Systems, methods, and computer-readable for multi-spatial scale object detection include generating one or more object trackers for tracking at least one object detected from on one or more images. One or more blobs are generated for the at least one object based on tracking motion associated with the at least one object. One or more tracklets are generated for the at least one object based on associating the one or more object trackers and the one or more blobs, the one or more tracklets including one or more scales of object tracking data for the at least one object. One or more uncertainty metrics are generated using the one or more object trackers and an embedding of the one or more tracklets. A training module for detecting and tracking the at least one object using the embedding and the one or more uncertainty metrics is generated using deep learning techniques.

Systems and methods for distributed training of deep learning models are disclosed. An example local device to train deep learning models includes a reference generator to label input data received at the local device to generate training data, a trainer to train a local deep learning model and to transmit the local deep learning model to a server that is to receive a plurality of local deep learning models from a plurality of local devices, the server to determine a set of weights for a global deep learning model, and an updater to update the local deep learning model based on the set of weights received from the server.

A device, system, and method is provided for storing a sparse neural network. A plurality of weights of the sparse neural network may be obtained. Each weight may represent a unique connection between a pair of a plurality of artificial neurons in different layers of a plurality of neuron layers. A minority of pairs of neurons in adjacent neuron layers are connected in the sparse neural network. Each of the plurality of weights of the sparse neural network may be stored with an association to a unique index. The unique index may uniquely identify a pair of artificial neurons that have a connection represented by the weight. Only non-zero weights may be stored that represent connections between pairs of neurons (and zero weights may not be stored that represent no connections between pairs of neurons).

The present approach relates to a system capable of life-long learning in a deep learning context. The system includes a deep learning network configured to process an input dataset and perform one or more tasks from among a first set of tasks. As an example, the deep learning network may be part of an imaging system, such as a medical imaging system, or may be used in industrial applications. The system further includes a learning unit communicatively coupled to the deep learning network 102 and configured to modify the deep learning network so as to enable it to perform one or more tasks in a second task list without losing the ability to perform the tasks from the first list.

Provided is a system, method, and computer program product for perforated backpropagation. The method includes segmenting a plurality of nodes into at least two sets including a set of first nodes and a set of second nodes, determining an error term for each node of the set of first nodes, the first set of nodes comprising a first and second subset of nodes, backpropagating the error terms for each node throughout the set of first nodes, determining an error term for each node of the first subset of nodes of the set of first nodes based on direct connections between the first subset of nodes and the second subset of nodes independent of error terms of the set of second nodes, determining an error term for each node of the set of second nodes, and updating weights of each node of the plurality of nodes based on the error term.

Provided is a system, method, and computer program product for perforated backpropagation. The method includes segmenting a plurality of nodes into at least two sets including a set of first nodes and a set of second nodes, determining an error term for each node of the set of first nodes, the first set of nodes comprising a first and second subset of nodes, backpropagating the error terms for each node throughout the set of first nodes, determining an error term for each node of the first subset of nodes of the set of first nodes based on direct connections between the first subset of nodes and the second subset of nodes independent of error terms of the set of second nodes, determining an error term for each node of the set of second nodes, and updating weights of each node of the plurality of nodes based on the error term.

A hybrid entity recognition system and accompanying method identify composite entities based on machine learning. An input sentence is received and is preprocessed to remove extraneous information, perform spelling correction, and perform grammar correction to generate a cleaned input sentence. A POS tagger tags parts of speech of the cleaned input sentence. A rules based entity recognizer module identifies first level entities in the cleaned input sentence. The cleaned input sentence is converted and translated into numeric vectors. Basic and composite entities are extracted from the cleaned input sentence using the numeric vectors.

According to one aspect, a long short-term memory (LSTM) cell for sensor fusion may include M number of forget gates, M number of input gates, and M number output gates. The M number of forget gates may receive M sets of sensor encoding data from M number of sensors and a shared hidden state. The M number of input gates may receive the corresponding M sets of sensor data and the shared hidden state. The M number output gates may generate M partial shared cell state outputs and M partial shared hidden state outputs based on the M sets of sensor encoding data, the shared hidden state, and a shared cell state.

According to one aspect, a long short-term memory (LSTM) cell for sensor fusion may include M number of forget gates, M number of input gates, and M number output gates. The M number of forget gates may receive M sets of sensor encoding data from M number of sensors and a shared hidden state. The M number of input gates may receive the corresponding M sets of sensor data and the shared hidden state. The M number output gates may generate M partial shared cell state outputs and M partial shared hidden state outputs based on the M sets of sensor encoding data, the shared hidden state, and a shared cell state.

A method for use with a computing device is provided. The method may include inputting an input data set into a first private artificial intelligence model generated using a first private data set and a second private artificial intelligence model generated using a second private data set. The method may further include receiving a first result data set from the first private artificial intelligence model and receiving a second result data set from the second private artificial intelligence model. The method may further include training an adaptive co-distillation model with the input data set and the first result data set. The method may further include training the adaptive co-distillation model with the input data set and the second result data set. The adaptive co-distillation model may not be trained on the first private data set or the second private data set.