CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The present application is a non-provisional of, and claims priority to, U.S. Provisional Application No. 62/014,201 , filed on June 19, 2014 and titled "Memri stive Neural Networks," which is incorporated by reference herein in its entirety.

BACKGROUND

[0002] A memristor is a two-terminal device that changes its resistance in response to the amount of electrical current that has previously flown through the device. Memristors may be used in crossbar neural network architectures. In a crossbar neural network, multiple memristors are connected in a perpendicular crossbar array with memristor synapses at each crossing. However, crossbar neural network architectures may require the use of complex designs in order to counteract parasitic leak paths. Additionally, redundant synapses do not exist in crossbar neural networks. Furthermore, a recurrent connection in a crossbar neural network requires complex circuit layouts, and from a footprint point of view, crossbar designs scale quadratically in size with the number of neurons.

BRIEF DESCRIPTION OF THE DRAWINGS

[0003] FIG. 1 is a drawing of a core-shell memristive nanofiber according to various embodiments of the present disclosure.

[0004] FIG. 2 is a drawing of an example of a nanofiber-based memristive neural network according to various embodiments of the present disclosure. [0005] FIG. 3 is a drawing of an example of a nanofiber-based memristive neural network according to various embodiments of the present disclosure.

[0006] FIG. 4 is a drawing of an example of a simulation of a circuit layout of a nanofiber-based memristive neural network according to various embodiments of the present disclosure.

[0007] FIG. 5 is a flowchart illustrating an example of a method of creating a nanofiber-based memristive neural network according to various embodiments of the present disclosure.

DETAILED DESCRIPTION

[0008] The present disclosure is directed towards neural networks that use memristive fibers. Generally, a neural network may comprise populations of simulated neurons with weighted connections between them. A neural network in accordance with various embodiments of the present disclosure may comprise an array of neural nodes that are interconnected using randomized connections of memristive fibers. Such a neural node may comprise, for example, a Complimentary Metal-Oxide-Semiconductor (CMOS) Leaky Integrate-and-Fire (LIF) neural circuit, or any other suitable type of neural circuit. Each neural node may output one or more signals that are responsive to one or more input signals that the neural node has received. For example, upon one or more input current signals reaching a threshold value, a neural node may output a voltage spike to one or more output paths.

[0009] As mentioned above, a neural network may also comprise memristive nanofibers. Memristive nanofibers may be used to form artificial synapses in neural networks. Each memristive nanofiber may couple one or more neural nodes to one or more other neural nodes. In this way, one or more output signals may be transmitted from a particular neural node to one or more other neural nodes. The particular neural nodes to which particular memristive nanofibers are connected may be randomized. In this regard, the particular neural nodes to which the memristive nanofibers are connected are not predetermined prior to the memristive fibers being connected to the one or more neural nodes. As a result of the connections being randomized, the network obtained may exhibit sparse, random connectivity, which has been shown to increase the performance and efficiency of neural networks. Thus, the neural network may be used, for example, to model a Liquid State Machine (LSM). Further description regarding LSMs is provided in Wolfgang Maass et al., Real-Time Computing without Stable States: A New Framework for Neural Computation Based on Perturbations, Neural Computation (Volume 14, Issue 1 1 ) (Nov. 1 1 , 2002), which is incorporated by reference herein in its entirety.

[0010] Each memristive nanofiber of the memristive neural network may comprise a conductive core, a memristive shell, and one or more electrodes. Memristive nanofibers having a conductive core, memristive shell, and one or more electrodes may be formed using electrospinning or any other suitable method. An electrode of the memristive nanofiber may serve as a conductive attachment point between the memristive nanofiber and an input or output terminal of a neural node. The conductive core of the memristive nanofiber in some embodiments may comprise T1O2 and/or any other suitable material. The memristive shell may at least partially surround the conductive core and thereby form a synapse between two or more neural nodes. In this regard, the memristive shell may cause the memristive nanofiber to form a connection that increases or decreases in strength in response to the past signals that have traveled through the memristive nanofiber. The memristive shell may comprise Ti0 ₂ and/or any other suitable material with memristive properties.

[0011] As previously mentioned, in some embodiments, the memristive nanofibers in the memristive neural network may form randomized connections between the neural nodes. Thus, the probability of two neurons being connected decreases as the distance between neural nodes increases. Additionally, patterned electric fields may be used so that particular connection types are more likely to be formed between neural nodes when the connections are made. Additionally, a neural network may be formed using patterned electric fields or other suitable methods so that multiple layers of memristive nanofibers are created. Such a neural network may also comprise connections that facilitate transmission of signals between various layers. In one particular embodiment, the layers and communication paths between layers are modeled after a neocortex of a brain.

[0012] Memristive neural networks in accordance with various embodiments of the present disclosure may provide various types of benefits. For example, such a memristive neural network may be capable of spike-timing-dependent plasticity (STDP). Additionally, the memristive neural network may comprise random, spatially dependent connections. Furthermore, the memristive neural network may comprise inhibitory outputs and/or recurrent connections. As such, the memristive neural networks in accordance with various embodiments of the present disclosure may have properties that are similar to biological neural networks.

