Companion analysis network in deep learning
Systems and methods analyze training of a first machine learning system with a second machine learning system. The first machine learning system comprises a neural network with a first inner layer node. The method includes connecting the first machine learning system to an input of the second machine learning system. The second machine learning system comprises a second objective function for analyzing an internal characteristic of the first machine learning system and which is different from a first objective function for the first machine learning system. The method further includes providing a training data item to the first machine learning system, collecting internal characteristic data from the first inner layer node of the first machine learning system associated with the internal characteristic, computing partial derivatives of the first objective function through the first machine learning system with respect to the training data item, and computing partial derivatives of the second objective function through both the second machine learning system and the first machine learning system with respect to the collected internal characteristic data.
1. A computer-implemented method for analyzing training of a first machine learning system with a second machine learning system, the first machine learning system comprising a first objective function, the method comprising:
connecting, by a computer system, the first machine learning system to an input of the second machine learning system, wherein:
the first machine learning system comprises a first neural network;
the first machine learning system comprises an input layer, an output layer, and one or more inner layers between the input and output layers;
each of the one or more inner layers comprises one or more nodes, such that the first machine learning system comprises at least a first inner layer node that is on one of the one or more inner layers of the first neural network;
the second machine learning system comprises a second objective function for analyzing an internal characteristic of the first machine learning system; and
the second objective function is different from the first objective function; and
iteratively training the first machine learning system through multiple iterations, wherein iteratively training the first machine learning system comprises:
for each iteration:
providing, by the computer system, a training data item to the first machine learning system;
collecting, by the computer system, internal characteristic data from the first inner layer node of the first machine learning system associated with the internal characteristic of the first machine learning system for the training data item;
computing, by the computer system, partial derivatives of the first objective function through the first machine learning system with respect to the training data item;
computing, by the computer system, partial derivatives of the second objective function through both the second machine learning system and the first machine learning system with respect to the collected internal characteristic data for the training data item; and
computing, by the computer system, combined partial derivatives for the first machine learning system, wherein the combined partial derivatives comprise a combination of the partial derivatives of the first objective function through the first machine learning system for the training data item and the partial derivatives of the second objective function through the first machine learning system; and
after the multiple iterations:
computing, by the computer system, accumulated partial derivatives for the first machine learning system, wherein the accumulated partial derivatives are computed from the combined partial derivatives across the training data items used in the multiple iterations; and
updating, by the computer system, learned parameters of the first machine learning system based on the accumulated partial derivatives.
2. The computer-implemented method of claim 1 , wherein connecting the first machine learning system to the second machine learning system comprises:
adding, by the computer system, an additional output node to the first machine learning system; and
connecting, by the computer system, the additional output node to the input of the second machine learning system.
3. The computer-implemented method of claim 1 , wherein connecting the first machine learning system to the second machine learning system comprises:
connecting, by the computer system, an internal element of the first machine learning system to the input of the second machine learning system.
4. The computer-implemented method of claim 3 , wherein the internal element comprises the first inner layer node of the neural network.
5. The computer-implemented method of claim 1 , wherein the internal characteristic data from the first inner layer node comprises a latent variable.
6. The computer-implemented method of claim 5 , wherein:
the latent variable comprises an activation value of the first inner layer node.
7. The computer-implemented method of claim 1 , wherein the internal characteristic data from the first inner layer node comprises a derivative calculated with respect to a parameter of the first inner layer node of the first machine learning system.
8. The computer-implemented method of claim 7 , wherein:
the derivative comprises a partial derivative of the first objective function with respect to an activation value of the first inner layer node.
9. The computer-implemented method of claim 7 , wherein:
the derivative comprises a partial derivative of the first objective function with respect to an input to the first inner layer node.
10. The computer-implemented method of claim 7 , wherein:
the neural network comprises a second node, and an arc from the second node to the first inner layer node; and
the derivative comprises a partial derivative of the first objective function with respect to a connection weight of the arc.
11. The computer-implemented method of claim 1 , wherein the internal characteristic data from the first inner layer node comprises a satisfaction score sat(n) defined as:
sat( n )=sign( A ( n )−0.5)*sign(−δ( n )),
where A(n) is an output activation of the first inner layer node of the first machine learning system for the data item, where “sign” represents a sigmoid function.
12. The computer-implemented method of claim 1 , wherein the internal characteristic data from the first inner layer node comprises a comparison between a first value of a latent variable for the first inner layer node for a current training iteration and a second value of the latent variable for a prior training iteration of the first machine learning system.
13. The computer-implemented method of claim 1 , wherein the internal characteristic data from the first inner layer node comprises a comparison between a first derivative calculated with respect to a parameter of the first inner layer node of the first machine learning system for a current training iteration and a second derivative calculated with respect to the parameter of the first inner layer node of the first machine learning system for a prior training iteration of the first machine learning system.
14. The computer-implemented method of claim 1 , wherein the second machine learning system comprises a second neural network.
