A first processor processes an instruction configured to perform a plurality of functions. The plurality of functions includes one or more functions to operate on one or more tensors. A determination is made of a function of the plurality of functions to be performed. The first processor provides to a second processor information related to the function. The second processor is to perform the function. The first processor and the second processor share memory providing memory coherence.

A tensor of a first select dimension is reformatted to provide one or more sub-tensors of a second select dimension. The reformatting includes determining a number of sub-tensors to be used to represent the tensor. The reformatting further includes creating the number of sub-tensors, in which a sub-tensor is to start on a boundary of a memory unit. Data of the tensor is rearranged to fit within the number of sub-tensors.

This disclosure relates generally to system and method for determining explainability of machine predicted decisions. Typical explainable AI (XAI) solutions are limited by type of data processed, such as structured, semi-structured and unstructured text. In addition, due to limited automation of the process of explainability, typical systems are cumbersome and time-consuming. The system and method provide an end to end solution for automating the determination of explainability of machine predicted decisions. The XAI process output an absolute relevance score indicative of relevance of the features associated with the prediction which is indicative of percentage relevance/contribution of individual feature. The system further computes relative relevance score of the features by adding up all the features and calculating how much each individual feature is contributing to the total score. The relative relevance scores are utilized for determining explainability of decisions of the prediction.

A method and system of graph feature extraction and graph classification based on adjacency matrix is provided. The invention first concentrates the connection information elements in the adjacency matrix into a specific diagonal region of the adjacency matrix which reduces the non-connection information elements in advance. Then the subgraph structure of the graph is further extracted along the diagonal direction using the filter matrix. Further, it uses a stacked convolutional neural network to extract a larger subgraph structure. On one hand, it greatly reduces the amount of computation and complexity, getting rid of the limitations caused by computational complexity and window size. On the other hand, it can capture large subgraph structure through a small window, as well as deep features from the implicit correlation structures at both vertex and edge level, which improves speed and accuracy of graph classification.

Techniques are described for a compiler scheduling algorithm/routine that utilizes backtracking to generate an execution schedule for a neural network computation graph using a neural network compiler intermediate representation of hardware synchronization counters. The hardware synchronization counters may be referred to as physical barriers, hardware (HW) barriers, or barriers and their intermediate representations may be referred to as barrier tasks or barriers. Backtracking is utilized to prevent an available number of hardware barriers from being exceeded during performance of an execution schedule. An execution schedule may be a computation workload schedule for neural network inference applications. An execution schedule may also be a first in first out (FIFO) schedule.

Systems, methods, and computer-executable instructions for determining a computation schedule for a recurrent neural network (RNN). A matrix multiplication (MM) directed-acyclic graph (DAG) is received for the RNN. Valid phased computation schedules for the RNN are generated. Each of the valid phase computation schedule includes an ordering of MM operations. For each of the plurality of valid phased computation schedules, each of the MM operations is partitioned to processor cores based on L3 cache to L2 cache data movement. The RNN is executed based on the valid phased computation schedules. A final computation schedule is stored. The final computation schedule is used for future executions of the RNN.

Various embodiments of systems and methods for attention models with random features for multi-layered graph embeddings are disclosed.

Various embodiments set forth systems and techniques for augmenting neural networks. The techniques include causing one or more neural networks to generate first output based on a first input; identifying one or more rules associated with the first input; processing the first output based on the one or more rules to generate a second output; and transmitting the second output, instead of the first output, as a result of processing the first input.

There is provided an information processing device which efficiently executes machine learning. The information processing device according to one embodiment includes: an obtaining unit which obtains a source code including a code which defines Forward processing of each layer constituting a neural network; a storage unit which stores an association relationship between each Forward processing and Backward processing associated with each Forward processing; and an executing unit which successively executes each code included in the source code, and which calculates an output value of the Forward processing defined by the code based on an input value at a time of execution of each code, and generates a reference structure for Backward processing in a layer associated with the code based on the association relationship stored in the storage unit.

A non-transitory computer-readable recording medium stores a learning program that causes a computer to execute a process including: extracting, from a plurality of data groups, a plurality of first data groups having an order; generating, for each data element corresponding to each of the first data groups, an ordered data matrix in which data elements having same order have value corresponding to relationship among the data elements and data elements having different orders have values corresponding to the different orders; and obtaining input tensor data by performing tensor decomposition with the ordered data matrix, inputting the input tensor data to a neural network at time of performing deep machine learning, performing deep machine learning of the neural network, and learning about method for the tensor decomposition.