Hailo Technologies Ltd.

Hallo Technologies Ltd.

Or Danon

Avi Baum

Daniel Chibotero

Daniel Ciubotariu

Hadar Zeitlin

Roi Seznayov

Guy Kaminitz

Rami Feig

Ori Katz

Nir Engelberg

Yuval Adelstein

Itai Resh

Mark Grobman

Gilad Nahor

Amir Shmul

Alex Finkelstein

A novel and useful system and method of improved power performance and lowered memory requirements for an artificial neural network based on packing memory utilizing several structured sparsity mechanisms. The invention applies to neural network (NN) processing engines adapted to implement mechanisms to search for structured sparsity in weights and activations, resulting in a considerably reduced memory usage. The sparsity guided training mechanism synthesizes and generates structured sparsity weights. A compiler mechanism within a software development kit (SDK), manipulates structured weight domain sparsity to generate a sparse set of static weights for the NN. The structured sparsity static weights are loaded into the NN after compilation and utilized by both the structured weight domain sparsity mechanism and the structured activation domain sparsity mechanism. The application of structured sparsity lowers the span of search options and creates a relatively loose coupling between the data and control planes.

Structured sparsity guided training in an artificial neural network

Novel and useful system and methods of functional safety mechanisms for use in an artificial neural network (ANN) processor. The mechanisms can be deployed individually or in combination to provide a desired level of safety in neural networks. Multiple strategies are applied involving redundancy by design, redundancy through spatial mapping as well as self-tuning procedures that modify static (weights) and monitor dynamic (activations) behavior. The NN processor incorporates functional safety concepts which reduce its risk of failure that occurs during operation from going unnoticed. The mechanisms function to detect and promptly flag and report the occurrence of an error with some mechanisms capable of correction as well. The safety mechanisms cover data stream fault detection, software defined redundant allocation, cluster interlayer safety, cluster intralayer safety, layer control unit (LCU) instruction addressing, weights storage safety, and neural network intermediate results storage safety.

Cluster interlayer safety mechanism in an artificial neural network processor

Novel and useful system and methods of several functional safety mechanisms for use in an artificial neural network (ANN) processor. The mechanisms can be deployed individually or in combination to provide a desired level of safety in neural networks. Multiple strategies are applied involving redundancy by design, redundancy through spatial mapping as well as self-tuning procedures that modify static (weights) and monitor dynamic (activations) behavior. The mechanisms address ANN system level safety in situ, as a system level strategy tightly coupled with the processor architecture. The NN processor incorporates several functional safety concepts that function to detect and promptly flag and report an error with some mechanisms capable of correction as well. The safety mechanisms cover data stream fault detection, software defined redundant allocation, cluster interlayer safety, cluster intralayer safety, layer control unit (LCU) instruction addressing, weights storage safety, and neural network intermediate results storage safety.

Weights safety mechanism in an artificial neural network processor

A novel and useful neural network (NN) processing core incorporating inter-device connectivity and adapted to implement artificial neural networks (ANNs). A chip-to-chip interface spreads a given ANN model across multiple devices in a seamless manner. The NN processor is constructed from self-contained computational units organized in a hierarchical architecture. The homogeneity enables simpler management and control of similar computational units, aggregated in multiple levels of hierarchy. Computational units are designed with minimal overhead as possible, where additional features and capabilities are aggregated at higher levels in the hierarchy. On-chip memory provides storage for content inherently required for basic operation at a particular hierarchy and is coupled with the computational resources in an optimal ratio. Lean control provides just enough signaling to manage only the operations required at a particular hierarchical level. Dynamic resource assignment agility is provided which can be adjusted as required depending on resource availability and capacity of the device.

Neural network processor incorporating inter-device connectivity

Structured weight based sparsity in an artificial neural network compiler

Structured weight based sparsity in an artificial neural network

Structured activation based sparsity in an artificial neural network

A novel and useful neural network (NN) processing core adapted to implement artificial neural networks (ANNs) and incorporating processing circuits having compute and local memory elements. The NN processor is constructed from self-contained computational units organized in a hierarchical architecture. The homogeneity enables simpler management and control of similar computational units, aggregated in multiple levels of hierarchy. Computational units are designed with minimal overhead as possible, where additional features and capabilities are aggregated at higher levels in the hierarchy. On-chip memory provides storage for content inherently required for basic operation at a particular hierarchy and is coupled with the computational resources in an optimal ratio. Lean control provides just enough signaling to manage only the operations required at a particular hierarchical level. Dynamic resource assignment agility is provided which can be adjusted as required depending on resource availability and capacity of the device.

