Abstract: Systems, apparatuses, and methods for adaptively mapping a machine learning model to a multi-core inference accelerator engine are disclosed. A computing system includes a multi-core inference accelerator engine with multiple inference cores coupled to a memory subsystem. The system also includes a control unit which determines how to adaptively map a machine learning model to the multi-core inference accelerator engine. In one implementation, the control unit selects a mapping scheme which minimizes the memory bandwidth utilization of the multi-core inference accelerator engine. In one implementation, this mapping scheme involves having one inference core of the multi-core inference accelerator engine fetch given data and broadcast the given data to other inference cores of the inference accelerator engine. Each inference core fetches second data unique to the respective inference core.

Abstract: Disclosed is a processor-implemented data processing method in a neural network. A data processing apparatus includes at least one processor, and at least one memory configured to store instructions to be executed by the processor and a neural network, wherein the processor is configured to, based on the instructions, input an input activation map into a current layer included in the neural network, output an output activation map by performing a convolution operation between the input activation map and a weight quantized with a first representation bit number of the current layer, and output a quantized activation map by quantizing the output activation map with a second representation bit number based on an activation quantization parameter.

Abstract: Methods and systems include a neural network system that includes a neural network accelerator. The neural network accelerator includes multiple processing engines coupled together to perform arithmetic operations in support of an inference performed using the deep neural network system. The neural network accelerator also includes a schedule-aware tensor data distribution circuitry or software that is configured to load tensor data into the multiple processing engines in a load phase, extract output data from the multiple processing engines in an extraction phase, reorganize the extracted output data, and store the reorganized extracted output data to memory.

Abstract: In a data processing system, at least one processing node is configured to perform computations for a multi-stage process whilst at least one other processor performs the load/unload operations required to calculate a subsequent stage of the multi stage process. An exchange of data then occurs between the processing nodes. At a later time, at least one processing node performs calculations using the data loaded from storage, whilst at least one other processor performs the load/unload operations required to calculate a subsequent stage of the multi stage process.

Abstract: An acceleration engine with multiple accelerators may share a common set of data that is used by each accelerator to perform computations on input data. The set of shared data can be loaded into the acceleration engine from an external memory. Instead of accessing the external memory multiple times to load the set of shared data into each accelerator, the external memory can be accessed once using direct memory access to load the set of shared data into the first accelerator. The set of shared data can then be serially loaded from one accelerator to the next accelerator in the acceleration engine using direct memory access. To achieve data parallelism and reduce computation time, a runtime driver may split the input data into data batches, and each accelerator can perform computations on a different batch of input data with the common set of shared data.

Abstract: A neuromorphic apparatus configured to process a multi-bit neuromorphic operation including a single axon circuit, a single synaptic circuit, a single neuron circuit, and a controller. The single axon circuit is configured to receive, as a first input, an i-th bit of an n-bit axon. The single synaptic circuit is configured to store, as a second input, a j-th bit of an m-bit synaptic weight and output a synaptic operation value between the first input and the second input. The single neuron circuit is configured to obtain each bit value of a multi-bit neuromorphic operation result between the n-bit axon and the m-bit synaptic weight, based on the output synaptic operation value. The controller is configured to respectively determine the i-th bit and the j-th bit to be sequentially assigned for each time period of different time periods to the single axon circuit and the single synaptic circuit.

Abstract: A data format converter rearranges data of an input image for input to a systolic array of multiply and accumulate processing elements. The image has a pixel height and a pixel width in a number of channels equal to a number of colors per pixel. The data format converter rearranges the data to a second, greater number of channels and inputs the second number of channels to one side of the systolic array. The second number of channels is less than or equal to the number of MAC PEs on the one side of the systolic array, and results in greater MAC PE utilization in the systolic array.

Abstract: Provided is an operation method for a memory device, the memory device being used for implementing an Artificial Neural Network (ANN). The operation method includes: reading from the memory device a weight matrix of a current layer of a plurality of layers of the ANN to extract a plurality of neuro values; determining whether to perform calibration; when it is determined to perform calibration, recalculating and updating a mean value and a variance value of the neuro values; and performing batch normalization based on the mean value and the variance value of the neuro values.

Abstract: A feature sub-network trainer improves robustness of interpretability of a deep neural network (DNN) by increasing the likelihood that the DNN will converge to a global minimum of a cost function of the DNN. After determining a plurality of correctly classified examples of a pre-trained DNN, the trainer extracts from the pre-trained DNN a feature sub-network that includes an input layer of the DNN and one or more subsequent sparsely-connected layers of the DNN. The trainer averages output signals from the sub-network to form an average representation of each class identifiable by the DNN. The trainer relabels each correctly classified example with the appropriate average representation, and then trains the feature sub-network with the relabeled examples. In one demonstration, the feature sub-network trainer improved classification accuracy of a seven-layer convolutional neural network, trained with two thousand examples, from 75% to 83% by reusing the training examples.

