Content analysis and retrieval using machine learning

An apparatus comprising:

• a processing device operatively coupled to a memory and configured to:
• train a neural network, wherein the training comprises:
• receiving an input dataset comprising training data;
• generating a distilled dataset by inputting the input dataset into a selected encoder of a plurality of encoders, the selected encoder corresponding to a format of the input dataset, wherein the training data of the input dataset comprises a plurality of first data samples, and the distilled dataset is generated by one or more machine learning models and comprises a plurality of second data samples based on an output of the selected encoder, wherein the plurality of second data samples is smaller than the plurality of first data samples, and wherein mutual information between the plurality of first data samples and the plurality of second data samples is maximized;
• and
• computing a loss to reduce similarities between the plurality of second data samples to make the plurality of second data samples orthogonal to each other;
• calculating a relation between the plurality of second data samples and the training data to obtain a probabilistic bag of images representation for each of the plurality of second data samples;
• and
• generating a label for each of the plurality of second data samples based on the probabilistic bag of images representation;
• receive, at the trained neural network, an input data sample;
• compute, using the trained neural network, similarities between the input data sample and respective ones of the plurality of second data samples based on the labels of the plurality of second data samples;
• select, using the trained neural network, one or more second data samples of the plurality of second data samples based, at least in part, on the computed similarities;
• and
• retrieve, using the trained neural network, one or more first data samples of the plurality of first data samples from the input dataset based, at least in part, on the selected one or more second data samples.

An article of manufacture comprising a non-transitory processor-readable storage medium having stored therein program code of one or more software programs, wherein the program code when executed by at least one processing device causes said at least one processing device to perform the steps of:

• training a neural network, wherein the training comprises:
• receiving an input dataset comprising training data;
• generating a distilled dataset by inputting the input dataset into a selected encoder of a plurality of encoders, the selected encoder corresponding to a format of the input dataset, wherein the training data of the input dataset comprises a plurality of first data samples, and the distilled dataset is generated by one or more machine learning models and comprises a plurality of second data samples based on an output of the selected encoder, wherein the plurality of second data samples is smaller than the plurality of first data samples, and wherein mutual information between the plurality of first data samples and the plurality of second data samples is maximized;
• and
• computing a loss to reduce similarities between the plurality of second data samples to make the plurality of second data samples orthogonal to each other;
• calculating a relation between the plurality of second data samples and the training data to obtain a probabilistic bag of images representation for each of the plurality of second data samples;
• and
• generating a label for each of the plurality of second data samples based on the probabilistic bag of images representation;
• receiving, at the trained neural network, an input data sample;
• computing, using the trained neural network, similarities between the input data sample and respective ones of the plurality of second data samples based on the labels of the plurality of second data samples;
• selecting, using the trained neural network, one or more second data samples of the plurality of second data samples based, at least in part, on the computed similarities;
• and
• retrieving, using the trained neural network, one or more first data samples of the plurality of first data samples from the input dataset based, at least in part, on the selected one or more second data samples.