Kernel sets normalization with capacitor charge sharing

1. A device for multiple copies of a set of multi-kernel set operations in a hardware accelerated neural network, the device including:

a word line for receiving a pixel value of an input image;

a bit line for communicating a modified pixel value;

an analog memory cell including a first capacitor for storing a first kernel weight of a first kernel in one of a plurality of kernel sets such that the pixel value is operated on by the first kernel weight to produce the modified pixel value;

a charge connection connecting the first capacitor to at least a second capacitor storing a second kernel weight of a related kernel of a second one of the plurality of kernel sets such that charge is shared between the first capacitor and at least the second capacitor to normalize the first kernel weight and the second kernel weight; and

a first cross-bar array having a first set of analog memory cells, including a first analog memory cell, for performing convolutions on a first patch of the input image; and

a second cross-bar array having a second set of analog memory cells, including a second analog memory cell corresponding to the second capacitor, for performing convolutions on a second patch of the input image concurrently with the convolutions on the first patch of the input image.

2. The device as recited in claim 1 , wherein the charge connection includes a pass-gate.

3. The device as recited in claim 1 , wherein the charge connection connects the first capacitor directly to at least the second capacitor.

4. The device as recited in claim 1 , further including a charge field effect transistor (FET) connecting the first capacitor to a charge voltage.

5. The device as recited in claim 1 , further including a discharge FET connecting the first capacitor to a ground.

6. The device as recited in claim 1 , further including a readout FET connecting the word line and the bit line to the first capacitor for a continuous modulation of a signal through the readout FET.

7. The device as recited in claim 1 , wherein the second capacitor is connected to a second word line and a second bit line in the second cross-bar array.

8. The device as recited in claim 7 , wherein:

the charge of the first capacitor is updated by a first update signal received by the bit line of the first analog memory cell;

the charge of the second capacitor is updated by a first update signal received by the second bit line; and

the charge of the first capacitor and the charge of the second capacitor are normalized by the charge connection to average kernel weights corresponding to each of the charge of the first capacitor and the charge of the second capacitor.

9. A device for multiple copies of a set of multi-kernel set operations in a hardware accelerated neural network, the device including:

a first cross-bar array for accelerating convolutions of at least one first patch of an input image, the first cross-bar array including:

a plurality of first word lines for receiving pixel values of the input image;

a plurality of first bit lines for communicating modified pixel values; and

a plurality of first analog memory cells, each analog memory cell including a first capacitor for storing a corresponding kernel weight of a first kernel in one of a plurality of kernel sets such that each pixel value is operated on by the first kernel weight to produce the modified pixel value;

at least a second cross-bar array separate from the first cross-bar array for accelerating convolutions of at least one second patch of the input image concurrently with the convolutions on the first patch of the input image; and

a plurality of charge connections connecting respective first capacitors to at least a respective second capacitor of at least the second cross-bar array, the respective second capacitor storing a second kernel weight of a related kernel of a second one of the plurality of kernel sets such that charge is shared between a respective first capacitor and at least the respective second capacitor to normalize the first kernel weight and the second kernel weight.

10. The device as recited in claim 9 , wherein the charge connection includes a pass-gate.

11. The device as recited in claim 9 , wherein the charge connection connects the first capacitor directly to at least the second capacitor.

12. The device as recited in claim 9 , further including a charge field effect transistor (FET) connecting the first capacitor to a charge voltage.

13. The device as recited in claim 9 , further including a discharge FET connecting the first capacitor to a ground.

14. The device as recited in claim 9 , further including a readout FET connecting a word line and a bit line of the plurality of word lines and bit lines, respectively, to the first capacitor for a continuous modulation of a signal through the readout FET.

15. The device as recited in claim 9 , wherein each of the first cross-bar array and at least the second cross-bar array includes a plurality analog memory cells to accelerate convolutions of the at least one first patch and the at least one second patch, respectively, by providing values corresponding to each pixel to corresponding ones of the plurality of analog memory cells through a corresponding plurality of first word lines.

16. The device as recited in claim 15 , wherein:

the charge of the first capacitor is updated by a first update signal received by the bit line of the first analog memory cell;

the charge of the second capacitor is updated by a first update signal received by a second bit line in the second cross-bar array; and

the charge of the first capacitor and the charge of the second capacitor are normalized by the charge connection to average kernel weights corresponding to each of the charge of the first capacitor and the charge of the second capacitor.

17. A method for multi-kernel set operations in a hardware accelerated neural network, the method including:

receiving a pixel value of an input image with a word line;

producing a modified pixel value from the pixel value with a first analog memory cell including a first capacitor for storing a first kernel weight of a first kernel in one of a plurality of kernel sets such that the pixel value is operated on by the kernel weight;

communicating the modified pixel value through a bit line;

sharing a charge between the first capacitor and at least a second capacitor storing a second kernel weight of a related kernel of a second one of the plurality of kernel sets using a charge connection connecting the first capacitor to at least the second capacitor to normalize the first kernel weight and at least the second kernel weight;

performing convolutions on a first patch of the input image by a first set of analog memory cells of a first cross-bar array including the first analog memory cell; and

performing convolutions on a second patch of the input image concurrently with the convolutions on the first patch of the input image, by a second set of analog memory cells of a second cross-bar array including the second capacitor in a second analog memory cell.

18. The method as recited in claim 17 , further including modulating the sharing of the charge between the first capacitor and at least the second capacitor with the charge connection by using a pass-gate.