Correlated electron switch elements for brain-based computing
Broadly speaking, the present techniques exploit the properties of correlated electron materials for artificial neural networks and neuromorphic computing. In particular, the present techniques provide apparatuses/devices that comprise at least one correlated electron switch (CES) element and which may be used as, or to form, an artificial neuron or an artificial synapse.
1. An apparatus for an artificial neural network, the apparatus comprising:
at least one correlated electron switch (CES) element to be configurable to:
store one or more synaptic weights;
operate in a high impedance state;
operate in one or more low impedance states;
establish a threshold current responsive to transition of the at least one CES element to one of the one or more low impedance states;
transition to the high impedance state responsive to a current in the at least one CES element meeting or exceeding the threshold current; and
accumulate currents inputted to the apparatus as an accumulated current and output a signal responsive to the accumulated current equaling or exceeding the established threshold current.
2. The apparatus as claimed in claim 1 wherein an impedance value of the high impedance state is at least 100 times larger than an impedance value of each of the one or more low impedance states.
3. The apparatus as claimed in claim 1 wherein, the at least one CES element is to be programmable to store a synaptic weight of the one or more synaptic weights as one of the one or more low impedance states depending on a compliance current applied to the at least one CES element.
4. The apparatus as claimed in claim 1 wherein, if configured as a synapse, the apparatus is adaptable to be in a low power consumption state responsive to placement of the at least one CES element in the high impedance state.
5. The apparatus as claimed in claim 1 wherein, if the at least one CES element is configured as a neuron, the at least one CES element is configured to transition to the high impedance state responsive to a reset voltage applied across the at least one CES element and a reset current density in the at least one CES element.
6. The apparatus as claimed in claim 1 wherein, if configured as a neuron, the at least one CES element configured to transition to a low impedance state of the one or more low impedance states responsive to a set voltage applied across the at least one CES element and a set current density in the at least one CES element.
7. The apparatus as claimed in claim 1 wherein the at least one CES element is capable of performing the transition to the high impedance state within 1.0 ms, and wherein the at least one CES element is capable of performing the transition to a low impedance state of the one or more low impedance states within 1.0 ms.
8. The apparatus as recited in claim 1 , wherein the apparatus is adapted to form a crosspoint array.
9. The apparatus as recited in claim 1 , wherein the apparatus is adapted to form a layered crosspoint array.
10. The apparatus of claim 1 , wherein the artificial neural network comprises a three-dimensional (3D) configurable array comprising:
at least two layers, each layer comprising an array of neurons, wherein the layers are arranged in a stack; and
electrical coupling between adjacent layers of the stack;
wherein a size of the 3D configurable array is adjustable by disconnecting or connecting layers of the stack.
11. The apparatus of claim 10 , wherein the 3D configurable array is to provide a synapse, and wherein a size of the synapse is adjustable.
12. The apparatus claimed in claim 1 wherein the outputted signal is to be a portion of a sum of a dot-product computation of synaptic weights and input signals received at the artificial neural network, the outputted signal is to provide an input signal to at least one second neuron of a plurality of neurons.
13. A configurable array for an artificial neural network, the configurable array comprising:
a plurality of apparatuses, each apparatus comprising one or more correlated electron switch (CES) elements, individual CES elements to be capable of functioning as either a synapse or as a neuron; and
a switching mechanism to selectively configure at least one of the one or more CES elements to function as the synapse or the neuron.
14. The configurable array as claimed in claim 13 wherein responsive to a CES element of one of the plurality of apparatuses being switched to function as a synapse, the CES element is capable of:
storing synaptic weights;
operating in a high impedance state; and
operating in one of a plurality of low impedance states.
15. The configurable array as claimed in claim 13 , wherein responsive to a CES element of one of the plurality of apparatuses being switched to function as a neuron, the CES element is capable of:
a transition to a high impedance state responsive to occurrence of a reset condition;
a transition to one of a plurality of low impedance states responsive to occurrence of a set condition; and
accumulate currents input into an apparatus of the plurality of apparatuses as an accumulated current.
16. The configurable array as claimed in claim 13 wherein each apparatus of the plurality of apparatuses is selectable, and wherein a size of the configurable array is adjustable by:
a selection of each of the plurality of apparatuses; and
a coupling of the selected apparatus into the configurable array or a decoupling of the selected apparatus from the configurable array.
17. The configurable array of claim 13 , wherein if at least a first one of the individual CES elements is configured as a synapse, the at least the first one of the individual CES elements configured to store a synaptic weight corresponding to a compliance current.
18. The configurable array of claim 17 , wherein if at least a second of the individual CES elements is configured as a neuron, the at least the second of the individual CES elements is capable of conducting a corresponding compliance current.
19. An apparatus for an artificial neural network, the apparatus comprising one or more correlated electron switch (CES) elements, wherein
at least a first CES element is adapted to:
establish a threshold current responsive to the first CES element transitioning to low impedance state; and
provide an output signal responsive to a transition of the first CES element to a high impedance state, wherein:
the transition of the first CES element is to occur responsive to a current in the first CES element equaling or exceeding the threshold current; and
the output signal is based, at least in part, on one or more stored synaptic weights.
20. The apparatus of claim 19 , wherein the one or more CES elements comprise a plurality of CES elements, and wherein the one or more synaptic weights are stored in one or more second CES elements of the plurality of CES elements.
21. An apparatus for an artificial neural network, the apparatus comprising a crosspoint array of a plurality of neurons, a first neuron of the plurality of neurons to comprise at least one correlated electron switch (CES) element, wherein the first neuron is adapted to:
establish a threshold current responsive to a transition of the at least one CES element to a low impedance state of one or more low impedance states;
transition to a high impedance state responsive to a current in the at least one CES element equaling or exceeding the threshold current; and
accumulate currents input to the first neuron as an accumulated current and output a signal responsive to the accumulated current equaling or exceeding the threshold current.