System and method for teaching compositionality to convolutional neural networks
A system for teaching compositionality to convolutional neural networks includes an unmasked convolutional neural network comprising a first set of convolutional neural network layers; a first masked convolutional neural network comprising a second set of convolutional neural network layers; the unmasked convolutional neural network and the first masked convolutional network sharing convolutional neural network weights; the system training the unmasked and first masked convolutional neural networks simultaneously based on an objective function that seeks to reduce both discriminative loss and compositional loss.
1. A system for teaching compositionality to convolutional neural networks, comprising:
an unmasked convolutional neural network comprising a first set of convolutional neural network layers; wherein input to the unmasked convolutional neural network is maskless;
a first masked convolutional neural network comprising a second set of convolutional neural network layers; wherein input to the first masked convolutional neural network is masked by a first object mask; wherein the unmasked convolutional neural network and the first masked convolutional network share convolutional neural network weights; wherein the first object mask suppresses activation outside of a region of a first object; and
a first penalty mask, corresponding to the first object mask, that operates on output of one of the first set of convolutional neural network layers; wherein output of the first penalty mask is used to calculate compositional loss;
wherein the system trains the unmasked and first masked convolutional neural networks simultaneously based on an objective function that seeks to reduce both discriminative loss and compositional loss.
2. The system of claim 1 , further comprising a second object mask, identical in function to the first object mask, located between two of the second set of convolutional neural network layers.
3. The system of claim 2 , wherein the first penalty mask is identical in function to the first and second object masks.
4. The system of claim 2 , wherein the first penalty mask is distinct in function from the first and second object masks.
5. The system of claim 4 , wherein the first penalty mask suppresses activation only in regions corresponding to any object mask other than the first object mask.
6. The system of claim 1 , wherein the system trains the unmasked and first masked convolutional neural networks simultaneously, based on the objective function, using a gradient descent algorithm.
7. The system of claim 6 , wherein the objective function models compositional loss as a sum of least square differences across activation outputs of the second set of convolutional neural network layers and a product of activation outputs of the second set of convolutional neural network layers and the first penalty mask.
8. The system of claim 7 , wherein the objective function models discriminative loss using softmax-cross entropy for joint class prediction.
9. The system of claim 7 , wherein the objective function models discriminative loss using sigmoid-cross entropy for independent class prediction.
10. The system of claim 1 , wherein the unmasked and first masked convolutional neural network are trained to receive data feature input by setting activation of child feature nodes according to image features of an image; wherein the inferred output includes a classification of the image.
11. The system of claim 1 , wherein the unmasked and first masked convolutional neural network are trained to receive data feature input by setting activation of child feature nodes according to audio features of an audio signal; wherein the inferred output includes a classification of the audio signal.
12. The system of claim 1 , wherein the unmasked convolutional neural network and the first masked convolutional network share all convolutional neural network weights.
13. A method for teaching compositionality to convolutional neural networks, comprising:
receiving an initial masked CNN structure, the structure comprising:
an unmasked convolutional neural network comprising a first set of convolutional neural network layers; wherein input to the unmasked convolutional neural network is maskless; and
a first masked convolutional neural network comprising a second set of convolutional neural network layers; wherein input to the first masked convolutional neural network is masked by a first object mask; wherein the unmasked convolutional neural network and the first masked convolutional network share convolutional neural network weights; wherein the first object mask suppresses activation outside of a region of a first object;
receiving a training dataset; and
learning the convolutional neural network weights; wherein learning the convolutional neural network weights comprises training the unmasked and first masked convolutional neural networks simultaneously on the training dataset based on an objective function that seeks to reduce both discriminative loss and compositional loss; wherein receiving an initial masked CNN structure further comprises receiving a first penalty mask, corresponding to the first object mask, that operates on output of one of the first set of convolutional neural network layers; wherein output of the first penalty mask is used to calculate compositional loss.
14. The method of claim 13 , wherein receiving an initial masked CNN structure further comprises receiving a second object mask, identical in function to the first object mask, located between two of the second set of convolutional neural network layers.
15. The method of claim 14 , wherein the first penalty mask is identical in function to the first and second object masks.
16. The method of claim 14 , wherein the first penalty mask is distinct in function from the first and second object masks.
17. The method of claim 15 , wherein the first penalty mask suppresses activation only in regions corresponding to any object mask other than the first object mask.
18. The method of claim 14 , wherein learning the convolutional neural network weights comprises training the unmasked and first masked convolutional neural networks simultaneously, based on the objective function, using a gradient descent algorithm.
19. The method of claim 18 , wherein the objective function models compositional loss as a sum of least square differences across activation outputs of the second set of convolutional neural network layers and a product of activation outputs of the second set of convolutional neural network layers and the first penalty mask.
20. A system for teaching compositionality to convolutional neural networks, comprising:
an unmasked convolutional neural network comprising a first set of convolutional neural network layers; wherein input to the unmasked convolutional neural network is maskless; and
a first masked convolutional neural network comprising a second set of convolutional neural network layers; wherein input to the first masked convolutional neural network is masked by a first object mask; wherein the unmasked convolutional neural network and the first masked convolutional network share convolutional neural network weights; wherein the first object mask suppresses activation outside of a region of a first object;
wherein the system trains the unmasked and first masked convolutional neural networks simultaneously based on an objective function that seeks to reduce both discriminative loss and compositional loss; wherein the unmasked and first masked convolutional neural network are trained to receive data feature input by setting activation of child feature nodes according to image features of an image; wherein the inferred output includes a classification of the image.
21. A system for teaching compositionality to convolutional neural networks, comprising:
an unmasked convolutional neural network comprising a first set of convolutional neural network layers; wherein input to the unmasked convolutional neural network is maskless; and
a first masked convolutional neural network comprising a second set of convolutional neural network layers; wherein input to the first masked convolutional neural network is masked by a first object mask; wherein the unmasked convolutional neural network and the first masked convolutional network share convolutional neural network weights; wherein the first object mask suppresses activation outside of a region of a first object;
wherein the system trains the unmasked and first masked convolutional neural networks simultaneously based on an objective function that seeks to reduce both discriminative loss and compositional loss; wherein the unmasked and first masked convolutional neural network are trained to receive data feature input by setting activation of child feature nodes according to audio features of an audio signal; wherein the inferred output includes a classification of the audio signal.