Selective training of deep learning modules
Machine-learning computer system breaks a neural network into a plurality of modules and tracks the training process module-by-module and datum-by-datum, recording auxiliary information during one iteration of the training process for retrieval during a later iteration. Based on this auxiliary information, the computer system can make decisions that can greatly reduce the amount of computation required by the training process. The auxiliary information allows the computer system to diagnose and fix problems that occur during the training process on a module-by-module and/or datum-by-datum basis.
1 . A method of training a parent neural network, wherein the parent neural network comprises a plurality of network modules and a plurality of nodes, including a first network module that comprises one or more nodes and a second network module that comprises one or more nodes, wherein both the first network module and the second network module are different subsets of the parent neural network, the method comprising, by a computer system:
iteratively training, by a programmed computer system that comprises multiple graphical processing units (GPUs) and a data store, through machine learning, the parent neural network over N iterations, where N is greater than one, wherein each of the N iterations comprises a plurality of training data items, wherein each of the training data items comprises a vector of numerical values, and wherein the training comprises, for each of the N iterations:
a forward computation, for each of the plurality of training data items for the iteration through the parent neural network in which activation values for each of the plurality of nodes are computed, wherein an activation value for each of the plurality of nodes is computed based on an activation function for the node;
a back-propagation of partial derivatives of an objective function through the parent neural network; and
an update of learned parameters for the parent neural network based on accumulation of estimates of the partial derivatives of the objective function summed over the plurality of training data items, wherein the learned parameters are parameters of the activation functions for the plurality of nodes; and
wherein a first set of one or more GPUs of the computer system's multiple GPUs process the forward computations and back-propagations for the first network module, and a second set of one or more GPUs of the computer system's multiple GPUs process the forward computations and back-propagations for the second network module, and
wherein iteratively training the parent neural network in the Nth iteration further comprises:
storing, in the data store, auxiliary data from an iteration prior to the Nth iteration, wherein the auxiliary data comprises datum-specific data for training data items in a training dataset for the parent neural network, wherein the datum-specific data comprise an activation value from an (N-n)th iteration for at least one of the plurality of nodes of the parent neural network for the training data items in the training dataset, where n is greater than or equal to one, and less than N; and
training the first network module in the Nth iteration, wherein training the first network module in the Nth iteration comprises, for each of the training data items in the training dataset:
determining whether training of the first network module should be skipped based on the activation value from the (N-n)th iteration stored in the auxiliary data for the training data item; and
skipping training of the first network module on the training data item upon a determination that the training of the first network module should be skipped and irrespective of whether the second network module is trained on the training data item.
2 . The method of claim 1 , wherein training the first network module in the Nth iteration comprises:
determining probabilistically whether training of the first network module should be skipped for the training data items.
3 . The method of claim 1 , wherein the N iterations comprise N epochs, such that the auxiliary data comprises data from epochs prior to the Nth epoch and the determination of whether to skip training of the first network module for training data items in the Nth epoch is based on datum-specific data from an epoch prior to the Nth epoch.
4 . The method of claim 3 , wherein determining whether to skip training of the first network module for each of the training data items in the Nth epoch is based on an estimate that a change in output of the first network module will be less than a threshold value for the training data item relative to an amount of change in the second network module of the parent neural network for the training data item.
5 . The method of claim 3 , wherein determining whether to skip training of the first network module for each of the training data items in the Nth epoch is based on whether the first network module has a stable extreme activation for the training data item for epochs prior to the Nth epoch.
6 . The method of claim 3 , wherein determining whether to skip training of the first network module for each of the training data items in the Nth epoch is based on whether the parent neural network has a stable output for the training data item for epochs prior to the Nth epoch.
7 . The method of claim 3 , wherein determining whether to skip training of the first network module for each of the training data items in the Nth epoch is based in part on a magnitude of a partial derivative of an error-loss function for the parent neural network with respect to an output of the first network module for the training data item.
8 . The method of claim 1 , further comprising computing forward propagation activation values for the first network module for the plurality of training data items in the Nth iteration but not computing back-propagation values for the first network module for the plurality of training data items in the Nth iteration.
9 . The method of claim 1 , wherein:
iteratively training the parent neural network comprises training the parent neural network in iterations prior to the Nth iteration according to a parent network objective function; and
training the first network module in the Nth iteration using a first network module-specific objective function for training the first network module in the Nth iteration.
10 . The method of claim 9 , further comprising, by the computer system, determining the first network module-specific objective function.
11 . The method of claim 1 , wherein each network module comprises a connected subnetwork of the parent neural network, wherein each subnetwork comprises one or more nodes and one or more arcs.
