Artificial neural network class-based pruning
Methods and apparatus, including computer program products, implementing and using techniques for configuring an artificial neural network to a particular surveillance situation. A number of object classes characteristic for the surveillance situation are selected. The object classes form a subset of the total number of object classes for which the artificial neural network is trained. A database is accessed that includes activation frequency values for the neurons within the artificial neural network. The activation frequency values are a function of the object class. Those neurons having activation frequency values lower than a threshold value for the subset of selected object classes are removed from the artificial neural network.
1. A method for configuring an artificial neural network in an embedded device to a particular surveillance situation, the method comprising:
selecting a number of object classes characteristic for the surveillance situation, wherein the object classes form a subset of the total number of object classes for which the artificial neural network is trained;
accessing a database including activation frequency values for the neurons within the artificial neural network, wherein the activation frequency values include at least one activation frequency value for each object class; and
removing from the artificial neural network those neurons having activation frequency values lower than a threshold activation frequency value for the subset of selected object classes, wherein the removed number of neurons is such that a reduction in computational resources needed for calculating the output from the artificial neural network outweighs an increase in computational resources needed for maintaining records about the removed neurons.
2. The method of claim 1 , wherein the threshold activation frequency value is a static value determined for each class during a training/test procedure for the artificial neural network.
3. The method of claim 1 , wherein the threshold activation frequency value is adaptive for each class based on artificial neural network performance.
4. The method of claim 1 , further comprising:
refining the selection of object classes characteristic for the surveillance situation, based on data collected over a certain time period; and
repeating the selecting, accessing and removing steps for the refined selection of object classes.
5. The method of claim 1 , wherein at least some of the selecting, accessing, and removing steps are performed in response to a user input received through a user interface.
6. The method of claim 1 , wherein, in a fully connected artificial neural network, removing neurons corresponds to removing rows and columns from matrices forming mathematical representations of individual layers of the artificial neural network, thereby reducing the number of computations needed in calculating the output from the artificial neural network.
7. A non-transitory computer-readable storage medium including instructions which, when executed in a processing system, cause the system to execute a method for configuring an artificial neural network employed in an embedded device to a particular surveillance situation, the method comprising:
selecting a number of object classes characteristic for the surveillance situation, wherein the object classes form a subset of the total number of object classes for which the artificial neural network is trained;
accessing a database including activation frequency values for the neurons within the artificial neural network, wherein the activation frequency values include at least one activation frequency value for each object class; and
removing from the artificial neural network those neurons having activation frequency values lower than a threshold activation frequency value for the subset of selected object classes, wherein the removed number of neurons is such that a reduction in computational resources needed for calculating the output from the artificial neural network outweighs an increase in computational resources needed for maintaining records about the removed neurons.