Neural network processing element incorporating compute and local memory elements

A novel and useful system and method of data driven quantization optimization of weights and input data in an artificial neural network (ANN). The system reduces quantization implications (i.e. error) in a limited resource system by employing the information available in the data actually observed by the system. Data counters in the layers of the network observe the data input thereto. The distribution of the data is used to determine an optimum quantization scheme to apply to the weights, input data, or both. The mechanism is sensitive to the data observed at the input layer of the network. As a result, the network auto-tunes to optimize the instance specific representation of the network. The network becomes customized (i.e. specialized) to the inputs it observes and better fits itself to the subset of the sample space that is applicable to its actual data flow. As a result, nominal process noise is reduced and detection accuracy improves. In addition, the mechanism enables the reduction of the representation space and further reduces the memory (and energy thereof) needed to represent the network properties.

Data driven quantization optimization of weights and input data in an artificial neural network

A novel and useful system and method of accessing multi-dimensional data in memory. The invention is applicable to neural network (NN) processing engines adapted to implement artificial neural networks (ANNs). The NN processor is constructed from self-contained computational units organized in a hierarchical architecture. The homogeneity enables simpler management and control of similar computational units, aggregated in multiple levels of hierarchy. Computational units are designed with minimal overhead as possible, where additional features and capabilities are aggregated at higher levels in the hierarchy. On-chip memory provides storage for content inherently required for basic operation at a particular hierarchy and is coupled with the computational resources in an optimal ratio. Lean control provides just enough signaling to manage only the operations required at a particular hierarchical level. Dynamic resource assignment agility is provided which can be adjusted as required depending on resource availability and capacity of the device.

System and method of memory access of multi-dimensional data

A novel and useful neural network (NN) processing core adapted to implement artificial neural networks (ANNs) and incorporating configurable and programmable sliding window based memory access. The memory mapping and allocation scheme trades off random and full access in favor of high parallelism and static mapping to a subset of the overall address space. The NN processor is constructed from self-contained computational units organized in a hierarchical architecture. The homogeneity enables simpler management and control of similar computational units, aggregated in multiple levels of hierarchy. Computational units are designed with minimal overhead as possible, where additional features and capabilities are aggregated at higher levels in the hierarchy. On-chip memory provides storage for content inherently required for basic operation at a particular hierarchy and is coupled with the computational resources in an optimal ratio. Lean control provides just enough signaling to manage only the operations required at a particular hierarchical level. Dynamic resource assignment agility is provided which can be adjusted as required depending on resource availability and capacity of the device.

Configurable and programmable sliding window based memory access in a neural network processor

Novel and useful system and methods of several functional safety mechanisms for use in an artificial neural network (ANN) processor. The mechanisms can be deployed individually or in combination to provide a desired level of safety in neural networks. Multiple strategies are applied involving redundancy by design, redundancy through spatial mapping as well as self-tuning procedures that modify static (weights) and monitor dynamic (activations) behavior. The various mechanisms of the present invention address ANN system level safety in situ, as a system level strategy that is tightly coupled with the processor architecture. The NN processor incorporates several functional safety concepts which reduce its risk of failure that occurs during operation from going unnoticed. The mechanisms function to detect and promptly flag and report the occurrence of an error with some mechanisms capable of correction as well. The safety mechanisms cover data stream fault detection, software defined redundant allocation, cluster interlayer safety, cluster intralayer safety, layer control unit (LCU) instruction addressing, weights storage safety, and neural network intermediate results storage safety.

Neural network intermediate results safety mechanism in an artificial neural network processor

A novel and useful neural network (NN) processing core adapted to implement artificial neural networks (ANNs) and incorporating strictly separate control and data planes. The NN processor is constructed from self-contained computational units organized in a hierarchical architecture. The homogeneity enables simpler management and control of similar computational units, aggregated in multiple levels of hierarchy. Computational units are designed with minimal overhead as possible, where additional features and capabilities are aggregated at higher levels in the hierarchy. On-chip memory provides storage for content inherently required for basic operation at a particular hierarchy and is coupled with the computational resources in an optimal ratio. Lean control provides just enough signaling to manage only the operations required at a particular hierarchical level. Dynamic resource assignment agility is provided which can be adjusted as required depending on resource availability and capacity of the device.

Neural network processor incorporating separate control and data fabric

Layer control unit instruction addressing safety mechanism in an artificial neural network processor

A novel and useful augmented artificial neural network (ANN) incorporating an existing artificial neural network (ANN) coupled to a supplemental ANN and a first-in first-out (FIFO) stack for storing historical output values of the network. The augmented ANN exploits the redundant nature of information present in an input data stream. The addition of the supplemental ANN along with a FIFO enables the augmented network to look back into the past in making a decision for the current frame. It provides context aware object presence as well as lowers the rate of false detections and misdetections. The output of the existing ANN is stored in a FIFO to create a lookahead system in which both past output values of the supplemental ANN and 'future' values of the output of the existing ANN are used in making a decision for the current frame. In addition, the mechanism does not require retraining the entire neural network nor does it require data set labeling.

Growing community of inventors

Kiryat Ono, Israel

Or Danon