12 . The method of claim 1 , wherein:
iteratively training the parent network comprises training the second network module in the Nth iteration, wherein training the second network module in the Nth iteration comprises:
determining whether training of the second network module should be skipped for each of the training data items in the training dataset based on the activation values from the (N-n)th iteration stored in the auxiliary data for the training data item; and
skipping training of the second network module on the training data item upon a determination that the training of the second network module should be skipped irrespective of whether the first network module is trained on the training data item.
13 . The method of claim 12 , wherein:
the N iterations comprise N epochs, such that the auxiliary data comprises data from epochs prior to the Nth epoch;
the determination of whether to skip training of the first network module for training data items in the Nth epoch is based on datum-specific data for the training data items from an epoch prior to the Nth epoch; and
the determination of whether to skip training of the second network module for training data items in the Nth epoch is based on datum-specific data for the training data items from an epoch prior to the Nth epoch.
14 . The method of claim 12 , wherein the steps of training the first and second network modules are performed simultaneously by the computer system.
15 . The method of claim 1 , wherein the first network module comprises a front-end first network module of the parent neural network.
16 . The method of claim 15 , wherein training the front-end first network module in the Nth iteration comprises training the front-end first network module in the Nth iteration using a training method other than stochastic gradient descent.
17 . The method of claim 1 , wherein the first network module comprises a back-end module of the parent neural network.
18 . The method of claim 1 , wherein the first network module comprises a hidden module of the parent neural network.
19 . The method of claim 1 , wherein each of the one or more nodes of the first network module are on a same layer of the parent neural network.
20 . The method of claim 1 , wherein:
the parent neural network comprises a plurality of layers; and
the first network module comprises at least:
a first node on a first layer of the parent neural network; and
a second node on a second layer of the parent neural network.
21 . The method of claim 1 , wherein the plurality of network modules of the parent neural network comprises a configuration of network modules, wherein the configuration comprises:
a first set of two or more front-end modules, wherein the computer system selects members of the first set from a first collection of three of more front-end modules; and
a second set of two or more back-end modules, wherein the computer system selects members of the second set from a second collection of three of more back-end modules.
22 . The method of claim 21 , wherein:
the computer system comprises a learning coach computer system; and
the method further comprises testing, by the learning coach computer system, different configurations for the parent neural network on training data.
23 . The method of claim 21 , wherein:
the computer system comprises a learning coach computer system; and
the method further comprises selecting, by the learning coach computer system, the members of the first and second sets based on performance of the front-end modules in the first collection and of the back-end modules in the second collection on development data.
24 . The method of claim 1 , further comprising adding a node to the first network module for the Nth iteration.
25 . The method of claim 24 , wherein the node added to the first network module comprises an error judgement node.
26 . The method of claim 1 , further comprising removing a node from the first network module for the Nth iteration.
27 . The method of claim 1 , further comprising splitting the first network module into multiple network modules for the Nth iteration.
28 . The method of claim 1 , further comprising:
splitting the data in the training dataset into a plurality of disjoint datasets; and
training each of the network modules of the parent neural network with a respective one of the disjoint datasets.
29 . The method of claim 1 , further comprising increasing regularization for the training of the first network module in the Nth iteration.
30 . The method of claim 1 , wherein the auxiliary data comprises an activation value from the (N-n)th iteration for at least one of the one or more nodes of the first network module for the training data items in the training dataset.
31 . The method of claim 1 , wherein the auxiliary data comprises an activation value, for the training data items, from the (N-n)th iteration for a node of the parent neural network that provides an input to the first network module.
32 . The method of claim 1 , wherein the auxiliary data comprises an estimated activation value for each training data item from the (N-n)th iteration of a source node in the parent neural network that is a source for a node in the first network module.
33 . The method of claim 1 , wherein the auxiliary data further comprises an estimate of variability for datum-specific data.
34 . The method of claim 1 , wherein:
the plurality of network modules comprises:
a plurality of front-end modules; and
a plurality of back-end modules; and
the first network module is one of the plurality of front-end modules.
35 . The method of claim 34 , wherein:
the plurality of front-end modules comprises a plurality of feature detector front-end modules; and
the plurality of back-end modules comprises a plurality of classifier back-end modules.
36 . The method of claim 35 , wherein the parent neural network further comprises a feature vector hidden layer between the plurality of feature detector front-end modules and the plurality of classifier back-end modules, wherein the feature vector hidden layer comprises a plurality of feature nodes.
37 . The method of claim 35 , wherein the parent neural network further comprises one or more interface modules between the plurality of feature detector front-end modules and the plurality of classifier back-end modules.
38 . The method of claim 37 , wherein the front-end modules comprise object detector front-end modules.
39 . The method of claim 37 , wherein the front-end modules comprise event detector front-end modules.