15. The computer-implemented method of claim 14 , wherein computing partial derivatives of the first objective function and computing partial derivatives of the second objective function each comprise performing a back propagation calculation.
16. The computer-implemented method of claim 1 , wherein computing partial derivatives of the first objective function and computing partial derivatives of the second objective function each comprise performing a numerical estimation calculation.
17. The computer-implemented method of claim 1 , wherein the internal characteristic of the first machine learning system is selected to cause the second machine learning system to alter an output of the first machine learning system when updating the learned parameters of the first machine learning system.
18. A computer system for analyzing training of a first machine learning system with a second machine learning system, the computer system comprising:
a processor; and
a memory coupled to the processor, the memory storing:
the first machine learning system comprising a first objective function, wherein:
the first machine learning system comprises a first neural network;
the machine learning system comprises an input layer, an output layer, and one or more inner layers between the input and output layers;
each of the one or more layers comprises one or more nodes, such that the first machine learning system comprises at least a first inner layer node;
the second machine learning system comprising a second objective function, wherein:
the second objective function is for analyzing an internal characteristics of the first machine learning system; and
the second objective function is different from the first objective function; and
instructions that, when executed by the processor, cause the computer system to:
connect the first machine learning system to an input of the second machine learning system; and
iteratively train the first machine learning system through multiple iterations by:
for each iteration
providing a training data item to the first machine learning system;
collecting internal characteristic data from the first inner layer node of the first machine learning system associated with the internal characteristic of the first machine learning system for the training data item;
computing partial derivatives of the first objective function through the first machine learning system with respect to the training data item;
computing partial derivatives of the second objective function through both the second machine learning system and the first machine learning system with respect to the collected internal characteristic data for the training data item;
computing combined partial derivatives for the first machine learning system, wherein the combined partial derivatives comprise a combination of the partial derivatives of the first objective function through the first machine learning system for the training data item and the partial derivatives of the second objective function through the first machine learning system; and
after the multiple iterations:
computing accumulated partial derivatives for the first machine learning system, wherein the accumulated partial derivatives are computed from the combined partial derivatives for the first machine learning system across the training data items used in the multiple iterations; and
updating learned parameters of the first machine learning system based on the accumulated partial derivatives.
19. The computer system of claim 18 , wherein the instructions cause the computer system to connect the first machine learning system to the second machine learning system by causing the computer system to:
add an additional output node to the first machine learning system; and
connect the additional output node to the input of the second machine learning system.
20. The computer system of claim 18 , wherein the instructions cause the computer system to connect the first machine learning system to the second machine learning system by causing the computer system to:
connect an internal element of the first machine learning system to the input of the second machine learning system.
21. The computer system of claim 20 , wherein the internal element comprises the first inner layer node of the neural network.
22. The computer system of claim 18 , wherein the internal characteristic data comprises a latent variable.
23. The computer system of claim 22 , wherein:
the latent variable comprises an activation value of the first inner layer node.
24. The computer system of claim 18 , wherein the internal characteristic data comprises a derivative calculated with respect to a parameter of the first inner layer node of the first machine learning system.
25. The computer system of claim 24 , wherein:
the derivative comprises a partial derivative of the first objective function with respect to an activation value of the first inner layer node.
26. The computer system of claim 24 , wherein:
the derivative comprises a partial derivative of the first objective function with respect to an input to the first inner layer node.
27. The computer system of claim 24 , wherein:
the neural network comprises a second node, and an arc from the second node to the first inner layer node; and
the derivative comprises a partial derivative of the first objective function with respect to a connection weight of the arc.
28. The computer system of claim 18 , wherein the internal characteristic data comprises a satisfaction score sat(n) defined as:
sat( n )=sign( A ( n )−0.5)*sign(−δ( n )),
where A(n) is an output activation of the first inner layer node of the first machine learning system for the data item, wherein “sign” represents a sigmoid function.
29. The computer system of claim 18 , wherein the internal characteristic data comprises a comparison between a first value of a latent variable for the first inner layer node for a current training iteration and a second value of the latent variable for the first inner layer node for a prior training iteration of the first machine learning system.
30. The computer system of claim 18 , wherein the internal characteristic data comprises a comparison between a first derivative calculated with respect to a parameter of the first inner layer node of the first machine learning system for a current training iteration and a second derivative calculated with respect to the parameter of the first inner layer node of the machine learning system for a prior training iteration of the first machine learning system.
31. The computer system of claim 18 , wherein the second machine learning system comprises a second neural network.
32. The computer system of claim 31 , wherein the instructions cause the computer system to compute partial derivatives of the first objective function and compute partial derivatives of the second objective function via a back propagation calculation.
33. The computer system of claim 18 , wherein the instructions cause the computer system to compute partial derivatives of the first objective function and compute partial derivatives of the second objective function via a numerical estimation calculation.
34. The computer system of claim 18 , wherein the internal characteristic is selected to cause the second machine learning system to alter an output of the first machine learning system when updating the learned parameters of the first machine learning system.
