Apparatuses and methods for operating neural networks
The present disclosure includes apparatuses and methods for operating neural networks. An example apparatus includes a plurality of neural networks, wherein the plurality of neural networks are configured to receive a particular portion of data and wherein each of the plurality of neural networks are configured to operate on the particular portion of data during a particular time period to make a determination regarding a characteristic of the particular portion of data.
1. An apparatus, comprising:
a plurality of neural networks, wherein each neural network of the plurality of neural networks operates within a corresponding bank of a plurality of banks of a processing in memory (PIM) device implemented on a chip, and wherein each neural network is configured to:
receive different training;
determine, via a plurality of compute components included in the corresponding bank, a characteristic of a particular portion of data simultaneously in response to receiving the different training; and
determine, via the plurality of compute components included in the corresponding bank, a training confidence factor of the characteristic of the particular portion of data based on the different training of each of the plurality of neural networks; and
a controller coupled to the plurality of neural networks, wherein the controller is configured to:
determine an accuracy of data recognition based on the training confidence factor of each of the plurality of neural networks using a voting scheme; and
determine the characteristic of the particular portion of data based on the accuracy of the data recognition.
2. The apparatus of claim 1 , wherein the plurality of neural networks can include an array of memory cells coupled to sensing circuitry, wherein the sensing circuitry includes a sense amplifier and a compute component.
3. The apparatus of claim 2 , wherein the array of memory cells is coupled to the sensing circuitry via a plurality of input/output (I/O) lines shared as a data path for in data path compute operations associated with the array configured to make the determination regarding the characteristic of the particular portion of data.
4. The apparatus of claim 1 , wherein the plurality of neural networks are configured to be independently trained.
5. The apparatus of claim 1 , wherein the plurality of neural networks are configured to simultaneously receive instructions to operate on the particular portion of data.
6. The apparatus of claim 1 , wherein the particular portion of data is configured to be an image, portion of an image, sound, or emotion.
7. The apparatus of claim 1 , wherein the plurality of neural networks are configured to operate in a fixed point or binary weighted network.
8. An apparatus, comprising:
a plurality of neural networks, wherein each neural network of the plurality of neural networks operates within a corresponding bank of a plurality of banks of a processing in memory (PIM) device implemented on a chip, and wherein each neural network is configured to:
receive different training;
receive a particular portion of data;
determine, via a plurality of compute components included in the corresponding bank, a characteristic of the particular portion of data simultaneously in response to receiving the different training; and
determine, via the plurality of compute components included in the corresponding bank, a training confidence factor of the characteristic of the particular portion of data based on the different training of each of the plurality of neural networks; and
a controller, wherein the controller is coupled to the plurality of neural networks and is configured to:
determine an accuracy of data recognition based on the training confidence factor of each of the plurality of neural networks using a voting scheme; and
determine the characteristic of the particular portion of data based on the accuracy of the data recognition.
9. The apparatus of claim 8 , wherein the controller is configured to receive a vote from each of the plurality of neural networks.
10. The apparatus of claim 9 , wherein the vote from each of the plurality of neural networks is evaluated by the controller.
11. The apparatus of claim 10 , wherein the vote from each of the plurality of neural networks is evaluated based on type of particular portion of data and particular training of each of the plurality of neural networks.
12. The apparatus of claim 8 , wherein the voting scheme is a majority rule.
13. The apparatus of claim 8 , wherein the voting scheme is an average where a result of each neural network is averaged.
14. The apparatus of claim 8 , wherein an output is provided by the controller based on the accuracy of the data recognition.
15. The apparatus of claim 14 , wherein the output is discarded if there is no uniform decision on the accuracy of the data recognition among the plurality of neural networks.
16. The apparatus of claim 8 , wherein the plurality of neural networks are configured as single-bit networks.
17. A method comprising:
training each of a plurality of neural networks differently, wherein each neural network of the plurality of neural networks operates within a corresponding bank of a plurality of banks of a processing in memory (PIM) device implemented on a chip;
determining, via the plurality of compute components included in the corresponding bank, a characteristic of a particular portion of data at each of the plurality of neural networks simultaneously in response to training each of the plurality of neural networks differently;
determining, via the plurality of compute components in the corresponding bank, a training confidence factor of the characteristic of the particular portion of data based on the different training of each of the plurality of neural networks;
determining an accuracy of data recognition based on the training confidence factor of each of the plurality of neural networks using a voting scheme at a controller; and
determining the characteristic of the particular portion of data based on the accuracy of the data recognition at the controller.
18. The method of claim 17 , wherein the method includes training each of the plurality of neural networks independently.
19. A method comprising:
training each of a plurality of neural networks differently, wherein each neural network of the plurality of neural networks operates within a corresponding bank of a plurality of banks of a processing in memory (PIM) device implemented on a chip;
operating the plurality of neural networks simultaneously in response to training each of the plurality of neural networks differently, wherein each of the plurality of neural networks receives a particular portion of data, determines, via a plurality of compute components included in the corresponding bank, a characteristic of the particular portion of data, and determines, via the plurality of compute components included in the corresponding bank, a training confidence factor of the characteristic of the particular portion of data based on the different training of each of the plurality of neural networks; and
operating a controller, wherein the controller determines an accuracy of data recognition based on the training confidence factor of each of the plurality of neural networks using a voting scheme, and wherein the controller determines the characteristic of the particular portion of data based on the accuracy of the data recognition.
20. The method of claim 19 , wherein a vote is received from each of the plurality of neural networks.
21. The method of claim 20 , wherein the vote received from each of the plurality of neural networks is evaluated.
22. The method of claim 21 , wherein the vote received from each of the plurality of neural networks is evaluated based on type of the particular portion of data and the different training of each of the plurality of neural networks.
23. The method of 19 , wherein a majority rule is used as the voting scheme.
24. The method of 19 , wherein an average is used as the voting scheme where a result of each of the plurality of neural networks is averaged.
25. The method of claim 19 , wherein an output is provided based on the accuracy of the data recognition.
26. The method of claim 25 , wherein no output is provided if there is no uniform decision on the accuracy of the data recognition among the plurality of neural networks.