System and method for unsupervised concept extraction from reinforcement learning agents
Described is a method for improved performance of agent-based machine learning. The method includes training a reinforcement learning (RL) agent on an image processing task. A dataset of states and corresponding actions is then extracted from the RL agent. A measure of attention is applied to an input space of the RL agent. During action selection by the RL agent, image patches of the input space are extracted based on the applied measure of attention. Portions of a set of inputs are clustered based on similarity to the image patches, generating a set of clusters having cluster centers. Non-semantic concept labels are provided as distances to the cluster centers for each state in the dataset.
1. A system for performance of agent-based machine learning with unsupervised concept extraction from reinforcement learning agents, 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 perform operations of:
training a reinforcement learning (RL) agent on an image processing task;
extracting a dataset of a plurality of states and corresponding actions from the RL agent, the plurality of states corresponding to a plurality of image frames;
perturbing each image frame by blurring each pixel in the image frame and focusing the blur around each pixel using a Gaussian filter centered on the pixel;
processing each image frame with the retrained RL agent to obtain an output;
comparing the output to the image frame prior to perturbation to determine a saliency score for each pixel in each image frame;
extracting, during action selection by the RL agent, a plurality of image patches of the input space based on the saliency score;
clustering, using an unsupervised clustering algorithm, portions of a first set of inputs based on similarity to the plurality of image patches, thereby generating a first set of clusters having cluster centers; and
providing non-semantic concept labels as distances to the cluster centers for each state in the dataset.
2. The system as set forth in claim 1 , wherein the one or more processors further perform an operation of using the non-semantic concept labels, retraining the RL agent on the image processing task.
3. The system as set forth in claim 1 , wherein the one or more processors further perform an operation of using the non-semantic concept labels, training another system that relies on the same input space.
4. The system as set forth in claim 1 , wherein the one or more processors perform operations of:
determining a saliency map for the image frame based on the saliency scores of the pixels in the image frame;
using the saliency map, extracting contiguous salient portions of the image frame; and
filtering the image frame using the extracted contiguous salient portions to highlight portions of the image frame for the RL agent to use for action selection.
5. A method for performance of agent-based machine learning with unsupervised concept extraction from reinforcement learning agents, the method comprising acts of:
training a reinforcement learning (RL) agent on an image processing task;
extracting a dataset of a plurality of states and corresponding actions from the RL agent, the plurality of states corresponding to a plurality of image frames;
perturbing each image frame by blurring each pixel in the image frame and focusing the blur around each pixel using a Gaussian filter centered on the pixel;
processing each image frame with the retrained RL agent to obtain an output;
comparing the output to the image frame prior to perturbation to determine a saliency score for each pixel in each image frame;
extracting, during action selection by the RL agent, a plurality of image patches of the input space based on the saliency score;
clustering, using an unsupervised clustering algorithm, portions of a first set of inputs based on similarity to the plurality of image patches, thereby generating a first set of clusters having cluster centers; and
providing non-semantic concept labels as distances to the cluster centers for each state in the dataset.
6. The method as set forth in claim 5 , further comprising an act of using the non-semantic concept labels, retraining the RL agent on the image processing.
7. The method as set forth in claim 5 , further comprising an act of using the non-semantic concept labels, training another system that relies on the same input space.
8. The method as set forth in claim 5 , further comprising acts of:
determining a saliency map for the image frame based on the saliency scores of the pixels in the image frame;
using the saliency map, extracting contiguous salient portions of the image frame; and
filtering the image frame using the extracted contiguous salient portions to highlight portions of the image frame for the RL agent to use for action selection.
9. A computer program product for performance of agent-based machine learning with unsupervised concept extraction from reinforcement learning agents, the computer program product comprising:
a non-transitory computer-readable medium having executable instructions encoded thereon, such that upon execution of the instructions by one or more processors, the one or more processors perform operations of:
training a reinforcement learning (RL) agent on an image processing task;
extracting a dataset of a plurality of states and corresponding actions from the RL agent, the plurality of states corresponding to a plurality of image frames;
perturbing each image frame by blurring each pixel in the image frame and focusing the blur around each pixel using a Gaussian filter centered on the pixel;
processing each image frame with the retrained RL agent to obtain an output;
comparing the output to the image frame prior to perturbation to determine a saliency score for each pixel in each image frame;
extracting, during action selection by the RL agent, a plurality of image patches of the input space based on the saliency score;
clustering, using an unsupervised clustering algorithm, portions of a first set of inputs based on similarity to the plurality of image patches, thereby generating a first set of clusters having cluster centers; and
providing non-semantic concept labels as distances to the cluster centers for each state in the dataset.
10. The computer program product as set forth in claim 9 , further comprising instructions for causing the one or more processors to perform an operation of using the non-semantic concept labels, retraining the RL agent on the image processing task.
11. The computer program product as set forth in claim 9 , further comprising instructions for causing the one or more processors to perform an operation of using the non-semantic concept labels, training another system that relies on the same input space.
12. The computer program product as set forth in claim 9 , further comprising instructions for causing the one or more processors to perform operations of:
determining a saliency map for the image frame based on the saliency scores of the pixels in the image frame;
using the saliency map, extracting contiguous salient portions of the image frame; and
filtering the image frame using the extracted contiguous salient portions to highlight portions of the image frame for the RL agent to use for action selection.