Continuous control of attention for a deep learning network

A system for reducing computation cost associated with a machine learning task performed by a computer system by implementing continuous control of attention for a deep learning network, comprising:

• a memory device for storing program code;
• and
• at least one processor operatively coupled to the memory device and configured to execute program code stored on the memory device to:
• initialize a control-value function, an observation-value function and a sequence of states associated with a current episode;
• if a current epoch of the current episode is odd:
• select an observation-action;
• execute the observation-action to observe a partial image;
• and
• update the observation-value function based on the partial image and the control-value function;
• and
• if the current epoch is even:
• select a control-action;
• execute the control-action to obtain a reward corresponding to the control-action;
• and
• update the control-value function based on the reward and the observation-value function.

A computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method for reducing computation cost associated with a machine learning task performed by a computer system by implementing continuous control of attention for a deep learning network, the method performed by the computer comprising:

• initializing a control-value function, an observation-value function and a sequence of states associated with a current episode;
• if a current epoch of the current episode is odd:
• selecting an observation-action;
• executing the observation-action to observe a partial image;
• and
• updating the observation-value function based on the partial image and the control-value function;
• and
• if the current epoch is even:
• selecting a control-action;
• executing the control-action to obtain a reward corresponding to the control-action;
• and
• updating the control-value function based on the reward and the observation-value function.