Information processing apparatus, information processing method and program
A learning network includes: a first extraction unit that extracts each feature vector of different pieces of teaching data designated in a support set using a first inception neural network that extracts a feature vector of time-series data by convolution layers with different filter sizes; a second extraction unit that extracts a feature vector of teaching data designated in a query set using a second inception neural network that extracts a feature vector of time-series data by a plurality of convolution layers with different filter sizes; and a learning unit that updates a network parameter of the learning network such that a relative distance between feature vectors of teaching data of the query set and of a positive sample of the support set is small and a relative distance between feature vectors of teaching data of the query set and of a negative sample of the support set is large.
1 . An information processing apparatus that learns or classifies time-series data in a learning network,
wherein the learning network includes:
a first extraction unit, including one or more processors, configured to extract each feature vector of a plurality of different pieces of teaching data designated in a support set using a first inception neural network that extracts a feature vector of time-series data using a plurality of convolution layers, each of which performs a convolution operation with different filter sizes,
a second extraction unit, including one or more processors, configured to extract a feature vector of teaching data designated in a query set using a second inception neural network that extracts a feature vector of time-series data using a plurality of convolution layers, each of which performs a convolution operation with different filter sizes, and
a learning unit, including one or more processors, configured to update a network parameter of the learning network such that a relative distance between feature vectors of teaching data of the query set and of a positive sample which coincides with or is similar to teaching data of the query set which is teaching data of the support set is small and a relative distance between feature vectors of the teaching data of the query set and of a negative sample that does not coincide with or is not similar to the teaching data of the query set which is the teaching data of the support set is large, using each of feature vectors of the plurality of pieces of teaching data and the feature vector of the teaching data.
2 . The information processing apparatus according to claim 1 , further comprising
a classification unit, including one or more processors, configured to perform estimation calculation for a classification category of time-series data as a classification target designated in the query set for the plurality of pieces of teaching data designated in the support set, using the learning network with the updated network parameter.
3 . The information processing apparatus according to claim 1 , wherein
the second extraction unit extracts a plurality of feature vectors for one piece of teaching data designated in the query set by using the plurality of convolution layers in parallel, and
the first extraction unit extracts a plurality of feature vectors for one piece of teaching data designated in the support set by using the plurality of convolution layers in parallel, calculates each attention score based on similarity between the plurality of feature vectors extracted by the first extraction unit and the plurality of feature vectors extracted by the second extraction unit, and applies each of the attention scores to each of the plurality of feature vectors extracted by the first extraction unit.
4 . An information processing method that learns or classifies time-series data in a learning network, the learning network including:
extracting each feature vector of a plurality of different pieces of teaching data designated in a support set using a first inception neural network that extracts a feature vector of time-series data using a plurality of convolution layers, each of which performs a convolution operation with different filter sizes;
extracting a feature vector of teaching data designated in a query set using a second inception neural network that extracts a feature vector of time-series data using a plurality of convolution layers, each of which performs a convolution operation with different filter sizes; and
updating a network parameter of the learning network such that a relative distance between feature vectors of teaching data of the query set and of a positive sample which coincides with or is similar to teaching data of the query set which is teaching data of the support set is small and a relative distance between feature vectors of the teaching data of the query set and of a negative sample that does not coincide with or is not similar to the teaching data of the query set which is the teaching data of the support set is large, using each of feature vectors of the plurality of pieces of teaching data and the feature vector of the teaching data.
5 . The information processing method according to claim 4 , further comprising
performing estimation calculation for a classification category of time-series data as a classification target designated in the query set for the plurality of pieces of teaching data designated in the support set, using the learning network with the updated network parameter.
6 . A non-transitory computer-readable storage medium storing an information processing program, wherein executing of the information processing program causes one or more computers to perform operations comprising:
extracting each feature vector of a plurality of different pieces of teaching data designated in a support set using a first inception neural network that extracts a feature vector of time-series data using a plurality of convolution layers, each of which performs a convolution operation with different filter sizes;
extracting a feature vector of teaching data designated in a query set using a second inception neural network that extracts a feature vector of time-series data using a plurality of convolution layers, each of which performs a convolution operation with different filter sizes; and
updating a network parameter of the learning network such that a relative distance between feature vectors of teaching data of the query set and of a positive sample which coincides with or is similar to teaching data of the query set which is teaching data of the support set is small and a relative distance between feature vectors of the teaching data of the query set and of a negative sample that does not coincide with or is not similar to the teaching data of the query set which is the teaching data of the support set is large, using each of feature vectors of the plurality of pieces of teaching data and the feature vector of the teaching data.