Training multimodal representation learning model on unannotated multimodal data
An example system includes a processor to receive unannotated multimodal data. The processor can estimate a probability an associated pair of different modalities in the unannotated multimodal data to be correctly associated using a multimodal similarity function and a local density estimation. The processor can also train a multimodal representation learning model on the unannotated multimodal data using the estimated probability as a weight for the associated pair in a loss function.
1 . A system comprising:
a processor;
a memory; and
program instructions stored on the memory for execution by the processor to:
receive unannotated multimodal data comprising two or more of visual modalities, auditory modalities, and language modalities, wherein the unannotated multimodal data comprises a plurality of pairs of concepts from different modalities;
estimate a respective probability that a respective pair of concepts, of the plurality of pairs of concepts, from the different modalities are of a same concept using a multimodal similarity function and a local density estimation;
detect, based on the estimation of the respective probability, noise in the unannotated multimodal data, wherein
the noise is detected based on at least one wrongly associated pair of concepts, of the plurality of pairs of concepts, that include different concepts;
integrate the detected noise into a loss function based on integration of a respective weight for the respective pair of concepts into the loss function, wherein
the respective weight corresponds to the estimated respective probability:
train, based on the integration of the detected noise into the loss function, a multimodal representation learning model on the unannotated multimodal data, wherein
the multimodal representation learning model is trained based on the respective weight for the detected noise from the at least one wrongly associated pair of concepts; and
execute, by the trained multimodal representation learning model, one or more downstream machine learning tasks involving multimodal data.
2 . The system of claim 1 , wherein the local density estimation comprises a k-nearest neighbor density estimation.
3 . The system of claim 1 , wherein the multimodal similarity function comprises a minimum of normalized similarity values for each modality of each sample pair of concepts of the plurality of pairs of concepts.
4 . The system of claim 1 , wherein the loss function comprises a soft max margin ranking loss function.
5 . A computer-implemented method, comprising:
receiving, via a processor, unannotated multimodal data comprising two or more of visual modalities, auditory modalities, and language modalities, wherein the unannotated multimodal data comprises a plurality of pairs of concepts from different modalities;
estimating, via the processor, that a respective probability of a respective pair of concepts, of the plurality of pairs of concepts, from the different modalities are of a same concept using a multimodal similarity function and a local density estimation;
detecting, based on the estimation of the respective probability, noise in the unannotated multimodal data, wherein
the noise is detected based on at least one wrongly associated pair of concepts, of the plurality of pairs of concepts, that include different concepts;
integrating the detected noise into a loss function based on integration of a respective weight for the respective pair of concepts into the loss function, wherein
the respective weight corresponds to the estimated respective probability;
training, via the processor, based on the integration of the detected noise into the loss function, a multimodal representation learning model on the unannotated multimodal data, wherein
the multimodal representation learning model is trained based on the respective weight for the detected noise from the at least one wrongly associated pair of concepts; and
executing, by the trained multimodal representation learning model, one or more downstream machine learning tasks involving multimodal data.
6 . The computer-implemented method of claim 5 , wherein the multimodal similarity function comprises calculating a minimum of normalized similarity values for each modality of each respective pair of concepts of the plurality of pairs of concepts from the different modalities in comparison with other values of corresponding modalities in the unannotated multimodal data.
7 . The computer-implemented method of claim 5 , wherein the local density estimation comprises performing a k-nearest neighbor density estimation in a multimodal space.
8 . The computer-implemented method of claim 5 , further comprising detecting, via the processor, that the respective pair of concepts from the different modalities is of the same concept based on detecting that the respective pair of concepts belongs to a cluster in a multimodal space with a density above a threshold value.
9 . The computer-implemented method of claim 5 , wherein the one or more downstream machine learning tasks further comprise: receiving a question regarding information in one or more of the visual modalities, and automatically generating an answer using the trained multimodal representation learning model.
10 . The computer-implemented method of claim 5 , wherein the one or more downstream machine learning tasks further comprise receiving a text including a description of a video to be retrieved, and automatically retrieving the video using the trained multimodal representation learning model.
11 . The computer-implemented method of claim 5 , wherein the visual modalities include extracted feature vectors representing visual features that are selected from a group consisting of edges, corners, internet points, blobs, and ridges.
12 . The computer-implemented method of claim 5 , wherein the auditory modalities include extracted feature vectors related to automatic speech recognition systems.
13 . A computer program product for training multimodal representation learning models, the computer program product comprising a computer-readable storage medium having program code embodied therewith, the program code executable by a processor to cause the processor to:
receive unannotated multimodal data comprising two or more of visual modalities, auditory modalities, and language modalities, wherein the unannotated multimodal data comprises a plurality of pairs of concepts from different modalities;
estimate a respective probability that a respective pair of concepts, of the plurality of pairs of concepts, from the different modalities are of a same concept using a multimodal similarity function and a local density estimation;
detect, based on the estimation of the respective probability, noise in the unannotated multimodal data, wherein
the noise is detected based on at least one wrongly associated pair of concepts, of the plurality of pairs of concepts, that include different concepts;
integrate the detected noise into a loss function based on integration of a respective weight for the respective pair of concepts into the loss function, wherein
the respective weight corresponds to the estimated respective probability;
train, based on the integration of the detected noise into the loss function, a multimodal representation learning model of the multimodal representation learning models on the unannotated multimodal data, wherein
the multimodal representation learning model is trained based on the respective weight for the detected noise from the at least one wrongly associated pair of concepts; and
execute, by the trained multimodal representation learning model, one or more downstream machine learning tasks involving multimodal data.
14 . The computer program product of claim 13 , wherein the program code executable by the processor further causes the processor to calculate a minimum of normalized similarity values for each modality of each respective pair of concepts of the plurality of pairs of concepts from the different modalities in comparison with other values of corresponding modalities in the unannotated multimodal data.
15 . The computer program product of claim 13 , wherein the program code executable by the processor further causes the processor to perform a k-nearest neighbor density estimation in a multimodal space.
16 . The computer program product of claim 13 , wherein the program code executable by the processor further causes the processor to detect that the respective pair of concepts from the different modalities is of the same concept based on a detection that the respective pair of concepts belongs to a cluster in a multimodal space with a density above a threshold value.
17 . The computer program product of claim 13 , wherein the one or more downstream machine learning tasks comprise reception of a question regarding information in one or more of the visual modalities of the unannotated multimodal data, and automatic generation of an answer using the trained multimodal representation learning model.
18 . The computer program product of claim 13 , wherein the one or more downstream machine learning tasks comprise reception of a text including a description of a video to be retrieved from the unannotated multimodal data, and automatic retrieval of the video using the trained multimodal representation learning model.
19 . The computer program product of claim 13 , wherein the visual modalities include extracted feature vectors representing visual features that are selected from a group consisting of edges, corners, internet points, blobs, and ridges.
20 . The computer program product of claim 13 , wherein the auditory modalities include extracted feature vectors related to automatic speech recognition systems.