Neural network systems

A system comprising:

• a memory configured to store a teacher network, task images, and a student network;
• a controller circuit configured to:
• obtain a teacher network that includes artificial neural layers configured to automatically identify one or more objects in an image examined by the artificial neural layers;
• receive a set of task images at the teacher network, each of the task images in the set depicting a common object;
• examine the set of task images with the teacher network, wherein each of the artificial neural layers in the teacher network is configured to apply one or more functions to each respective pixel of multiple pixels within the task images in the set to generate a score vector for the respective pixel, the score vector representing probabilities that the respective pixel represents different object classes;
• identify a subset of the artificial neural layers of the teacher network which, during the examination of the set of task images, calculate that one or more of the pixels within the task images represent the common object, wherein the controller circuit is configured to identify the subset by selecting the artificial neural layers that, upon applying the one or more functions to a pixel that depicts the common object, generate the score vector indicating that the pixel has a greater probability of representing an object class associated with the common object than any other object class of the different object classes;
• and
• define a student network to include at least the subset of the artificial neural layers of the teacher network, wherein the student network is configured to automatically identify one or more objects in an input image examined by the student network;
• wherein a storage size of the student network is smaller than the teacher network.

A method comprising:

• obtaining a teacher network that includes artificial neural layers configured to automatically identify one or more objects in an image examined by the artificial neural layers;
• receiving a set of task images at the teacher network, each of the task images in the set depicting a common object;
• examining the set of task images with the teacher network, wherein examining the set includes each of the artificial neural layers in the teacher network applying one or more functions to each respective pixel of multiple pixels within the task images in the set to generate a score vector for the respective pixel, the score vector representing probabilities that the respective pixel represents different object classes;
• identifying a subset of the artificial neural layers of the teacher network which, during the examination of the set of task images, calculate that one or more of the pixels within the task images represent the common object, wherein identifying the subset includes selecting the artificial neural layers that, upon applying the one or more functions to a pixel that depicts the common object, generate the score vector indicating that the pixel has a greater probability of representing an object class associated with the common object than any other object class of the different object classes;
• and
• defining a student network to include at least the subset of the artificial neural layers of the teacher network, the student network defined by removing the artificial neural layers of the teacher network that are not a part of the subset such that the student network includes fewer artificial neural layers than the teacher network, wherein the student network is configured to automatically identify one or more objects in an input image examined by the student network, and a storage size of the student network in a memory is smaller than the teacher network.