Multimodal data inference
Computer-implemented methods are provided for generating machine learning model for multimodal data inference tasks. Such a method includes, for each sample in a training dataset of multimodal data samples, encoding the sample to produce a compressed vector representation of the sample in a k-dimensional latent space, and perturbing features of the sample to identify, for each dimension of the latent space, a set of active features perturbation of each of which produces more than a threshold change in the vector representation in that dimension. The method further comprises generating a sample graph having nodes interconnected by edges, wherein the nodes comprise nodes representing respective said features of the sample and edges interconnecting nodes indicate the active features for each dimension. The sample graph is then used to train a graph neural network model to perform the multimodal data inference task. Multimodal data inference systems employing such models are also provided.
1 . A computer-implemented method for generating a machine learning model for a multimodal data inference task, the method comprising, for each sample in a training dataset including a plurality of data samples having different modalities:
encoding the sample to produce a compressed vector representation of the sample in a k-dimensional latent space;
perturbing features of the sample to identify, for each of a plurality of dimensions of the k-dimensional latent space, a set of active features, perturbation of each of which produces more than a threshold change in the compressed vector representation in that dimension;
generating an active feature map based on the set of active features using each of the plurality of dimensions of the k-dimensional latent space;
creating a node map having a plurality of sides, each side of the node map including a plurality of nodes from a respective one of the different modalities of the training dataset, wherein the nodes represent respective features of the sample;
for each of the plurality of dimensions of the k-dimensional latent space and based on the active feature map, creating a sub-sample graph defining edges interconnecting the nodes of the node map for that dimension, wherein the edges indicate the active features for each dimension;
generating a sample graph by superimposing each of the sub-sample graphs for each of the plurality of dimensions; and
using the sample graph to train a graph neural network model to perform the multimodal data inference task.
2 . The computer-implemented method of claim 1 , further comprising:
after training the graph neural network model, performing the multimodal data inference task for a second multimodal data sample by generating a second sample graph for the second multimodal data sample and supplying that second sample graph to the graph neural network model to obtain an inference output for the second multimodal data sample.
3 . The computer-implemented method of claim 1 , wherein encoding the sample to produce the compressed vector representation comprises:
for each modality of data in the sample, encoding that data using an intra-modality encoder, pretrained in an autoencoder architecture for that data modality, to produce a latent feature vector;
concatenating the latent feature vectors for the sample to obtain a multimodal feature vector; and
encoding the multimodal feature vector using an inter-modality encoder, pretrained in an autoencoder architecture for such multimodal feature vectors, to produce the compressed vector representation of the sample.
4 . The computer-implemented method of claim 3 , further comprising:
perturbing features of the multimodal feature vector for the sample to identify the set of active features for each dimension.
5 . The computer-implemented method of claim 4 , further comprising:
perturbing features of the multimodal feature vector by setting each feature in turn to zero.
6 . The computer-implemented method of claim 3 , further comprising:
pre-training the intra-modality encoder and the inter-modality encoder in the respective autoencoder architectures.
7 . The computer-implemented method of claim 3 , further comprising:
during training of the graph neural network, further training the intra-modality encoder and the inter-modality encoder in dependence on an output error of the graph neural network.
8 . The computer-implemented method of claim 1 , wherein edges indicating active features for different dimensions of the k-dimensional latent space are distinguished in the sample graph.
9 . The computer-implemented method of claim 8 , wherein:
the sample graph includes a node representing each dimension of the k-dimensional latent space;
edges indicating active features for each dimension interconnect the node representing that dimension with nodes representing those active features; and
the sample graph includes additional edges interconnecting the nodes representing the dimensions.
10 . The computer-implemented method of claim 9 , wherein a weight is defined in the sample graph for each edge interconnecting a node representing a dimension and a node representing an active feature, the weight being dependent on the amount of change in the vector representation produced by perturbation of that active feature.
11 . The computer-implemented method of claim 1 , wherein, for each dimension of the k-dimensional latent space, nodes representing active features for that dimension are interconnected by edges in the sample graph.
12 . The computer-implemented method of claim 11 , wherein different types of edges are defined in the sample graph to indicate active features for different dimensions of the k-dimensional latent space.
