Machine-learned verification pipeline for sensor inputs

1 . A computer-implemented method for verification of sensor inputs with a machine-learned verification pipeline, the method comprising:

obtaining, by one or more sensors of a computing device during a first session on the computing device, unverified interaction data descriptive of interactions of a user with the computing device;

generating, using a machine-learned embedding network of the machine-learned verification pipeline, feature values that embed characteristics of the unverified interaction data;

determining, using a verification model of the machine-learned verification pipeline, and based on the feature values, a user account associated with the unverified interaction data;

associating, by the computing device, the first session on the computing device with the user account;

obtaining, by the computing device during a second session on the computing device, second unverified interaction data from interactions of a second user;

generating, by the machine-learned embedding network, second feature values that embed characteristics of the second unverified interaction data; and

associating, using the verification model and based on the second feature values, in the background during the second session, the second session on the computing device with a second user account associated with the second user.

2 . The method of claim 1 , comprising:

initiating a prompt for rendering by the computing device to request confirmation, from the user, of the determined user account.

3 . The method of claim 2 , comprising:

updating one or more parameters of the machine-learned verification pipeline based on the confirmation.

4 . The method of claim 1 , wherein the determining the user account associated with the unverified interaction data comprises:

comparing the feature values against a reference library comprising one or more verified embeddings associated with the user account.

5 . The method of claim 1 , wherein the determining the user account associated with the unverified interaction data comprises:

comparing the feature values against a reference set of feature values, the reference set comprising one or more clustered embeddings; and

determining an association with the user account based on a distance between the feature values and the one or more clustered embeddings.

6 . The method of claim 1 , comprising:

rendering, by the computing device, content selected based on the determined user account.

7 . The method of claim 6 , wherein the content is selected by real-time content arbitration among a plurality of content providers.

8 . The method of claim 1 ,

wherein the computing device is associated with multiple user accounts;

wherein the computing device is configured to operate multiple user sessions respectively with the multiple user accounts; and

wherein the computing device is configured to operate the multiple user sessions without requiring a sign-in for each session.

9 . The method of claim 1 , wherein one or more parameters of the machine-learned verification pipeline was trained based on prior inputs to a different computing device associated with the user account.

10 . The method of claim 9 , wherein the prior inputs were verified based on a sign-on procedure associated with unlocking an interface of the different computing device.

11 . The method of claim 1 , wherein the machine-learned verification pipeline performs image-based processing of the unverified interaction data.

12 . The method of claim 1 , wherein the machine-learned verification pipeline performs sequence-based processing of the unverified interaction data.

13 . The method of claim 1 , wherein the machine-learned verification pipeline processes a map of input locations.

14 . The method of claim 1 , wherein the machine-learned verification pipeline comprises a transformer model.

15 . The method of claim 1 , wherein the machine-learned verification pipeline comprises a convolutional neural network.

16 . The method of claim 1 , wherein the machine-learned verification pipeline comprises an encoder and a decoder, and wherein the encoder comprises the machine-learned embedding network and the decoder comprises the verification model.

17 . The method of claim 1 , wherein the machine-learned verification pipeline provides, via an API, account verification as a service to one or more applications executing on the computing device.

18 . A computer-implemented method for verification of sensor inputs with a machine-learned verification pipeline, the method comprising:

obtaining, by a plurality of sensors of a computing device, unverified interaction data descriptive of interactions of a user with the computing device, wherein the plurality of sensors record interaction data at different data resolutions;

merging the unverified interaction data into a multidimensional input format that indexes data from the plurality of sensors over a plurality of time steps at a shared data resolution;

inputting the merged unverified interaction data to a machine-learned embedding network of the machine-learned verification pipeline;

generating, using the machine-learned embedding network of the machine-learned verification pipeline, feature values that embed characteristics of the unverified interaction data; and

determining, using a verification model of the machine-learned verification pipeline, and based on the feature values, a user account associated with the unverified interaction data.

19 . A computing device, comprising:

a plurality of sensors;

one or more processors; and

a memory device comprising one or more non-transitory computer-readable media storing:

a machine-learned verification pipeline comprising:

a machine-learned embedding network configured to receive a multidimensional input format that indexes data from the plurality of sensors over a plurality of time steps at a shared data resolution; and

a verification model;

instructions that are executable to cause one or more processors to perform operations comprising:

obtaining, from the plurality of sensors, unverified interaction data descriptive of interactions of a user with the computing device, wherein the plurality of sensors record interaction data at different data resolutions;

merging the unverified interaction data into the multidimensional input format;

inputting the merged unverified interaction data to the machine-learned embedding network;

generating, using the machine-learned embedding network, feature values that embed characteristics of the unverified interaction data; and

determining, using the verification model and based on the feature values, a user account associated with the unverified interaction data.

20 . The method of claim 1 , wherein the computing device associates the second session on the computing device with the second user account associated with the second user without requiring explicit confirmation from the second user.

21 . The method of claim 1 , wherein the feature values correspond to a plurality of image frames in a buffer representing snapshots of user interactions over time.