Abstract: A reservoir calculation device according to an embodiment includes a reservoir circuit and an output circuit. The reservoir circuit receives input data and outputs intermediate signals, each undergoing a transient change when the input data changes. The output circuit outputs an output signal obtained by combining the intermediate signals. The reservoir circuit includes intermediate circuits, each including a neuron circuit and an intermediate output circuit. The neuron circuit generates an intermediate voltage undergoing a transient change corresponding to weight data and the input data when the input data changes. The intermediate output circuit outputs an intermediate signal representing a level of the intermediate voltage from the neuron circuit. The neuron circuit includes a time constant circuit capable of changing a time constant. The time constant circuit is connected between a reference potential and an intermediate terminal outputting the intermediate voltage.

Abstract: According to an embodiment, an inference system includes a recurrent neural network circuit, an inference neural network, and a control circuit. The recurrent neural network circuit receives M input signals and outputs N intermediate signals, where M is an integer of 2 or more and N is an integer of 2 or more. The inference neural network circuit receives the N intermediate signals and outputs L output signals, where L is an integer of 2 or more. The control circuit adjusts a plurality of coefficients that are set to the recurrent neural network circuit and adjusts a plurality of coefficients that are set to the inference neural network circuit. The control circuit adjusts the coefficients set to the recurrent neural network circuit according to a total delay time period from timing for applying the M input signals until timing for firing the L output signals.

Abstract: In a gate electrode of a nonvolatile memory device of an embodiment, a tunnel insulating film covers a channel region. A first current collector file is disposed on the side opposite to the channel region with respect to the tunnel insulating film. An ion conductor film is disposed between the tunnel insulating film and the first current collector film. A first electrode film is disposed between the tunnel insulating film and the ion conductor film. The first electrode film is in contact with the ion conductor film. A second electrode film is disposed between the ion conductor film and the first current collector film. The second electrode film is in contact with the ion conductor film. A second current collector film is disposed between the tunnel insulating film and the second electrode film.

Abstract: A multiplication device according to one embodiment includes a short-term memory circuit, a long-term memory circuit, a conversion circuit, and a control circuit. The short-term memory circuit generates a first control voltage in accordance with a weight value. The long-term memory circuit generates a second control voltage by a circuit with a larger time constant than the short-term memory circuit. The conversion circuit outputs an output current by multiplying an input voltage by a conductance. The output current is output by that, the first control voltage is applied to a control terminal of the conversion circuit, and an input voltage according to an input value is applied to an input terminal of the conversion circuit. The control circuit executes a calibration process of matching the first control voltage with the second control voltage by transferring an electric charge from the long-term memory circuit to the short-term memory circuit.

Abstract: According to an embodiment, an inference system includes a recurrent neural network circuit, an inference neural network, and a control circuit. The recurrent neural network circuit receives M input signals and outputs N intermediate signals, where M is an integer of 2 or more and N is an integer of 2 or more. The inference neural network circuit receives the N intermediate signals and outputs L output signals, where L is an integer of 2 or more. The control circuit adjusts a plurality of coefficients that are set to the recurrent neural network circuit and adjusts a plurality of coefficients that are set to the inference neural network circuit. The control circuit adjusts the coefficients set to the recurrent neural network circuit according to a total delay time period from timing for applying the M input signals until timing for firing the L output signals.

Abstract: A neural network apparatus according to an embodiment includes neuron circuits, synaptic circuits, and a control circuit. A firing circuit of each neuron circuit outputs a firing signal when absolute value of the internal potential is larger than a firing threshold. A firing threshold adjustment circuit of each neuron circuit changes the firing threshold in accordance with frequency of the firing signal. When the firing signal is output from a pre-synaptic neuron circuit, the synaptic circuit changes the synaptic weight in accordance with a contrast between a learning threshold and the absolute value of the internal potential held in a post-synaptic neuron circuit. The control circuit changes the learning threshold in accordance with frequency of the firing signal from a target neuron circuit. The learning threshold is used for changing the synaptic weight stored in one or more synaptic circuits each outputting the output signal to the target neuron.