Abstract: A computational method simulating the motion of elements within a multi-element system using a graph neural network (GNN). The method includes converting a molecular dynamics snapshot of the elements into a directed graph comprised of nodes and edges. The method further includes the step of initially embedding the nodes and the edges to obtain initially embedded nodes and edges. The method also includes updating the initially embedded nodes and edges by passing a first message from a first edge to a first node using a first message function and passing a second message from the first node to the first edge using a second message function to obtain updated embedded nodes and edges, and predicting a force vector for one or more elements based on the updated embedded edges and a unit vector pointing from the first node to a second node or the second node to the first node.

Abstract: Various embodiments are generally directed to an apparatus, system, and other techniques for dynamic and intelligent deployment of a neural network or any inference model on a hardware executor or a combination of hardware executors. Computational costs for one or more operations involved in executing the neural network or inference model may be determined. Based on the computational costs, an optimal distribution of the computational workload involved in running the one or more operations among multiple hardware executors may be determined.

Abstract: A method includes generating a training data set comprising a plurality of training examples, wherein each training example is generated by receiving map data associated with a road portion, receiving sensor data associated with a road agent located on the road portion, defining one or more corridors associated with the road portion based on the map data and the sensor data, extracting a plurality of agent features associated with the road agent based on the sensor data, extracting a plurality of corridor features associated with each of the one or more corridors based on the sensor data, and for each corridor, labeling the training example based on the position of the road agent with respect to the corridor, and training a neural network using the training data set.

Abstract: A feature map caching method of a convolutional neural network includes a connection analyzing step and a plurality of layer operation steps. The connection analyzing step is for analyzing a network to establish a convolutional neural network connection list. The convolutional neural network connection list includes a plurality of tensors and a plurality of layer operation coefficients. Each of the layer operation coefficients includes a step index, at least one input operand label and an output operand label. The step index as a processing order for the layer operation step. At least one of the layer operation steps is for flushing at least one of the tensors in a cache according to a distance between the at least one of the layer operation steps and a future layer operation step of the layer operation steps. The distance is calculated according to the convolutional neural network connection list.

Abstract: A method for receiving an ownership graph, wherein the ownership graph comprises a first set of nodes and a first set of directional edges, and wherein each of the first set of directional edges connects two nodes and indicates ownership of a first node by a second node, each node having at most one owner, the ownership graph being acyclic. The method further includes receiving a dependency graph that also comprises a set of nodes and a set of directional edges. The method further includes creating a respective enumerating variable declaration for each node in a path from an owner node to a root node in the ownership graph. The method further includes creating a respective accessing variable declaration for each owner node in the dependency graph of the current node.

Abstract: A method of generating training data for training a neural network, method of training a neural network and using a neural network for autonomous operations, related devices and systems. In one aspect, a neural network for autonomous operation of an object in an environment is trained. Policy values are generated based on a sample data set. An approximate action-value function is generated from the policy values. A set of approximated policy values is generated using the approximate action-value function for all states in the sample data set for all possible actions. A training target for the neural network is calculated based on the approximated policy values. A training error is calculated as the difference between the training target and the policy value for the corresponding state-action pair in the sample data set. At least some of the parameters of the neural network are updated to minimize the training error.

Abstract: A reservoir element of the first aspect of the present disclosure includes: a spin conduction layer containing a non-magnetic conductor; ferromagnetic layers positioned in a first direction with respect to the spin conduction layer and spaced apart from each other in a plan view from the first direction; and via wirings electrically connected to spin conduction layer on a surface opposite to a surface with the ferromagnetic layers.

Abstract: In an example, a neural network program corresponding to a neural network model is received. The neural network program includes matrices, vectors, and matrix-vector multiplication (MVM) operations. A computation graph corresponding to the neural network model is generated. The computation graph includes a plurality of nodes, each node representing a MVM operation, a matrix, or a vector. Further, a class model corresponding to the neural network model is populated with a data structure pointing to the computation graph. The computation graph is traversed based on the class model. Based on the traversal, the plurality of MVM operations are assigned to MVM units of a neural network accelerator. Each MVM unit can perform a MVM operation. Based on assignment of the plurality of MVM operations, an executable file is generated for execution by the neural network accelerator.

Abstract: Aspects of the present disclosure address systems and methods for fixed-point quantization using a dynamic quantization level adjustment scheme. Consistent with some embodiments, a method comprises accessing a neural network comprising floating-point representations of filter weights corresponding to one or more convolution layers. The method further includes determining a peak value of interest from the filter weights and determining a quantization level for the filter weights based on a number of bits in a quantization scheme. The method further includes dynamically adjusting the quantization level based on one or more constraints. The method further includes determining a quantization scale of the filter weights based on the peak value of interest and the adjusted quantization level. The method further includes quantizing the floating-point representations of the filter weights using the quantization scale to generate fixed-point representations of the filter weights.

Abstract: A system includes a data collection engine, a plurality of items including radio-frequency identification chips, a plurality of third party data and insight sources, a plurality of interfaces, client devices, a server and method thereof for preventing suicide. The server includes trained machine learning models, business logic and attributes of a plurality of patient events. The data collection engine sends attributes of new patient events to the server. The server can predict an adverse event risk of the new patient events based upon the attributes of the new patient events utilizing the trained machine learning models.

Abstract: Methods, systems, and circuits for training a neural network include applying noise to a set of training data across wordlines using a respective noise switch on each wordline. A neural network is trained using the noise-applied training data to generate a classifier that is robust against adversarial training.