System and method for online deep learning in an ultra-low power consumption state

A method for online deep learning, the method comprising causing one or more processors to execute instructions encoded on a non-transitory computer-readable medium, such that upon execution of the instructions, the one or more processors perform acts of:

• generating one or more feature maps by filtering an input image having one or more objects, each feature map having feature values;
• generating decreased size feature maps by dividing the one or more feature maps into non-overlapping small regions and pooling the feature values in each small region;
• generating a collection of cell maps having connections to the decreased sized feature maps by dividing the decreased size feature maps into overlapping patches, with the overlapping patches being joined together such that overlapping patches centered at a same position across all decreased size feature maps are joined together to connect with a single cell in the same position, with all of the single cells collectively forming the collection of cell maps;
• generating a decreased sized collection of cell maps by dividing the collection of cell maps into non-overlapping small regions and pooling feature values in each small region;
• mapping the decreased sized collection of cell maps to a single cell using an ARTMAP inference module, the ARTMAP inference module mapping the decreased collection of cell maps to a single cell by performing operations of:
• checking whether a winner label matches an input data label:
• if they match, then u dating a template corresponding to the winner:
• otherwise, retrieving a degree-of-match from a degree-of-match generator and checking whether the degree of match for the winner is greater than a threshold;
• if so, then doing nothing;
• and
• otherwise, using a new cell to represent the input template and attach the input data label to it;
• and
• generating a class label for the one or more objects based on the single cell.

A computer program product for online deep learning, 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:
• generating one or more feature maps by filtering an input image having one or more objets, each feature map having feature values;
• generating decreased size feature maps by dividing the one or more feature maps into non-overlapping small regions and pooling the feature values in each small region;
• generating a collection of cell maps having connections to the decreased sized feature maps by dividing the decreased size feature maps into overlapping patches, with the overlapping patches being joined together such that overlapping patches centered at a same position across all decreased size feature maps are joined together to connect with a single cell in the same position, with all of the single cells collectively forming the collection of cell maps;
• generating a decreased sized collection of cell maps by dividing the collection of cell maps into non-overlapping small regions and pooling feature values in each small region;
• mapping the decreased sized collection of cell maps to a single cell using an ARTMAP inference module, the ARTMAP inference module mapping the decreased collection of cell maps to a single cell by performing operations of:
• checking whether a winner label matches an input data label;
• if they match, then updating a template corresponding to the winner;
• otherwise, retrieving a degree-of-match from a degree-of-match generator and checking whether the degree of match for the winner is greater than a threshold;
• if so, then doing nothing;
• and
• otherwise, using a new cell to represent the input template and attach the input data label to it;
• and
• generating a class label for the one or more objects based on the single cell.