Abstract: Examples described herein relate to a decision tree computation system in which a hardware accelerator for a decision tree is implemented in the form of an analog Content Addressable Memory (a-CAM) array. The hardware accelerator accesses a decision tree. The decision tree comprises of multiple paths and each path of the multiple paths includes a set of nodes. Each node of the decision tree is associated with a feature variable of multiple feature variables of the decision tree. The hardware accelerator combines multiple nodes among the set of nodes with a same feature variable into a combined single node. Wildcard values are replaced for feature variables not being evaluated in each path. Each combined single node associated with each feature variable is mapped to a corresponding column in the a-CAM array and the multiple paths of the decision tree to rows of the a-CAM array.

Abstract: Systems and methods are provided for implementing a hardware accelerator. The hardware accelerator emulates a neural network, and includes a memristor crossbar array, and a non-linear filter. The memristor crossbar array can be programmed to calculate node values of the neural network. The nodes values can be calculated in accordance with rules to reduce an energy function associated with the neural network. The non-linear filter is coupled to the memristor crossbar array and programmed to harness noise signals that may be present in analog circuitry of the hardware accelerator. The noise signals can be harnessed such that the energy function associated with the neural network converges towards a global minimum and modifies the calculated node values. In some embodiments, the non-liner filter is implemented as a Schmidt trigger comparator.

Abstract: Recurrent neural networks, and methods therefor, are provided with diagonal and programming fluctuation to find energy global minima. The method may include storing the matrix of weights in memory cells of a crossbar array of a recursive neural network prior to operation of the recursive neural network; altering the weights according to a probability distribution; setting the weights to non-zero values in at least one of the memory cells in a diagonal of the memory cells in the crossbar array; and operating the recursive neural network.

Abstract: Systems are provided for implementing a hardware accelerator. The hardware accelerator emulate a stochastic neural network, and includes a first memristor crossbar array, and a second memristor crossbar array. The first memristor crossbar array can be programmed to calculate node values of the neural network. The nodes values can be calculated in accordance with rules to reduce an energy function associated with the neural network. The second memristor crossbar array is coupled to the first memristor crossbar array and programmed to introduce noise signals into the neural network. The noise signals can be introduced such that the energy function associated with the neural network converges towards a global minimum and modifies the calculated node values.

Abstract: In one example, a device to process analog sensor data is described. For example, a device may include at least one analog sensor to generate a first set of analog voltage signals and a crossbar array including a plurality of memristors. In one example, the crossbar array is to receive an input vector of the first set of analog voltage signals, generate an output vector comprising a second set of analog voltage signals that is based upon a dot product of the input vector and a matrix comprising resistance values of the plurality of memristors, detect a pattern of the output vector, and activate a processor upon a detection of the pattern.

Abstract: Encoding of domain logic rules in an analog content addressable memory (aCAM) is disclosed. By encoding domain logic in an aCAM, rapid and flexible search capabilities are enabled, including the capability to search ranges of analog values, fuzzy match capabilities, and optimized parameter search capabilities. This is achieved with low latency by using only a small number of clock cycles at low power. A domain logic ruleset may be represented using various data structures such as decision trees, directed graphs, or the like. These representations can be converted to a table of values, where each table column can be directly mapped to a corresponding row of the aCAM.

Abstract: Systems and methods are configured to provide a first problem to be solved to a network of memristors. A second problem to be solved can be gradually provided to the network of memristors. Controlled noise can be applied to the network of memristors for at least a portion of time during which the second problem is “gradually” provided to the network of memristors. A solution to the second problem can be determined.

Abstract: Staged neural networks and methods are described herein. In some embodiments, the methods may identify a plurality of second NP hard/complete problems that are similar to the first NP hard/complete problem and identify solutions to the second NP hard/complete problems. The methods may train a deep neural network with the second NP hard/complete problems and the solutions. The methods may provide the first NP hard/complete problem to the trained deep neural network to generate a preliminary solution to the first NP hard/complete problem and provide the preliminary solution to a recursive neural network configured to execute an energy minimization search. The recursive neural network may generate a final solution to the problem based on the preliminary solution.

Abstract: Systems, devices, circuits, methods, and non-transitory computer readable media that enable storing and searching arbitrary segments of ranges of analog values are disclosed. Various analog content addressable memory (aCAM) circuit implementations having the capability to store and search outside of a range of values, within any of multiple disjoint ranges, or outside of multiple ranges are disclosed. The disclosed aCAM circuit implementations make searching for complex input features more flexible and efficient, thereby yielding a technological improvement over conventional solutions. In some implementations, an aCAM may include multiple pull-down transistors connected in series to a match line that is pre-charged, in which case, the aCAM detects a match if the match line is not discharged by the pull-down transistors, which occurs if at least one pull-down transistor is in an OFF state. In other implementations, an aCAM includes pass gates connected to a match line to detect a match.

Abstract: Systems and methods are provided for implementing hardware optimization for a hardware accelerator. The hardware accelerator emulates a neural network. Training of the neural network integrates a regularized pruning technique to systematically reduce a number of weights. A crossbar array included in hardware accelerator can be programmed to calculate node values of the pruned neural network to selectively reduce the number of weight column lines in the crossbar array. During deployment, the hardware accelerator can be programmed to power off periphery circuit elements that correspond to a pruned weight column line to optimize the hardware accelerator for power. Alternatively, before deployment, the hardware accelerator can be optimized for area by including a finite number of weight column line. Then, regularized pruning of the neural network selectively reduces the number of weights for consistency with the finite number of weight columns lines in the hardware accelerator.