40 . The method of claim 37 , wherein the plurality of back-end modules comprises a softmax back-end module.
41 . The method of claim 37 , wherein the plurality of back-end modules comprises an ontology back-end module.
42 . The method of claim 37 , wherein the plurality of back-end modules comprises a mereology back-end module.
43 . The method of claim 37 , wherein the plurality of back-end modules comprises a back-end recurrent network module.
44 . The method of claim 1 , wherein:
the parent neural network comprises:
a plurality of ensemble members; and
one or more combining networks; and
the first network module is one of the plurality of ensemble members or one of the one or more combining networks.
45 . A computer system for training a parent neural network, wherein the parent neural network comprises a plurality of network modules and a plurality of nodes, including a first network module that comprises one or more nodes and a second network module that comprises one or more nodes, wherein both the first network module and the second network module are different subsets of the parent neural network, the computer system comprising:
a data store;
multiple GPUs; and
a memory in communication with the multiple GPUs, wherein in the memory stores computer instructions that when executed by the multiple GPUs cause the multiple GPUs to iteratively train, through machine learning, the parent neural network over N iterations, where Nis greater than one, wherein each of the N iterations comprises a plurality of training data items, wherein each of the training data items comprises a vector of numerical values, and wherein the computer instructions, when executed by the multiple GPUs, cause the multiple GPUs to train the parent neural network via iterative training that comprises, for each of the N iterations:
a forward computation, for each of the plurality of training data items for the iteration through the parent neural network in which activation values for each of the plurality of nodes are computed, wherein an activation value for each of the plurality of nodes is computed based on an activation function for the node;
a back-propagation of partial derivatives of an objective function through the parent neural network; and
an update of learned parameters for the parent neural network based on accumulation of estimates of the partial derivatives of the objective function summed over the plurality of training data items, wherein the learned parameters are parameters of the activation functions for the plurality of nodes, such that a first set of one or more GPUs of the multiple GPUs process the forward computations and back-propagations for the first network module, and a second set of one or more GPUs of the multiple GPUs process the forward computations and back-propagations for the second network module; and
wherein the computer instructions, when executed by the multiple GPUs, further cause the multiple GPUs to train the parent neural network by:
storing auxiliary data in the data store from an iteration prior to the Nth iteration, wherein the auxiliary data comprises datum-specific data for training data items in a training dataset for the parent neural network, wherein the datum-specific data comprise an activation value from an (N-n)th iteration for at least one of the plurality of nodes of the parent neural network for the training data items in the training dataset, where n is greater than or equal to one, and less than N; and
training, by the first set of one or more GPUs, the first network module in the Nth iteration by, for each of the training data items:
determining whether training of the first network module should be skipped based on the activation value from the (N-n)th iteration stored in the auxiliary data for the training data item; and
skipping training of the first network module on the training data item upon a determination that the training of the first network module should be skipped and irrespective of whether the second network module is trained on the training data item.
46 . The computer system of claim 45 , wherein the N iterations comprise N epochs, such that the auxiliary data comprises data from epochs prior to the Nth epoch and the determination of whether to skip training of the first network module for training data items in the Nth epoch is based on datum-specific data from an epoch prior to the Nth epoch.
47 . The computer system of claim 45 , wherein determining whether to skip training of the first network module for training data items in the Nth epoch is based on an estimate that a change in output of the first network module will be less than a threshold value for the training data items relative to an amount of change in other network modules of the parent neural network.
48 . The computer system of claim 45 , wherein determining whether to skip training of the first network module-for training data items in the Nth epoch is based on whether the first network module has a stable extreme activation for the training data items for epochs prior to the Nth epoch.
49 . The computer system of claim 45 , wherein determining whether to skip training of the first network module for training data items in the Nth epoch is based on whether the parent neural network has a stable output for the training data items for epochs prior to the Nth epoch.
50 . The computer system of claim 45 , wherein determining whether to skip training of the first network module for training data items in the Nth epoch is based in part on a magnitude of a partial derivative of an error-loss function for the parent neural network with respect to an output of the first network module.
51 . The computer system of claim 45 , wherein the multiple GPUs are programmed further to compute forward propagation activation values for the first network module for the training data items in the Nth iteration but not computing back-propagation values for the first network module for certain training data items in the Nth iteration.
52 . The computer system of claim 45 , wherein the multiple GPUs are configured:
to train the parent neural network in iterations prior to the Nth iteration according to a parent network objective function; and
to train the first network module with a first network module-specific objective function in the Nth iteration.