Abstract: According to an embodiment, an operation apparatus includes a first neural network, a second neural network, an evaluation circuit, and a coefficient-updating circuit. The first neural network performs an operation in a first mode. The second neural network performs an operation in a second mode and has a same layer structure as the first neural network. The evaluation circuit evaluates an error of the operation of the first neural network in the first mode and evaluates an error of the operation of the second neural network in the second mode. The coefficient-updating circuit updates, in the first mode, coefficients set for the second neural network based on an evaluating result of the error of the operation of the first neural network, and updates, in the second mode, coefficients set for the first neural network based on an evaluating result of the error of the operation of the second neural network.

Abstract: According to an embodiment, a synapse circuit includes: a buffer that changes an output signal to a second logical value at a timing when an input signal exceeds a first threshold level, in a case where the output signal has a first logical value in a first mode, and changes the output signal to the second logical value at a timing when the input signal exceeds a reference level lower than the first threshold level, in a case where the output signal has the first logical value in a second mode; an adjusting unit that adjusts the first threshold level depending on a stored coefficient; and a mode switching unit that operates the buffer in the first mode during a period in which an acquired spike is not generated, and operates the buffer in the second mode during a period in which the spike is generated.

Abstract: A spiking neural network device according to an embodiment includes a synaptic element, a neuron circuit, a synaptic potentiator, and a synaptic depressor. The synaptic element has a variable weight. The neuron circuit inputs a spike voltage having a magnitude adjusted in accordance with the weight of the synaptic element via the synaptic element, and fires when a predetermined condition is satisfied. The synaptic potentiator performs a potentiating operation for potentiating the weight of the synaptic element depending on input timing of the spike voltage and firing timing of the neuron circuit. The synaptic depressor performs a depression operation for depressing the weight of the synaptic element in accordance with a schedule independent from the input timing of the spike voltage and the firing timing of the neuron circuit.

Abstract: According to an embodiment, a neural network apparatus includes cores, routers, a tree path, and a short-cut path. The cores are provided according to leaves in a tree structure, each core serving as a circuit that performs calculation or processing for part of elements of the neural network. The routers are provided according to nodes other than the leaves in the tree structure. The tree path connects the cores and the routers such that data is transferred along the tree structure. The short-cut path connects part of the routers such that data is transferred on a route differing from the tree path. The routers transmit data output from each core to any of the cores serving as a transmission destination on one of routes in the tree path and the short-cut path such that the calculation or the processing is performed according to a structure of the neural network.

Abstract: A variable resistance element according to an embodiment serves to change to a low resistance state or a high resistance state. The variable resistance element includes a first transition metal compound layer, a second transition metal compound layer, and a lithium ion conductor layer. The first transition metal compound layer is connected to a first electrode. The first transition metal compound layer is a metal compound containing lithium ions in lattice interstices. The second transition metal compound layer is connected to a second electrode. The second transition metal compound layer is a metal compound containing lithium ions in lattice interstices. The lithium ion conductor layer is provided between the first transition metal compound layer and the second transition metal compound layer. The lithium ion conductor layer is a solid substance that is permeable to lithium ions and is less permeable to electrons.

Abstract: A synaptic circuit according to an embodiment includes a weight storage circuit and a transmission circuit. The weight storage circuit stores a synaptic weight indicating a first value or a second value. The transmission circuit receives a firing signal output from a pre-synaptic neuron circuit, and supplies an output signal to a post-synaptic neuron circuit. The output signal is obtained by adding, to the firing signal, influence of the synaptic weight. The post-synaptic neuron circuit holds an internal potential. When the firing signal is received, the weight storage circuit causes the synaptic weight to change to indicate the first or second value with a first probability in accordance with a comparison result between the internal potential and a set potential. The weight storage circuit causes the synaptic weight to change to indicate the first value with a second probability regardless of whether or not the firing signal is received.

Abstract: According to an embodiment, a reinforcement learning system includes a memristor array in which each of a plurality of first direction lines corresponds to one of a plurality of states, and each of a plurality of second direction lines corresponds to one of a plurality of actions, a first voltage application unit that individually applies voltage to the first direction lines, a second voltage application unit that individually applies voltage to the second direction lines, a action decision circuit that decides action to be selected by an agent in a state corresponding to a first direction line to which a readout voltage is applied, a action storage unit that stores action selected by the agent in each state that can be caused in an environment, and a trace storage unit that stores a time at which the state is caused by action selected by the agent.

