阅读说明：本技术 二分忆阻网络中的深度学习 (Deep learning in a two-memristive network ) 是由 杰克·D·肯德尔 胡安·C·尼诺 劳拉·E·苏亚雷斯 于 2018-05-21 设计创作，主要内容包括：本文描述了一种二分忆阻网络和训练这种网络的方法。在一个示例情况下,忆阻网络可以包括多个纳米纤维,其中,每个纳米纤维包括金属核和忆阻壳。忆阻网络还可以包括被放置在纳米纤维上的多个电极。第一组多个电极可以包括在忆阻网络中的输入电极,以及第二组多个电极可以包括在忆阻网络中的输出电极。忆阻网络可以体现为二分忆阻网络,并根据本文描述的训练方法进行训练。(A two-memristive network and a method of training such a network are described herein. In one example case, the memristive network may include a plurality of nanofibers, where each nanofiber includes a metal core and a memristive shell. The memristive network may also include a plurality of electrodes disposed on the nanofibers. The first plurality of electrodes may include input electrodes in a memristive network, and the second plurality of electrodes may include output electrodes in the memristive network. The memristive network may be embodied as a two-memristive network and trained according to the training methods described herein.)

1. A method of training a memristive network including a plurality of input nodes and a plurality of output nodes, the method comprising:

applying an input voltage or an input current to an input node of the plurality of input nodes;

grounding an output node of the plurality of output nodes;

measuring an output current or an output voltage at the output node;

comparing the output current or the output voltage to a target current or a target voltage to determine an error δ; and

applying a threshold voltage or a threshold current to the output node for a period of time proportional to a magnitude of the error δ.

2. The method of claim 1, wherein applying the threshold voltage or the threshold current to the output node when the error δ is negative comprises:

applying a positive threshold voltage or a positive threshold current to the output node for the period of time proportional to the error δ; and

applying a negative threshold voltage or a negative threshold current to the output node for the period of time proportional to the error δ.

3. The method of claim 1, wherein applying the threshold voltage or the threshold current to the output node when the error δ is positive comprises:

inverting a polarity of the input voltage or the input current applied to the input node;

applying a positive threshold voltage or a positive threshold current to the output node for the period of time proportional to the error δ; and

applying a negative threshold voltage or a negative threshold current to the output node for the period of time proportional to the error δ.

4. The method of claim 1, further comprising:

transforming the error δ into an error δ voltage or an error δ current;

applying the error delta voltage or the error delta current to the output node; and

applying the threshold voltage or the threshold current to the input node for a second time period proportional to an absolute value of the error delta voltage or the error delta current.

5. The method of claim 4, wherein applying the threshold voltage or the threshold current to the input node when the input voltage or the input current applied to the input node is positive comprises:

applying a positive threshold voltage or a positive threshold current to the input node for the second time period proportional to an absolute value of the error delta voltage or the error delta current; and

applying a negative threshold voltage or a negative threshold current to the input node for the second time period proportional to an absolute value of the error delta voltage or the error delta current.

6. The method of claim 4, wherein applying the threshold voltage or the threshold current to the input node when the input voltage or the input current applied to the input node is negative comprises:

inverting a polarity of the error delta voltage or the error delta current applied to the output node;

applying a positive threshold voltage or a positive threshold current to the input node for the second time period proportional to an absolute value of the error delta voltage or the error delta current; and

applying a negative threshold voltage or a negative threshold current to the input node for the second time period proportional to the absolute value of the error delta voltage.

7. The method of claim 1, in which the memristive network comprises a two-memristive network.

8. The method of claim 1, wherein the method reproduces a back propagation algorithm used to train the memristive network.

9. A memristive network, comprising:

a plurality of nanofibers, wherein each nanofiber comprises a metal core and a memristive shell;

a plurality of electrodes disposed on the nanofibers, wherein the plurality of electrodes comprises a plurality of input nodes and a plurality of output nodes; and

a training processor configured to:

applying an input voltage or an input current to an input node of the plurality of input nodes;

grounding an output node of the plurality of output nodes;

measuring an output current or an output voltage at the output node;

comparing the output current or the output voltage to a target current or a target voltage to determine an error δ; and

applying a threshold voltage or a threshold current to the output node for a period of time proportional to a magnitude of the error δ.

10. The memristive network of claim 9, wherein when the error δ is negative, the training processor is further configured to:

applying a positive threshold voltage or a positive threshold current to the output node for the period of time proportional to the error δ; and

applying a negative threshold voltage or a negative threshold current to the output node for the period of time proportional to the error δ.

11. The memristive network of claim 9, wherein when the error δ is positive, the training processor is further configured to:

inverting a polarity of the input voltage or the input current applied to the input node;

applying a positive threshold voltage or a positive threshold current to the output node for the period of time proportional to the error δ; and

applying a negative threshold voltage or a negative threshold current to the output node for the period of time proportional to the error δ.

12. The memristive network of claim 9, wherein the training processor is further configured to:

transforming the error δ into an error δ voltage or an error δ current;

applying the error delta voltage or the error delta current to the output node; and

applying the threshold voltage or the threshold current to the input node for a second time period proportional to an absolute value of the error delta voltage or the error delta current.

13. The memristive network of claim 12, wherein when the input voltage or the input current applied to the input node is positive, the training processor is further configured to:

applying a positive threshold voltage or a positive threshold current to the input node for the second time period proportional to an absolute value of the error delta voltage or the error delta current; and

applying a negative threshold voltage or a negative threshold current to the input node for the second time period proportional to an absolute value of the error delta voltage or the error delta current.

14. The memristive network of claim 12, wherein when the input voltage or the input current applied to the input node is negative, the training processor is further configured to:

inverting a polarity of the error delta voltage or the error delta current applied to the output node;

applying a positive threshold voltage or a positive threshold current to the input node for the second time period proportional to an absolute value of the error delta voltage or the error delta current; and

applying a negative threshold voltage or a negative threshold current to the input node for the second time period proportional to the absolute value of the error delta voltage.

15. The memristive network of claim 9, wherein the memristive network comprises a two-memristive network.