Methods, systems and apparatus, including computer programs encoded on computer storage medium, for training a neurosymbolic data imputation system on training data inputs in a domain to impute missing data in a data input from the data domain. In one aspect a method includes, for each training data input, adding random noise to missing fields of the training data input;

generating an embedding data input for the training data input using concept embeddings from the domain; processing the noisy data input and the embedding data input through a correlation network to obtain correlation data; applying attention to the noisy training data input and the correlation data to generate a combined data input; processing, by an autoencoder, the combined data input to obtain a decoded data output; computing a difference between the decoded data output and the training data input; and updating parameters of the data imputation system using the difference.

Systems and computer-implemented methods of performing image registration. One method includes receiving a first image and a second image acquired from a patient at different times and, in each of the first image and the second image, detecting an upper boundary of an imaged object in an image coordinate system and detecting a lower boundary of the imaged object in the image coordinate system. The method further includes, based on the upper boundary and the lower boundary of each of the first image and the second image, cropping and padding at least one of the first image and the second image to create an aligned first image and an aligned second image and executing a registration model on the aligned first image and the aligned second image to compute a deformation field between the aligned first image and the aligned second image.

A method for determining hard example sensor data inputs for training a task neural network is described. The task neural network is configured to receive a sensor data input and to generate a respective output for the sensor data input to perform a machine learning task. The method includes: receiving one or more sensor data inputs depicting a same scene of an environment, wherein the one or more sensor data inputs are taken during a predetermined time period; generating a plurality of predictions about a characteristic of an object of the scene; determining a level of inconsistency between the plurality of predictions; determining that the level of inconsistency exceeds a threshold level; and in response to the determining that the level of inconsistency exceeds a threshold level, determining that the one or more sensor data inputs comprise a hard example sensor data input.

Neural network performance is improved in terms of training speed and/or accuracy by encoding (mapping) inputs to the neural network into a higher dimensional space via a hash function. The input comprises coordinates used to identify a point within a d-dimensional space (e.g., 3D space). The point is quantized and a set of vertex coordinates corresponding to the point are input to a hash function. For example, for d=3, space may be partitioned into axis-aligned voxels of identical size and vertex coordinates of a voxel containing the point are input to the hash function to produce a set of encoded coordinates. The set of encoded coordinates is used to lookup D-dimensional feature vectors in a table of size T that have been learned. The learned feature vectors are filtered (e.g., linearly interpolated, etc.) based on the coordinates of the point to compute a feature vector corresponding to the point.

Neural network performance is improved in terms of training speed and/or accuracy by encoding (mapping) inputs to the neural network into a higher dimensional space via a hash function. The input comprises coordinates used to identify a point within a d-dimensional space (e.g., 3D space). The point is quantized and a set of vertex coordinates corresponding to the point are input to a hash function. For example, for d=3, space may be partitioned into axis-aligned voxels of identical size and vertex coordinates of a voxel containing the point are input to the hash function to produce a set of encoded coordinates. The set of encoded coordinates is used to lookup D-dimensional feature vectors in a table of size T that have been learned. The learned feature vectors are filtered (e.g., linearly interpolated, etc.) based on the coordinates of the point to compute a feature vector corresponding to the point.

Techniques for implementing a performance-adaptive sampling strategy towards fast and accurate graph neural networks are provided. In one technique, a graph that comprises multiple nodes and edges connecting the nodes is stored. An embedding for each node is initialized, as well as a sampling policy for sampling neighbors of nodes. One or more machine learning techniques are used to train a graph neural network and learn embeddings for the nodes. Using the one or more machine learning techniques comprises, for each node: (1) selecting, based on the sampling policy, a set of neighbors of the node; (2) based on the graph neural network and embeddings for the node and the set of neighbors, computing a performance loss; and (3) based on a gradient of the performance loss, modifying the sampling policy.

Techniques for implementing a performance-adaptive sampling strategy towards fast and accurate graph neural networks are provided. In one technique, a graph that comprises multiple nodes and edges connecting the nodes is stored. An embedding for each node is initialized, as well as a sampling policy for sampling neighbors of nodes. One or more machine learning techniques are used to train a graph neural network and learn embeddings for the nodes. Using the one or more machine learning techniques comprises, for each node: (1) selecting, based on the sampling policy, a set of neighbors of the node; (2) based on the graph neural network and embeddings for the node and the set of neighbors, computing a performance loss; and (3) based on a gradient of the performance loss, modifying the sampling policy.

A computer system trains a federated learning model. A federated learning model is distributed to a plurality of computing nodes, each having a set of local training data comprising labeled data samples. Statistical data is received from each computing node that indicates the node's count of data samples for each label, and is analyzed to identify one or more computing nodes having local training data in which a label category is underrepresented beyond a threshold value with respect to data samples. Additional data samples labeled with the underrepresented labels are provided, and the computing nodes perform training. Results of training are received and are processed to generate a trained global model. Embodiments of the present invention further include a method and program product for training a federated learning model in substantially the same manner described above.

A computer system trains a federated learning model. A federated learning model is distributed to a plurality of computing nodes, each having a set of local training data comprising labeled data samples. Statistical data is received from each computing node that indicates the node's count of data samples for each label, and is analyzed to identify one or more computing nodes having local training data in which a label category is underrepresented beyond a threshold value with respect to data samples. Additional data samples labeled with the underrepresented labels are provided, and the computing nodes perform training. Results of training are received and are processed to generate a trained global model. Embodiments of the present invention further include a method and program product for training a federated learning model in substantially the same manner described above.

Various methods and systems are provided for computed tomography imaging. In one embodiment, a method includes acquiring, with an x-ray detector and an x-ray source coupled to a gantry, a three-dimensional image volume of a subject while the subject moves through a bore of the gantry and the gantry rotates the x-ray detector and the x-ray source around the subject, inputting the three-dimensional image volume to a trained deep neural network to generate a corrected three-dimensional image volume with a reduction in aliasing artifacts present in the three-dimensional image volume, and outputting the corrected three-dimensional image volume. In this way, aliasing artifacts caused by sub-sampling may be removed from computed tomography images while preserving details, texture, and sharpness in the computed tomography images.