Quantized inputs for machine learning models

1. A method of pre-processing first data for use with a trained machine learning model, the method comprising:

accessing the first data, wherein the first data has a first precision represented by a floating-point value of at least 8 bits;

generating, based on at least a first portion of the first data, second data having a second precision, represented by an integer value, lower than the first precision; and

providing the second data as input to the trained machine learning model to generate model output, wherein:

the trained machine learning model is trained using training data having a greater precision than the second precision;

the trained machine learning model is implemented by a photonic processor; and

providing the second data as input to the trained machine learning model comprises sending the second data from a host device to a second device having the photonic processor; and

using the photonic processor to provide the model output based at least in part on the second data.

2. The method of claim 1 , wherein the first precision comprises a first number of bits of at least 8 bits, and the second precision comprises a second number of bits, wherein the second number of bits is lower than the first number of bits.

3. The method of claim 1 , wherein the trained machine learning model was trained using training data having a same precision as the first precision.

4. The method of claim 1 , wherein the at least the first portion of the first data comprises all of the first data.

5. The method of claim 1 , further comprising scaling and/or normalizing the second data.

6. The method of claim 1 , wherein generating the second data is performed as part of an analog to digital conversion of the at least the first portion of the first data.

7. The method of claim 2 , wherein the second number of bits is less than 65% of the first number of bits, and wherein an accuracy of the trained machine learning model using the second data is at least 95% of an accuracy of the trained machine learning model using data having a same precision as the first precision.

8. The method of claim 1 , wherein generating the second data comprises performing a bit shifting operation on the at least the first portion of the first data.

9. The method of claim 1 , further comprising:

generating, based on a second portion of the first data, third data having a third precision lower than the first precision, wherein the third precision is different than the second precision.

10. The method of claim 9 , wherein the first data comprises image data, wherein generating the second data comprises reducing a precision of a value for a first pixel of the image data, and wherein generating the third data comprises reducing a precision of a value for a second pixel of the image data.

11. The method of claim 9 , wherein the first data comprises image data, and wherein generating the second data comprises reducing a precision of a value for a first channel of the image data, and wherein generating the third data comprises reducing a precision of a value for a second channel of the image data.

12. The method of claim 1 , wherein the first data comprises image data.

13. The method of claim 1 , wherein the first data comprises time series data.

14. The method of claim 5 , wherein:

accessing the first data comprises reading a value from an image file; and

generating the second data comprises right-shifting, by a number of bits, the read value to produce a reduced-bit value.

15. The method of claim 14 , wherein scaling and/or normalizing the first portion of the first data comprises:

scaling the reduced-bit value to a within a range of 0 to 1; and

normalizing the scaled reduced-bit value according to a dataset mean and standard deviation.

16. A method for using a trained machine learning model with reduced-precision input, the method comprising:

accessing reduced-precision data having a precision, represented by an integer value, less than a precision of training data used to train the trained machine learning model, wherein the precision of training data is a floating-point value of at least 8 bits;

providing the reduced-precision data as input to the trained machine learning model to generate model output, wherein the trained machine learning model is implemented by a photonic processor and providing the reduced-precision data as input to the trained machine learning model comprises sending the reduced-precision data from a host device to a second device having the photonic processor; and

using the model output to determine at least one characteristic of the reduced-precision data.

17. The method of claim 16 , further comprising scaling and/or normalizing the reduced-precision data.

18. A method of pre-processing first image data for use with a trained machine learning model, the method comprising:

accessing the first image data, wherein the first image data has a first precision represented by a floating-point value of at least 8 bits;

generating, based on at least a first portion of the first image data, second image data having a second precision, represented by an integer value, lower than the first precision; and

providing the second image data as input to the trained machine learning model to generate model output, wherein:

the trained machine learning model is trained using training data having a greater precision than the second precision;

the trained machine learning model is implemented by a photonic processor; and

providing the second image data as input to the trained machine learning model comprises sending the second image data from a host device to a second device having the photonic processor; and

using the photonic processor to provide the model output based on the second image data.