platform based on the assigned object label.

10. The computer program product as set forth in claim 9 , wherein controlling the autonomous platform further comprises:

generating an executable control script appropriate for the object label; and

causing the autonomous platform to execute the control script and perform an action corresponding to the control script.

11. The computer program product as set forth in claim 9 , wherein the salient components are extracted using unsupervised data decomposition.

12. The computer program product as set forth in claim 9 , wherein the autonomous platform is a vehicle, and wherein controlling the autonomous platform further comprises causing the vehicle to perform a driving operation in accordance with the assigned object label.