Abstract: According to an embodiment, a neural network apparatus includes a plurality of neuron circuits, each including an integration circuit, a firing circuit, and a secondary battery. The integration circuit is configured to output an integral signal obtained by integrating input signals. The firing circuit is configured to generate, in accordance with the integral signal, a pulse signal to be transmitted to the neuron circuit provided at a subsequent layer. The secondary battery is configured to supply the firing circuit with drive electric power used for generating the pulse signal.

Abstract: According to an embodiment, an inference system includes a recurrent neural network circuit, an inference neural network, and a control circuit. The recurrent neural network circuit receives M input signals and outputs N intermediate signals, where M is an integer of 2 or more and N is an integer of 2 or more. The inference neural network circuit receives the N intermediate signals and outputs L output signals, where L is an integer of 2 or more. The control circuit adjusts a plurality of coefficients that are set to the recurrent neural network circuit and adjusts a plurality of coefficients that are set to the inference neural network circuit. The control circuit adjusts the coefficients set to the recurrent neural network circuit according to a total delay time period from timing for applying the M input signals until timing for firing the L output signals.

Abstract: According to an embodiment, a neural network device includes a plurality of cores, and a plurality of routers. Each of the plurality of routers includes an input circuit and an output circuit. Each of the plurality of cores transmits at least one of forward direction data propagating in the neural network in a forward direction and reverse direction data propagating in the neural network in a reverse direction. The input circuit receives the forward direction data and the reverse direction data from any one of the plurality of cores and the plurality of routers. The output circuit or the input circuit selectively deletes the reverse direction data stored based on a request signal for requesting reception of data.

Abstract: According to an embodiment, an arithmetic device includes a comparator, M cross switches, and M coefficient circuits. The comparator compares a first voltage generated at a first comparison terminal and a second voltage generated at a second comparison terminal. The M cross switches are provided corresponding to the M input signals. The M coefficient circuits are provided corresponding to the M coefficients, and each includes a first constant current source and a second constant current source. Each of the M cross switches performs switching between a straight state and a reverse state. In each of the M coefficient circuits, the first constant current source is connected between a positive output terminal of the corresponding coefficient circuit and a reference potential, and the second constant current source is connected between a negative output terminal of the corresponding coefficient circuit and the reference potential.

Abstract: According to an embodiment, a neural network device includes: a plurality of cores each executing computation and processing of a partial component in a neural network; and a plurality of routers transmitting data output from each core to one of the plurality of cores such that computation and processing are executed according to structure of the neural network. Each of the plurality of cores outputs at least one of a forward data and a backward data propagated through the neural network in a forward direction and a backward direction, respectively. Each of the plurality of routers is included in one of a plurality of partial regions each being a forward region or a backward region. A router included in the forward region and a router included in the backward region transmit the forward data and the backward data to other routers in the same partial regions, respectively.

Abstract: A memory device according to an embodiment can be used for storing weights for a neural network. An update circuit changes a difference between charge amounts accumulated in first/second accumulation circuits in the memory device. An output circuit outputs, as a weight, a signal corresponding to the difference between the charge amounts. The update circuit performs the change of the difference by changing, when the update amount is positive, the electric charges accumulated in the first accumulation circuit in a first direction by a charge amount corresponding to an absolute value of the update amount, the first direction being either an increasing direction or a decreasing direction, and changing, when the update amount is negative, the electric charges accumulated in the second accumulation circuit in the first direction by a charge amount corresponding to an absolute value of the update amount.

Abstract: A synaptic circuit according to an embodiment is a circuit in which a weight value changed by learning is set. The synaptic circuit receives a binary input signal from a pre-synaptic neuron circuit and outputs an output signal to a post-synaptic neuron circuit. The synaptic circuit includes a propagation circuit and a control circuit. The propagation circuit supplies, to the post-synaptic neuron circuit, the output signal obtained by adding an influence of the weight value to the input signal. The control circuit stops output of the output signal from the propagation circuit to the post-synaptic neuron circuit when the weight value is smaller than a predetermined reference value.