Initialization of classification layers in neural networks
Example methods disclosed herein to initialize classification vectors of a neural network include ranking a plurality of classes to be represented by the classification vectors based on respective numbers of instances of training data associated with corresponding ones of the classes. Disclosed example methods also include initializing the classification vectors to span a classification space corresponding to the classes. Disclosed example methods further include assigning respective ones of the classes to corresponding ones of the classification vectors based on the ranking of the classes.
1. A tangible computer readable storage medium comprising instructions which, when executed, cause one or more processors to at least:
rank a plurality of classes to be represented by classification vectors based on respective numbers of instances of training data associated with corresponding ones of the classes, the classification vectors to define classification areas in a classification space corresponding to the classes;
initialize the classification vectors to be angled substantially equidistant from each other or to define the classification areas to be substantially equal in size, the classification vectors to span the classification space; and
assign respective ones of the classes to corresponding ones of the classification vectors based on the ranking of the classes.
2. The computer readable storage medium of claim 1 , wherein the instructions cause the one or more processors to train an artificial intelligence model based on the assignment of the classes to the classification vectors.
3. The computer readable storage medium of claim 2 , wherein the artificial intelligence model is a neural network.
4. A tangible computer readable storage medium comprising instructions which, when executed, cause one or more processors to at least:
rank a plurality of classes to be represented by classification vectors based on respective numbers of instances of training data associated with corresponding ones of the classes, the rank of the plurality of classes to be in order from a highest number of instances of the training data to a lowest number of the instances for the training data or from the lowest number of the instances of the training data to the highest number of the instances of the training data;
initialize the classification vectors to span a classification space corresponding to the classes; and
assign respective ones of the classes to corresponding ones of the classification vectors based on the ranking of the classes by:
selecting a first class of the classes, the first class corresponding to a first number of training instances;
assigning the first class to a first classification vector;
selecting a second class of the classes, the second class being ranked consecutively relative to the first class; and
assigning the second class to a second classification vector neighboring the first classification vector.
5. A tangible computer readable storage medium comprising instructions which, when executed, cause one or more processors to at least:
rank a plurality of classes to be represented by classification vectors based on respective numbers of instances of training data associated with corresponding ones of the classes, the rank of the plurality of classes to be in order from a highest number of instances of the training data to a lowest number of the instances for the training data or from the lowest number of the instances of the training data to the highest number of the instances of the training data;
initialize the classification vectors to span a classification space corresponding to the classes; and
assign respective ones of the classes to corresponding ones of the classification vectors based on the ranking of the classes to ensure a first class with a low rank does not neighbor a second class with a high rank of training instances.
6. An apparatus comprising:
a comparator to rank a plurality of classes to be represented by classification vectors based on respective numbers of instances of training data associated with corresponding ones of the classes, the classification vectors defining classification areas in a classification space corresponding to the classes; and
a classification space determiner to:
initialize the classification vectors to be angled substantially equidistant from each other or to define the classification areas to be substantially equal in size, the classification vectors to span the classification space; and
assign respective ones of the classes to corresponding ones of the classification vectors based on the ranking of the classes.
7. The apparatus of claim 6 , further including a model trainer to train an artificial intelligence model based on the assignment of the classes to the classification vectors.
8. The apparatus of claim 7 , wherein the artificial intelligence model is a neural network.
9. An apparatus comprising:
a comparator to rank a plurality of classes to be represented by classification vectors based on respective numbers of instances of training data associated with corresponding ones of the classes, rank of the plurality of classes to be in order from a highest number of instances of the training data to a lowest number of the instances for the training data or from the lowest number of the instances of the training data to the highest number of the instances of the training data; and
a classification space determiner to:
initialize the classification vectors to span a classification space corresponding to the classes; and
assign respective ones of the classes to corresponding ones of the classification vectors based on the ranking of the classes by:
selecting a first class of the classes, the first class corresponding to a first number of training instances;
assigning the first class to a first classification vector;
selecting a second class of the classes, the second class being ranked consecutively relative to the first class; and
assigning the second class to a second classification vector neighboring the first classification vector.
10. An apparatus comprising:
a comparator to rank a plurality of classes to be represented by classification vectors based on respective numbers of instances of training data associated with corresponding ones of the classes, the rank of the plurality of classes to be in order from a highest number of instances of the training data to a lowest number of the instances for the training data or from the lowest number of the instances of the training data to the highest number of the instances of the training data; and
a classification space determiner to:
initialize the classification vectors to span a classification space corresponding to the classes; and
assign respective ones of the classes to corresponding ones of the classification vectors based on the ranking of the classes to ensure a first class with a low rank does not neighbor a second class with a high rank of training instances.
11. A method comprising:
ranking, by executing an instruction with one or more processors, a plurality of classes to be represented by classification vectors based on respective numbers of instances of training data associated with corresponding ones of the classes, the classification vectors defining classification areas in a classification space corresponding to the classes;
initializing, by executing an instruction with the one or more processors, the classification vectors to be angled substantially equidistant from each other or to define the classification areas to be substantially equal in size, the classification vectors to span the classification space; and
assigning, by executing an instruction with the one or more processors, respective ones of the classes to corresponding ones of the classification vectors based on the ranking of the classes.
12. The method of claim 11 , further including training an artificial intelligence model based on the assignment of the classes to the classification vectors.
13. The method of claim 12 , wherein the artificial intelligence model is a neural network.
14. The method of claim 11 , wherein the ranking of the plurality of classes includes ranking the plurality of classes in order from a highest number of the instances of the training data to a lowest number of the training data or from the lowest number of the instances of the training data to the highest number of the instances of the training data.
15. The method of claim 14 , wherein the assigning of the respective ones of the classes to the corresponding ones of the classification vectors includes:
selecting a first class of the classes, the first class corresponding to a first number of training instances;
assigning the first class to a first classification vector;
selecting a second class of the classes, the second class being ranked consecutively relative to the first class; and
assigning the second class to a second classification vector neighboring the first classification vector.
16. The method of claim 14 , wherein the assigning the respective ones of the classes to the corresponding ones of the classification vectors is to ensure a first class with a low rank does not neighbor a second class with a high rank of training instances.