53 . The computer system of claim 45 , wherein:
the second set of one or more GPUs are further configured to train the second network module in the Nth iteration by, for each of the training data items in the training dataset:
determining whether training of the second network module should be skipped based on the activation value from the (N-n)th iteration stored in the auxiliary data for the training data item; and
skipping training of the second network module on the training data item upon a determination that the training of the second network module should be skipped and irrespective of whether the first network module is trained on the training data item.
54 . The computer system of claim 53 , wherein:
the N iterations comprise N epochs, such that the auxiliary data comprises data from epochs prior to the Nth epoch;
the first set of one or more GPU are configured to determine whether to skip training of the first network module for the training data items in the Nth epoch based on datum-specific data for the training data items from an epoch prior to the Nth epoch; and
the first set of one or more GPU are configured to determine whether to skip training the second network module for the training data items in the Nth epoch based on datum-specific data for the training data items from an epoch prior to the Nth epoch.
55 . The computer system of claim 53 , wherein the steps of training the first and second network modules are performed simultaneously by the computer system.
56 . The computer system of claim 45 , wherein the plurality of network modules of the parent neural network comprises a configuration of network modules, wherein the configuration comprises:
a first set of two or more front-end modules, wherein the computer system selects members of the first set from a first collection of three of more front-end modules; and
a second set of two or more back-end modules, wherein the computer system selects members of the second set from a second collection of three of more back-end modules.
57 . The computer system of claim 56 , wherein:
the computer system comprises a learning coach computer system; and
the learning coach computer system is configured to test different configurations for the parent neural network on training data.
58 . The computer system of claim 56 , wherein:
the computer system comprises a learning coach computer system; and
the learning coach computer system is configured to select the members of the first and second sets based on performance of the front-end modules in the first collection and of the back-end modules in the second collection on development data.
59 . The computer system of claim 45 , wherein the first set of one or more GPUs is configured to add a node to the first network module in the Nth iteration.
60 . The computer system of claim 59 , wherein the node added to the first network module comprises an error judgement node.
61 . The computer system of claim 45 , wherein the first set of one or more GPUs is configured to remove a node from the first network module in the Nth iteration.
62 . The computer system of claim 45 , wherein the first set of one or more GPUs is configured to split the first network module into multiple network modules in the Nth iteration.
63 . The computer system of claim 45 , wherein the multiple GPUs are configured to:
split the data in the training dataset into a plurality of disjoint datasets; and
train each of the network modules of the parent neural network with a respective one of the disjoint datasets.
64 . The computer system of claim 45 , wherein the first set of one or more GPUs is configured to increase regularization for the training of the first network module in the Nth iteration.
65 . The computer system of claim 45 , wherein the auxiliary data comprises an activation value from the (N-n)th iteration for at least one of the one or more nodes of the first network module for the training data items.
66 . The computer system of claim 45 , wherein the auxiliary data comprises an activation value, for each training data item, from the (N-n)th iteration for a node of the parent neural network that provides an input to the first network module.
67 . The computer system of claim 45 , wherein the auxiliary data comprises an estimated activation value for each training data item from the (N-n)th iteration of a source node in the parent neural network that is a source for a node in the first network module.
68 . The computer system of claim 45 , wherein the auxiliary data further comprises an estimate of variability for datum-specific data.
69 . The computer system of claim 45 , wherein:
the plurality of network modules comprises:
a plurality of front-end modules; and
a plurality of back-end modules; and
the first network module is one of the plurality of front-end modules.
70 . The computer system of claim 69 , wherein:
the plurality of front-end modules comprises a plurality of feature detector front-end modules; and
the plurality of back-end modules comprises a plurality of classifier back-end modules.
71 . The computer system of claim 70 , wherein the parent neural network further comprises a feature vector hidden layer between the plurality of feature detector front-end modules and the plurality of classifier back-end modules, wherein the feature vector hidden layer comprises a plurality of feature nodes.
72 . The computer system of claim 70 , wherein the parent neural network further comprises one or more interface modules between the plurality of feature detector front-end modules and the plurality of classifier back-end modules.
73 . The computer system of claim 72 , wherein the front-end modules comprise object detector front-end modules.
74 . The computer system of claim 72 , wherein the front-end modules comprise event detector front-end modules.
75 . The computer system of claim 72 , wherein the plurality of back-end modules comprises a softmax back-end module.
76 . The computer system of claim 72 , wherein the plurality of back-end modules comprises an ontology back-end module.
77 . The computer system of claim 72 , wherein the plurality of back-end modules comprises a mereology back-end module.
78 . The computer system of claim 72 , wherein the plurality of back-end modules comprises a back-end recurrent network module.
79 . The computer system of claim 45 , wherein:
the parent neural network comprises:
a plurality of ensemble members; and
one or more combining networks; and
the first network module is one of the plurality of ensemble members or one of the one or more combining networks.