Deep learning with judgment
Computer systems and computer-implemented methods for modifying a machine learning network, such as a deep neural network, to introduce judgment to the network are disclosed. A “combining” node is added to the network, to thereby generate a modified network, where activation of the combining node is based, at least in part, on output from a subject node of the network. The computer system then trains the modified network by, for each training data item in a set of training data, performing forward and back propagation computations through the modified network, where the backward propagation computation through the modified network comprises computing estimated partial derivatives of an error function of an objective for the network, except that the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node, even though activation of the combining node is based on the activation of the subject node.
1. A method of accelerating training of a deep neural network, the method comprising:
initial iterative training, by a programmed computer system, through machine learning, of the deep neural network over a mega-batch of training data, wherein the mega-batch comprises multiple mini-batches of training data, and the initial iterative training comprises updating learned parameters of the deep neural network based on each mini-batch of training data;
accumulating, by the programmed computer system, mega-batch-accumulated statistics over the mega-batch, wherein the mega-batch-accumulated statistics comprise mega-batch-accumulated statistics for a subject node of the deep neural network; and
after the initial iterative training, training, by the programmed computer system, one or more nodes of the deep neural network based on the mega-batch statistics, wherein training the one or more nodes based on the mega-batch statistics comprises:
adding a combining node and an error judgement node to the deep neural network, wherein:
the error judgement node is for producing an activation value for a datum indicative of whether an activation value of the subject node of the deep neural is erroneous for the datum;
an activation value of the combining node for the datum is based on the activation values of both the subject and error judgement nodes; and
the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node; and
training the error judgement node based on the mega-batch-accumulated statistics.
2. The method of claim 1 , wherein training the one or more added nodes of the deep neural network based on the mega-batch-accumulated statistics further comprises initializing learned parameters for the one or more nodes based on the mega-batch-accumulated statistics.
3. The method of claim 2 , wherein initializing the learned parameters for the one or more added nodes comprises learning the learned parameters for the one or more added nodes based on the mega-batch-accumulated statistics.
4. The method of claim 3 , wherein learning the learned parameters for the one or more nodes comprises fixing other learned parameters of the deep neural network.
5. The method of claim 1 , wherein:
the one or more nodes is a parametric probability distribution; and
the mega-batch-accumulated statistics comprise sufficient statistics for the parametric probability distribution.
6. The method of claim 1 , wherein:
the one or more nodes comprises a first node; and
training the first node based on the mega-batch-accumulated statistics comprises learning learned parameters for the first node using iteratively reweighted least squares on the mega-batch-accumulated statistics.
7. A system for accelerating training of a deep neural network, the system comprising:
one or more processor cores; and
computer memory in communication with the one or more processor cores, wherein the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to:
iteratively train through machine learning, the deep neural network in an initial training, wherein the initial training comprises iteratively training the deep neural network over a mega-batch of training data, wherein the mega-batch comprises multiple mini-batches of training data, and the initial training comprises updating learned parameters of the deep neural network based on each mini-batch of training data;
accumulate mega-batch-accumulated statistics over the mega-batch, wherein the mega-batch-accumulated statistics comprise mega-batch-accumulated statistics for a subject node of the deep neural network; and
after the initial training, train one or more nodes of the deep neural network based on the mega-batch-accumulated statistics by, in part:
adding a combining node and an error judgement node to the deep neural network, wherein:
the error judgement node is for producing an activation value for a datum indicative of whether an activation value of the subject node of the deep neural is erroneous for the datum;
an activation value of the combining node for the datum is based on the activation values of both the subject and error judgement nodes; and
the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node; and
training the error judgement node based on the mega-batch-accumulated statistics.
8. The system of claim 7 , wherein the computer memory stores computer instructions that when executed by the one or more processor cores further cause the one or more processor cores to train the one or more added nodes of the deep neural network based on the mega-batch-accumulated statistics by initializing learned parameters for the one or more nodes based on the mega-batch-accumulated statistics.
9. The system of claim 8 , wherein the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to initialize the learned parameters for the one or more added nodes by learning the learned parameters for the one or more added nodes based on the mega-batch-accumulated statistics.
10. The system of claim 9 , wherein the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to learn the learned parameters for the one or more nodes by fixing other learned parameters of the deep neural network.
11. The system of claim 7 , wherein:
the one or more nodes is a parametric probability distribution; and
the mega-batch-accumulated statistics comprise sufficient statistics for the parametric probability distribution.
12. The system of claim 7 , wherein:
the one or more nodes comprises a first node; and
the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to train the first node based on the mega-batch-accumulated statistics by learning learned parameters for the first node using iteratively reweighted least squares on the mega-batch-accumulated statistics.
13. A method of accelerating training of a deep neural network, the method comprising:
initial iterative training, by a programmed computer system, through machine learning, of the deep neural network over a mega-batch of training data, wherein the mega-batch comprises multiple mini-batches of training data, and the initial iterative training comprises updating learned parameters of the deep neural network based on each mini-batch of training data;
accumulating, by the programmed computer system, mega-batch-accumulated statistics over the mega-batch; and
after the initial iterative training:
partitioning a domain of an activation function for a subject node of the neural network into multiple intervals;
creating a subnetwork of nodes, wherein the subnetwork comprises a new partition node for each of the multiple intervals; and
connecting the subnetwork to the deep neural network.
14. The method of claim 13 , wherein creating the subnetwork of nodes comprises defining an activation function for each new partition node, such that the activation function for each new partition node has a same value as the activation function of the subject node for data for which an input to the activation function is within an interval corresponding to the new partition node.
15. The method of claim 14 , wherein connecting the subnetwork to the deep neural network comprises connecting the subnetwork to the subject node.
16. The method of claim 13 , wherein creating the subnetwork of nodes comprises defining an activation function for each new partition node, such that the activation function for each new partition node differs by a constant value from the activation function of the subject node for data for which an input to the activation function is within an interval corresponding to the new partition node.
17. The method of claim 16 , wherein connecting the subnetwork to the deep neural network comprises connecting the subnetwork to the subject node.
18. A system for accelerating training of a deep neural network, the system comprising:
one or more processor cores; and
computer memory in communication with the one or more processor cores, wherein the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to:
iteratively train through machine learning, the deep neural network in an initial training, wherein the initial training comprises iteratively training the deep neural network over a mega-batch of training data, wherein the mega-batch comprises multiple mini-batches of training data, and the initial training comprises updating learned parameters of the deep neural network based on each mini-batch of training data;
accumulate mega-batch-accumulated statistics over the mega-batch; and
after the initial training:
partition a domain of an activation function for a subject node of the deep neural network into multiple intervals;
create a subnetwork of nodes, wherein the subnetwork comprises a new partition node for each of the multiple intervals; and
connect the subnetwork to the deep neural network.
19. The system of claim 18 , wherein the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to create the subnetwork of nodes by defining an activation function for each new partition node, such that the activation function for each new partition node has a same value as the activation function of the subject node for data for which an input to the activation function is within an interval corresponding to the new partition node.
20. The system of claim 19 , wherein the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to connect the subnetwork to the deep neural network by connecting the subnetwork to the subject node.
21. The system of claim 18 , wherein the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to create the subnetwork of nodes by defining an activation function for each new partition node, such that the activation function for each new partition node differs by a constant value from the activation function of the subject node for data for which an input to the activation function is within an interval corresponding to the new partition node.
22. The system of claim 21 , wherein the computer memory stores computer instructions that when executed by the one or more processor cores cause the one or more processor cores to connect the subnetwork to the deep neural network by connecting the subnetwork to the subject node.