35. A computer-implemented method for analyzing a first machine learning system via a second machine learning system, the first machine learning system comprising a first objective function, the method comprising:
connecting, by a computer system, the first machine learning system to an input of the second machine learning system, wherein the first machine learning system comprises a neural network comprising a plurality of output nodes;
wherein the second machine learning system comprises a second objective function for analyzing an internal characteristic of the first machine learning system;
providing, by the computer system, a data item to the first machine learning system;
collecting, by the computer system, internal characteristic data from the first machine learning system associated with the internal characteristic of the first machine learning system, wherein:
the internal characteristic data comprises a derivative calculated with respect to a parameter of the first machine learning system; and
the derivative comprises a partial derivative of a difference between a largest activation and a second largest activation of the plurality of output nodes for the data item;
computing, by the computer system, partial derivatives of the first objective function through the first machine learning system with respect to the data item; and
computing, by the computer system, partial derivatives of the second objective function through both the second machine learning system and the first machine learning system with respect to the collected internal characteristic data.
36. A computer-implemented method for analyzing a first machine learning system via a second machine learning system, the first machine learning system comprising a first objective function, the method comprising:
connecting, by a computer system, the first machine learning system to an input of the second machine learning system;
wherein the second machine learning system comprises a second objective function for analyzing an internal characteristic of the first machine learning system;
providing, by the computer system, a data item to the first machine learning system;
collecting, by the computer system, internal characteristic data from the first machine learning system associated with the internal characteristic of the first machine learning system, wherein the internal characteristic data comprises a satisfaction score sat(n) defined as:
sat( n )=sign( A ( n )−0.5)*sign(−δ( n )),
where A(n) is an output activation of a node n of the first machine learning system for the data item and sign represents a sigmoid function;
computing, by the computer system, partial derivatives of the first objective function through the first machine learning system with respect to the data item; and
computing, by the computer system, partial derivatives of the second objective function through both the second machine learning system and the first machine learning system with respect to the collected internal characteristic data.
37. A computer system for analyzing a first machine learning system via a second machine learning system, the computer system comprising:
a processor; and
a memory coupled to the processor, the memory storing:
the first machine learning system comprising a first objective function, wherein the first machine learning system comprises a neural network comprising a plurality of output nodes;
the second machine learning system comprising a second objective function; and
instructions that, when executed by the processor, cause the computer system to:
connect the first machine learning system to an input of the second machine learning system;
wherein the second objective function is for analyzing an internal characteristic of the first machine learning system;
provide a data item to the first machine learning system;
collect internal characteristic data from the first machine learning system associated with the internal characteristic of the first machine learning system, wherein:
the internal characteristic data comprises a derivative calculated with respect to a parameter of the first machine learning system; and
the derivative comprises a partial derivative of a difference between a largest activation and a second largest activation of the plurality of output nodes for the data item;
compute partial derivatives of the first objective function through the first machine learning system with respect to the data item; and
compute partial derivatives of the second objective function through both the second machine learning system and the first machine learning system with respect to the collected internal characteristic data.
38. A computer system for analyzing a first machine learning system via a second machine learning system, the computer system comprising:
a processor; and
a memory coupled to the processor, the memory storing:
the first machine learning system comprising a first objective function;
the second machine learning system comprising a second objective function; and
instructions that, when executed by the processor, cause the computer system to:
connect the first machine learning system to an input of the second machine learning system;
wherein the second objective function is for analyzing an internal characteristic of the first machine learning system;
provide a data item to the first machine learning system;
collect internal characteristic data from the first machine learning system associated with the internal characteristic of the first machine learning system, wherein the internal characteristic data comprise a satisfaction score sat(n) defined as:
sat( n )=sign( A ( n )−0.5)*sign(−δ( n )),
where A(n) is an output activation of a node n of the first machine learning system for the data item and sign represents a sigmoid function;
compute partial derivatives of the first objective function through the first machine learning system with respect to the data item; and
compute partial derivatives of the second objective function through both the second machine learning system and the first machine learning system with respect to the collected internal characteristic data.
39. The method of claim 1 , wherein the combined partial derivatives comprise a weighted average of the partial derivatives of the first objective function through the first machine learning system for the training data item and the partial derivatives of the second objective function through the first machine learning system.
40. The method of claim 39 , wherein weighted average comprises:
the partial derivatives of the first objective function through the first machine learning system weighted by a first hyperparameter; and
the partial derivatives of the second objective function through the first machine learning system weighted by a second hyperparameter.
41. The computer system of claim 18 , wherein the combined partial derivatives comprise a weighted average of the partial derivatives of the first objective function through the first machine learning system for the training data item and the partial derivatives of the second objective function through the first machine learning system.
42. The computer system of claim 41 , wherein weighted average comprises:
the partial derivatives of the first objective function through the first machine learning system weighted by a first hyperparameter; and
the partial derivatives of the second objective function through the first machine learning system weighted by a second hyperparameter.