Abstract: A DPE memristor crossbar array system includes a plurality of partitioned memristor crossbar arrays. Each of the plurality of partitioned memristor crossbar arrays includes a primary memristor crossbar array and a redundant memristor crossbar array. The redundant memristor crossbar array includes values that are mathematically related to values within the primary memristor crossbar array. In addition, the plurality of partitioned memristor crossbar arrays includes a block of shared analog circuits coupled to the plurality of partitioned memristor crossbar arrays. The block of shared analog circuits is to determine a dot product value of voltage values generated by at least one partitioned memristor crossbar array of the plurality of partitioned memristor crossbar arrays.

Abstract: Systems, devices, circuits, methods, and non-transitory computer readable media that enable storing and searching arbitrary segments of ranges of analog values are disclosed. Various analog content addressable memory (aCAM) circuit implementations having the capability to store and search outside of a range of values, within any of multiple disjoint ranges, or outside of multiple ranges are disclosed. The disclosed aCAM circuit implementations make searching for complex input features more flexible and efficient, thereby yielding a technological improvement over conventional solutions. In some implementations, an aCAM may include multiple pull-down transistors connected in series to a match line that is pre-charged, in which case, the aCAM detects a match if the match line is not discharged by the pull-down transistors, which occurs if at least one pull-down transistor is in an OFF state. In other implementations, an aCAM includes pass gates connected to a match line to detect a match.

Abstract: An analog content addressable memory (aCAM) that enables parallel searching of analog ranges of values and generates analog outputs that quantify matches between input data and stored data is disclosed. The input data can be compared with the stored data, and the input data can be determined to match the stored data based on a value associated with the input data being within a range associated with the stored data. The aCAM can generate an analog output that represents a number of matches and a number of mismatches between the input data and the stored data. Based on the analog output, whether the input data matches the stored data and a degree to which the input data matches the stored data can be determined.

Abstract: Examples described herein relate to a security system consistent with the disclosure. For instance, the security system may comprise a sensor interface bridge connecting a gateway to an input/output (I/O) card, a Field Programmable Gate Array (FPGA) to scan data to detect an anomaly in the data while the data is in the sensor interface bridge, where a learning neural network accelerator Application-Specific Integrated Circuit (ASIC) is integrated with the FPGA and send the data without an anomaly to the gateway.

Abstract: Examples disclosed herein relate to digital hash code generation. A digital hash code generating device comprising a plurality of variable conductance elements. Each variable conductance element is coupled to a selected row line and to a selected column line of a crossbar array. Each variable conductance element comprises a conductance from a stochastic distribution of conductance. A plurality of comparator elements and each comparator element is connected to a set of at least two column lines. The plurality of comparator elements generates a hash code in response to a vector input applied to the plurality of row lines of the crossbar array.

Abstract: Disclosed techniques provide for dynamically changing precision of a multi-stage compute process. For example, changing neural network (NN) parameters on a per-layer basis depending on properties of incoming data streams and per-layer performance of an NN among other considerations. NNs include multiple layers that may each be calculated with a different degree of accuracy and therefore, compute resource overhead (e.g., memory, processor resources, etc.). NNs are usually trained with 32-bit or 16-bit floating-point numbers. Once trained, an NN may be deployed in production. One approach to reduce compute overhead is to reduce parameter precision of NNs to 16 or 8 for deployment. The conversion to an acceptable lower precision is usually determined manually before deployment and precision levels are fixed while deployed.

Abstract: An optical device includes a light-emitting device integrated with a memory device. The memory device include a first electrode and a second electrode, and the light-emitting device includes a third electrode and the second electrode. In such configuration, a first voltage between the second electrode and the third electrode causes the light-emitting device to emit light of a first wavelength, and a second voltage between the first electrode and the second electrode while the memory device is at OFF state causes the light-emitting device to emit light of a second wavelength shorter than the first wavelength or while the memory device is at ON state causes the light-emitting device to emit light of a third wavelength longer than the first wavelength.

Abstract: Encoding of domain logic rules in an analog content addressable memory (aCAM) is disclosed. By encoding domain logic in an aCAM, rapid and flexible search capabilities are enabled, including the capability to search ranges of analog values, fuzzy match capabilities, and optimized parameter search capabilities. This is achieved with low latency by using only a small number of clock cycles at low power. A domain logic ruleset may be represented using various data structures such as decision trees, directed graphs, or the like. These representations can be converted to a table of values, where each table column can be directly mapped to a corresponding row of the aCAM.

Abstract: Devices and methods are provided. In one aspect, a device for driving a memristor array includes a substrate including a well having a bottom layer, a first wall and a second wall. The substrate is formed of a strained layer of a first semiconductor material. A vertical JFET is formed in the well. The vertical JFET includes a vertical gate region formed in a middle portion of the well with a gate region height less than a depth of the well. A channel region is formed of an epitaxial layer of a second semiconductor wrapped around the vertical gate region. Vertical source regions are formed on both sides of a first end of the vertical gate region, and vertical drain regions are formed on both sides of a second end of the vertical gate region.

Abstract: A fault-tolerant analog computing device includes a crossbar array having a number l rows and a number n columns intersecting the l rows to form l×n memory locations. The l rows of the crossbar array receive an input signal as a vector of length l. The n columns output an output signal as a vector of length n that is a dot product of the input signal and the matrix values defined in the l×n memory locations. Each memory location is programmed with a matrix value. A first set of k columns of the n columns is programmed with continuous analog target matrix values with which the input signal is to be multiplied, where k<n. A second set of m columns of the n columns is programmed with continuous analog matrix values for detecting an error in the output signal that exceeds a threshold error value, where m<n.