Regularized multi-label classification from partially labeled training data

A computer-implemented method, comprising:

• accessing data representing an image;
• accessing data representing an indication that a first class of a plurality of classes is depicted in the image, wherein the indication does not specify whether the image depicts a second class of the plurality of classes;
• create a data record that labels the first class with a first label value representing known positive ground truth and labels the second class with a second label value representing unknown ground truth;
• training a machine learning model to predict the first value from the data representing the image at least partly by:
• distinguishing between the first label value and the second label value;
• performing a forward pass of data representing the image through the machine learning model to generate a first probability that the first class is depicted in the image and a second probability that the second class is depicted in the image;
• and
• computing updates to parameters of the machine learning model differently for predictions relating to the first class and the second class based at least partly on the distinguishing, wherein a first portion of the updates are computed based at least partly on the first probability, and wherein a second portion of the updates are computed based on applying a regularization penalty to an update of the parameters based on a difference between the second probability and the second label value representing the unknown ground truth;
• and
• storing a representation of the machine learning model, the representation including the updates to the parameters.

A computer-implemented method, comprising:

• accessing data representing an image;
• accessing data representing an indication that a first class of a plurality of classes is depicted in the image, wherein the indication does not specify whether the image depicts a second class of the plurality of classes;
• creating a data record that labels the first class with a first label value representing known positive ground truth and labels the second class with a second label value representing unknown ground truth;
• training a machine learning model to predict the first value from the data representing the image at least partly by:
• distinguishing between the first label value and the second label value;
• performing a forward pass of data representing the image through the machine learning model to generate a first probability that the first class is depicted in the image and a second probability that the second class is depicted in the image;
• and
• computing updates to parameters of the machine learning model differently for predictions relating to the first class and the second class based at least partly on the distinguishing, wherein computing the updates comprises using a loss function to compute the updates to the parameters based on a difference between the first probability and the first label value, and not updating the parameters based on a difference between the second probability and the second label value representing the unknown ground truth;
• and
• storing a representation of the machine learning model, the representation including the updates to the parameters.

A computer-implemented method, comprising:

• accessing data representing an image;
• accessing data representing an indication that a first class of a plurality of classes is depicted in the image, wherein the indication does not specify whether the image depicts a second class of the plurality of classes;
• creating a data record that labels the first class with a first label value representing known positive ground truth and labels the second class with a second label value representing unknown ground truth;
• training a machine learning model to predict the first value from the data representing the image at least partly by:
• distinguishing between the first label value and the second label value, and
• computing updates to parameters of the machine learning model differently for predictions relating to the first class and the second class based at least partly on the distinguishing;
• and
• storing a representation of the machine learning model, the representation including the updates to the parameters;
• and
• performing a forward pass of data representing the image through the machine learning model to generate a first probability that the first class is depicted in the image and a second probability that the second class is depicted in the image;
• wherein the machine learning model comprises a neural network, the computer-implemented method further comprising:
• obtaining an n-dimensional output from a final layer of the neural network, wherein a number n of the n-dimensional output corresponds to the number n of the plurality of classes;
• applying a sigmoid function to a first dimension of the n-dimensional output to generate the first probability that the first class is depicted in the image;
• and
• applying the sigmoid function to a second dimension of the n-dimensional output to generate the second probability that the second class is depicted in the image.

A computer-implemented method, comprising:

• accessing data representing an image;
• accessing data representing an indication that a first class of a plurality of classes is depicted in the image, wherein the indication does not specify whether the image depicts a second class of the plurality of classes;
• creating a data record that labels the first class with a first label value representing known positive ground truth and labels the second class with a second label value representing unknown ground truth;
• training a machine learning model to predict the first value from the data representing the image at least partly by:
• distinguishing between the first label value and the second label value, and
• computing updates to parameters of the machine learning model differently for predictions relating to the first class and the second class based at least partly on the distinguishing;
• storing a representation of the machine learning model, the representation including the updates to the parameters;
• accessing a loss function for computing the updates to the parameters of the machine learning model, wherein the loss function includes an indicator function;
• using the indicator function to identify that the first class has the first label value and that the second class has the second label value;
• and
• distinguishing between the first value and the second value based on the indicator function.

A computer-implemented method, comprising:

• accessing data representing an image;
• accessing data representing an indication that a first class of a plurality of classes is depicted in the image, wherein the indication does not specify whether the image depicts a second class of the plurality of classes;
• creating a data record that labels the first class with a first label value representing known positive ground truth and labels the second class with a second label value representing unknown ground truth;
• training a machine learning model to predict the first value from the data representing the image at least partly by:
• distinguishing between the first label value and the second label value, and
• computing updates to parameters of the machine learning model differently for predictions relating to the first class and the second class based at least partly on the distinguishing;
• storing a representation of the machine learning model, the representation including the updates to the parameters;
• accessing data representing an additional image and a corresponding additional indication that the additional image depicts none of the plurality of classes;
• labeling the plurality of classes with a third label value representing known negative ground truth;
• and
• computing the updates to the parameters based additionally on differences between the third label value and predictions of the machine learning model regarding whether the plurality of classes are depicted in the additional image.

A non-transitory computer-readable medium storing instructions that, when executed, cause at least one processor to perform operations comprising:

• accessing a machine learning model trained to predict whether each of a plurality of classes is depicted in an image, wherein parameters of the machine learning model represent training by a learning framework that updates the parameters differently for known and unknown ground truth labels relating to the plurality of classes;
• forward passing pixel values of an image through the machine learning model;
• based on an output of the machine learning model, determining that at least one class of the plurality of classes is depicted in an image;
• and
• storing the image in association with data indicating that the image depicts the at least one class;
• and
• updating weights of the machine learning model using a loss function that includes an indicator function, wherein the indicator function identifies that a first class of the plurality of classes has a first label value and that a second class of the plurality of classes has a second label value;
• wherein the loss function determines updates to the weights of the machine learning model differently for the first class and the second class based on the indicator function.