Systems and methods for unsupervised continual learning
Described is a system for continual adaptation of a machine learning model implemented in an autonomous platform. The system adapts knowledge previously learned by the machine learning model for performance in a new domain. The system receives a consecutive sequence of new domains comprising new task data. The new task data and past learned tasks are forced to share a data distribution in an embedding space, resulting in a shared generative data distribution. The shared generative data distribution is used to generate a set of pseudo-data points for the past learned tasks. Each new domain is learned using both the set of pseudo-data points and the new task data. The machine learning model is updated using both the set of pseudo-data points and the new task data.
1. A system for continual adaptation of a machine learning model implemented in an autonomous platform, the system comprising:
one or more processors and one or more associated memories, each associated memory being a non-transitory computer-readable medium having executable instructions encoded thereon such that when executed, the one or more processors perform an operation of:
adapting a set of knowledge previously learned by a machine learning model for performance in a new domain, wherein adapting the set of knowledge comprises:
receiving a consecutive sequence of new domains, where each new domain comprises new task data;
forcing, through dimensionality reduction, the new task data and a plurality of past learned tasks to share a same parametric data distribution in a task invariant embedding space as clusters of consolidated classes, resulting in a shared generative data distribution;
using the shared generative data distribution, generating a set of pseudo-data points for the past learned tasks;
learning each new domain using both the set of pseudo-data points and the new task data such that each new domain learned has an empirical data distribution in the task invariant embedding space that matches the same parametric data distribution;
updating the machine learning model using both the set of pseudo-data points and the new task data; and
causing one or more mechanical components of the autonomous platform to actuate and, in doing so, causing the autonomous platform to perform a physical driving operation based on the new task data.
2. The system as set forth in claim 1 , wherein the one or more processors further perform an operation of using knowledge of data distributions obtained from the plurality of past learned tasks to match a data distribution of new task data in the embedding space.
3. The system as set forth in claim 1 , wherein the embedding space is invariant with respect to any learned task, such that new task data does not interfere with remembering any past learned task.
4. The system as set forth in claim 1 , wherein the one or more processors further perform an operation of using a Sliced Wasserstein Distance metric to force the new task data and the plurality of past learned tasks to share the data distribution in the embedding space.
5. The system as set forth in claim 1 , wherein the shared generative data distribution is a multi-modal distribution modeled as a Gaussian mixture model.
6. A computer implemented method for continual adaptation of a machine learning model implemented in an autonomous platform, the method comprising an act of:
causing one or more processors to execute instructions encoded on one or more associated memories, each associated memory being a non-transitory computer-readable medium, such that upon execution, the one or more processors perform operations of:
adapting a set of knowledge previously learned by a machine learning model for performance in a new domain, wherein adapting the set of knowledge comprises:
receiving a consecutive sequence of new domains, where each new domain comprises new task data;
forcing, through dimensionality reduction, the new task data and a plurality of past learned tasks to share a same parametric data distribution in a task invariant embedding space as clusters of consolidated classes, resulting in a shared generative data distribution;
using the shared generative data distribution, generating a set of pseudo-data points for the past learned tasks;
learning each new domain using both the set of pseudo-data points and the new task data such that each new domain learned has an empirical data distribution in the task invariant embedding space that matches the same parametric data distribution;
updating the machine learning model using both the set of pseudo-data points and the new task data; and
causing one or more mechanical components of the autonomous platform to actuate and, in doing so, causing the autonomous platform to perform a physical driving operation based on the new task data.
7. The method as set forth in claim 6 , wherein the one or more processors further perform an operation of using knowledge of data distributions obtained from the plurality of past learned tasks to match a data distribution of new task data in the embedding space.
8. The method as set forth in claim 6 , wherein the embedding space is invariant with respect to any learned task, such that new task data does not interfere with remembering any past learned task.
9. The method as set forth in claim 6 , wherein the one or more processors further perform an operation of using a Sliced Wasserstein Distance metric to force the new task data and the plurality of past learned tasks to share the data distribution in the embedding space.
10. The method as set forth in claim 6 , wherein the shared generative data distribution is a multi-modal distribution modeled as a Gaussian mixture model.
11. A computer program product for continual adaptation of a machine learning model implemented in 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:
adapting a set of knowledge previously learned by a machine learning model for performance in a new domain, wherein adapting the set of knowledge comprises:
receiving a consecutive sequence of new domains, where each new domain comprises new task data;
forcing, through dimensionality reduction, the new task data and a plurality of past learned tasks to share a same parametric data distribution in a task invariant embedding space as clusters of consolidated classes, resulting in a shared generative data distribution;
using the shared generative data distribution, generating a set of pseudo-data points for the past learned tasks;
learning each new domain using both the set of pseudo-data points and the new task data such that each new domain learned has an empirical data distribution in the task invariant embedding space that matches the same parametric data distribution;
updating the machine learning model using both the set of pseudo-data points and the new task data; and
causing one or more mechanical components of the autonomous platform to actuate and, in doing so, causing the autonomous platform to perform a physical driving operation based on the new task data.
12. The computer program product as set forth in claim 11 , wherein the one or more processors further perform an operation of using knowledge of data distributions obtained from the plurality of past learned tasks to match a data distribution of new task data in the embedding space.
13. The computer program product as set forth in claim 11 , wherein the embedding space is invariant with respect to any learned task, such that new task data does not interfere with remembering any past learned task.
14. The computer program product as set forth in claim 11 , wherein the one or more processors further perform an operation of using a Sliced Wasserstein Distance metric to force the new task data and the plurality of past learned tasks to share the data distribution in the embedding space.
15. The computer program product as set forth in claim 11 , wherein the shared generative data distribution is a multi-modal distribution modeled as a Gaussian mixture model.