[0013] With reference to FIG. 1 , shown is an example of a core-shell memristive nanofiber 100 according to various embodiments of the present disclosure. Each memristive nanofiber 100 of a memristive neural network may comprise one or more electrodes 103, a conductive core 106, and a memristive shell 109. An electrode 103 of the memristive nanofiber 100 may serve as a conductive attachment point between the memristive nanofiber 100 and an input or output terminal of a neural node. The conductive core 106 of the memristive nanofiber 100 in some embodiments may comprise ΤΊΟ2 and/or any other suitable material with memristive properties. The memristive shell 109 may at least partially surround the conductive core 106 and thereby form a synapse between two or more neural nodes. In this regard, the memristive shell 109 may cause the memristive nanofiber 100 to form a connection that increases or decreases in strength in response to the past signals that have traveled through the memristive nanofiber 100. The memristive shell 109 may comprise Ti0 ₂ and/or any other suitable material, such as polyaniline.

[0014] With reference to FIG. 2, shown is an example of a nanofiber-based memristive neural network 200 according to various embodiments of the present disclosure. The memristive nanofibers 100 (FIG. 1 ) can be used as memristive connections 206A-206J between CMOS-based neuron arrays 203A-203E in the nanofiber-based memristive neural network 200. The memristive nanofibers 100 may form randomized memristive connections 206A-206J between inputs 212A- 212E and outputs 209A-209E of the CMOS-based neuron arrays 203A-203E.

[0015] With reference to FIG. 3, shown is a drawing of an example of a nanofiber-based memristive neural network 200 according to various embodiments of the present disclosure. The nanofiber-based memristive neural network 200 depicts examples of memristive nanofibers 100, referred to herein as memristive nanofibers 100A-100D, comprising conductive cores 106, referred to herein as the conductive cores 106A-106D, and memristive shells 109, referred to herein as the memristive shells 109A-109D. The memristive nanofibers 100A-100D may be used as memristive connections 206A-206J between CMOS neurons 327A and 327B located on the silicon substrate 330. Each memristive shell 109A-109D partially surrounds each conductive core 106A-106D and thereby forms a synapse 318A-318D between two or more neural nodes. The input electrodes 321 A and 321 B and output electrodes 324A and 324B may serve as conductive attachment points between memristive nanofibers 100A-100D and input terminals 212A-212E or output terminals 209A-209E of the neural nodes.

[0016] With reference to FIG. 4, shown is a simulation 403 of an example of a circuit 406 of a nanofiber-based memristive neural connection according to various embodiments of the present disclosure. The circuit 406 depicts a nanofiber-based memristive neural connection including memristive shells 109A-109D, voltage source V1 421 , voltage source V2 409, resistor R3 412, resistor R4 415, and resistor R1 418. The simulation 403 shows that driving current through a nanofiber- based memristive neural connection will not cause the effects from opposing memristors on a nanofiber to cancel each other out.

[0017] With reference to FIG. 5, shown is a flowchart illustrating one example of a method of creating a nanofiber-based memristive neural network 200 (FIG. 2) according to various embodiments of the present disclosure. Beginning with box 503, stoichiometric nanofibers are synthesized using a precursor. The stoichiometric nanofibers may comprise, for example, Ti0 ₂ and/or any other suitable material. The precursor may be for example, titanium isopropoxide, titanium butoxide, or another suitable precursor.

[0018] Next, in box 506, a core-shell memristive nanofiber 100 (FIG. 1 ) is created. The core-shell memristive nanofiber 100 (FIG. 1 ) may be created by electrospinning a stoichiometric Ti0 ₂ outer shell 109 with a doped conductive Ti0 ₂ -x core 106. Then, in box 509, electrodes 103 may be deposited on the core-shell memristive nanofiber 100 (FIG. 1 ).

[0019] Next, in box 512, memristive properties are verified and a spike-timing dependent plasticity is implemented to create a computational model based on the nanofiber network response. Then, in box 515, a physical prototype of a memristive nanofiber neural network 200 using CMOS neurons 327A-327B is created. Thereafter, the process ends.

[0020] Disjunctive language used herein, such as the phrase "at least one of X, Y, or Z," unless specifically stated otherwise, is otherwise understood with the context as used in general to present that an item, term, etc., may be either X, Y, or Z, or any combination thereof (e.g., X, Y, and/or Z). Thus, such disjunctive language does not imply that certain embodiments require at least one of X, at least one of Y, or at least one of Z to each be present.

[0021] The above-described embodiments of the present disclosure are merely examples of implementations set forth for a clear understanding of the principles of the disclosure. Many variations and modifications may be made to the above- described embodiments without departing substantially from the spirit and principles of the disclosure. All such modifications and variations are intended to be included herein within the scope of this disclosure.