13 . The computer-implemented method of claim 1 , wherein each multimodal data sample comprises multimodal medical data for a patient, and wherein the multimodal data inference task provides a medical evaluation.
14 . The computer-implemented method of claim 1 , wherein each multimodal data sample comprises multimodal measurement data for a physical system, and wherein the multimodal data inference task comprises evaluating a state of the system.
15 . A computer program product for generating a machine learning model for a multimodal data inference task, the computer program product comprising a computer readable storage medium having program instructions embodied therein, the program instructions being executable by a computing system to cause the computing system, for each sample in a training dataset of multimodal data samples, to:
encode the sample to produce a compressed vector representation of the sample in a k-dimensional latent space;
perturb features of the sample to identify, for each of a plurality of dimensions of the k-dimensional latent space, a set of active features, perturbation of each of which produces more than a threshold change in the compressed vector representation in that dimension;
generate an active feature map based on the set of active features using each of the plurality of dimensions of the k-dimensional latent space;
create a node map having a plurality of sides, each side of the node map including a plurality of nodes from a respective one of the different modalities of the training dataset, wherein the nodes represent respective features of the sample;
for each of the plurality of dimensions of the k-dimensional latent space and based on the active feature map, create a sub-sample graph defining edges interconnecting the nodes of the node map for that dimension, wherein the edges indicate the active features for each dimension;
generate a sample graph by superimposing each of the sub-sample graphs for each of the plurality of dimensions; and
use the sample graph to train a graph neural network model to perform the multimodal data inference task.
16 . The computer program product of claim 15 , wherein the program instructions are further executable to cause the computing system, after training the graph neural network model, to perform the multimodal data inference task for a second multimodal data sample by generating a second sample graph for the second multimodal data sample and supplying that second sample graph to the graph neural network model to obtain an inference output for the second multimodal data sample.
17 . The computer program product of claim 15 , wherein the program instructions are executable to cause the computing system to encode the sample to produce the compressed vector representation by:
for each modality of data in the sample, encoding that data using an intra-modality encoder, pretrained in an autoencoder architecture for that data modality, to produce a latent feature vector;
concatenating the latent feature vectors for the sample to obtain a multimodal feature vector; and
encoding the multimodal feature vector using an inter-modality encoder, pretrained in an autoencoder architecture for such multimodal feature vectors, to produce the compressed vector representation of the sample.
18 . The computer program product of claim 17 , wherein the program instructions are executable to cause the computing system to perturb features of the multimodal feature vector for the sample to identify the set of active features for each dimension.
19 . A multimodal data inference system comprising:
a sample graph generator adapted to, for each sample in a training dataset including a plurality of data samples having different modalities:
encode an input multimodal data the sample to produce a compressed vector representation of the sample in a k-dimensional latent space,
perturb features of the sample to identify, for each of a plurality of dimensions of the k-dimensional latent space, a set of active features, perturbation of each of which produces more than a threshold change in the compressed vector representation in that dimension,
generate an active feature map based on the set of active features using each of the plurality of dimensions of the k-dimensional latent space;
create a node map having a plurality of sides, each side of the node map including a plurality of nodes from a respective one of the different modalities of the training dataset, wherein the nodes represent respective features of the sample;
for each of the plurality of dimensions of the k-dimensional latent space and based on the active feature map, create a sub-sample graph defining edges interconnecting the nodes of the node map for that dimension, wherein the edges indicate the active features for each dimension;
generate a sample graph by superimposing each of the sub-sample graphs for each of the plurality of dimensions; and
a graph neural network model, pretrained to perform a multimodal data inference task using sample graphs generated by the sample graph generator for respective samples in a training dataset of multimodal data samples;
wherein the sample graph generator is adapted to supply the sample graph for the input multimodal data sample to the graph neural network model to obtain an inference output for the sample.
20 . The system of claim 19 , wherein the sample graph generator is adapted to encode the input sample to produce the compressed vector representation by:
for each modality of data in the sample, encoding that data using an intra-modality encoder, pretrained in an autoencoder architecture for that data modality, to produce a latent feature vector;
concatenating the latent feature vectors for the sample to obtain a multimodal feature vector; and
encoding the multimodal feature vector using an inter-modality encoder, pretrained in an autoencoder architecture for such multimodal feature vectors, to produce the compressed vector representation of the sample.