Information processing method, information processing apparatus, and computer readable storage medium

1. An information processing method comprising:

reading a layer structure and parameters of layers from each model of a respective one of two neural networks, wherein one of the models is modified by changing a structure of a plurality of filters while keeping a plurality of effects of a weight filter, the changing of the structure of the filters comprising performing at least one of increasing a filter size, changing a sequence of the filters, adding another filter, or deleting one of the filters; and

determining a degree of matching between the models of the two neural networks, by comparing layers, of the respective models of the two neural networks, that are configured as a graph-like form in respective hidden layers, in order from an input layer using breadth first search or depth first search, based on similarities between respective layers,

wherein the determining the degree of matching between the models of the two neural networks includes, when convolutional layers are compared,

estimating a true filter size with respect to each weight filter of the convolutional layers,

modifying parameters of weight filters to be compared to respective estimated true filter sizes,

expressing the convolutional layers by vector sets by regarding parameters of each weight filter as one vector, and

setting a similarity between the vector sets of the respective convolutional layers to be compared as a similarity between the convolutional layers, and

wherein the estimating the true filter size includes

accumulating absolute values of parameters of respective channels in each of the weight filters, and

estimating a minimum rectangle that includes all positions at which accumulated values are a predetermined threshold value or more, as a true filter size.

2. The information processing method according to claim 1 , wherein the determining the degree of matching between the models of the two neural networks includes, if types of layers to be compared are different, setting a similarity to 0, and not performing comparison in layers in a later stage than the layers subjected to the comparison.

3. The information processing method according to claim 1 ,

wherein the determining the degree of matching between the models of the two neural networks includes, when full-connected layers are compared,

expressing each of the full-connected layers, by regarding weights of each of the full-connected layers as a feature vector, using a vector set, and

setting a similarity between the vector sets of the respective full-connected layers to be compared as a similarity between the full-connected layers.

4. The information processing method according to claim 1 ,

wherein the similarity between the vector sets is obtained by

configuring a bipartite graph by obtaining pairs of feature vectors whose distance is a predetermined threshold value or less, and

calculating a maximum number of matches by solving a maximum matching problem from the bipartite graph.

5. The information processing method according to claim 1 ,

wherein the similarity between the vector sets is obtained by

quantizing respective feature vectors, and

obtaining a similarity between quantization histograms.

6. The information processing method according to claim 1 ,

wherein the determining the degree of matching between the models of the two neural networks includes, when activation layers are compared,

setting, if types of the activation layers are the same, and a distance between parameters is a predetermined threshold value or less, the distance as a similarity, and in other cases, 0 to the similarity.

7. The information processing method according to claim 6 , wherein the type of the activation layer is a linear connection, a sigmoid function, a hard sigmoid function, a tanh function (hyperbolic tangent function), a softsign function, a softplus function, or a ReLU (Rectified Linear Unit).

8. The information processing method according to claim 1 ,

wherein the determining the degree of matching between the models of the two neural networks includes, when pooling layers are compared,

setting, if types of the pooling layers are the same, and a distance between parameters is a predetermined threshold value or less, the distance as a similarity, and in other cases, 0 to the similarity.

9. The information processing method according to claim 8 ,

wherein the type of the pooling layer is max pooling or average pooling, and

the parameters are a filter size and an interval at which a filter is applied.

10. An information processing apparatus comprising:

one or more processors; and

one or more memory devices that store one or more programs,

wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to perform:

reading a layer structure and parameters of layers from each model of a respective one of two neural networks, wherein one of the models is modified by changing a structure of a plurality of filters while keeping a plurality of effects of a weight filter, the changing of the structure of the filters comprising performing at least one of increasing a filter size, changing a sequence of the filters, adding another filter, or deleting one of the filters; and

determining a degree of matching between the models of the two neural networks, by comparing layers, of the respective models of the two neural networks, that are configured as a graph-like form in respective hidden layers, in order from an input layer using breadth first search or depth first search, based on similarities between respective layers,

wherein the determining the degree of matching between the models of the two neural networks includes, when convolutional layers are compared,

estimating a true filter size with respect to each weight filter of the convolutional layers,

modifying parameters of weight filters to be compared to respective estimated true filter sizes,

expressing the convolutional layers by vector sets by regarding parameters of each weight filter as one vector, and

setting a similarity between the vector sets of the respective convolutional layers to be compared as a similarity between the convolutional layers, and

wherein the estimating the true filter size includes

accumulating absolute values of parameters of respective channels in each of the weight filters, and

estimating a minimum rectangle that includes all positions at which accumulated values are a predetermined threshold value or more, as a true filter size.

11. A non-transitory computer readable storage medium storing a program, the program causes, upon being executed by one or more processors of a computer, the computer to execute:

reading a layer structure and parameters of layers from each model of a respective one of two neural networks, wherein one of the models is modified by changing a structure of a plurality of filters while keeping a plurality of effects of a weight filter, the changing of the structure of the filters comprising performing at least one of increasing a filter size, changing a sequence of the filters, adding another filter, or deleting one of the filters; and

determining a degree of matching between the models of the two neural networks, by comparing layers, of the respective models of the two neural networks, that are configured as a graph-like form in respective hidden layers, in order from an input layer using breadth first search or depth first search, based on similarities between respective layers,

wherein the determining the degree of matching between the models of the two neural networks includes, when convolutional layers are compared,

estimating a true filter size with respect to each weight filter of the convolutional layers,

modifying parameters of weight filters to be compared to respective estimated true filter sizes,

expressing the convolutional layers by vector sets by regarding parameters of each weight filter as one vector, and

setting a similarity between the vector sets of the respective convolutional layers to be compared as a similarity between the convolutional layers, and

wherein the estimating the true filter size includes

accumulating absolute values of parameters of respective channels in each of the weight filters, and

estimating a minimum rectangle that includes all positions at which accumulated values are a predetermined threshold value or more, as a true filter size.