Techniques for pruning neural networks

1 . One or more processors to perform at least one of object detection or language model inferencing based, at least in part, on:

updating a first neural network model in a data storage device communicatively coupled with the one or more processors, the updating comprising at least:

removing one or more neurons of the first neural network model in the data storage device communicatively coupled with the one or more processors; and

adjusting one or more subsequent layers of the first neural network model to compensate for the removed one or more neurons based, at least in part, on one or more operations to be performed in at least one layer of the one or more subsequent layers; and

using the updated neural network model to perform the at least one of object detection or language model inferencing using one or more data inputs.

2 . The one or more processors of claim 1 , wherein adjusting the one or more subsequent layers of the first neural network model includes propagating one or more bias terms of the removed one or more neurons to a subsequent layer of the one or more subsequent layers.

3 . The one or more processors of claim 2 , wherein propagating the one or more bias terms includes calculating one or more activation function results of the one or more bias terms.

4 . The one or more processors of claim 1 , wherein removing the one or more neurons includes identifying zero-weighted neurons in the first neural network model, and removing the zero-weighted neurons.

5 . The one or more processors of claim 1 , wherein adjusting the one or more subsequent layers of the first neural network model includes determining that removing the one or more neurons would result in an element-wise operation performed between layers having operands of different length, and inserting one or more scatter operations to generate operands of equal length for the element-wise operation.

6 . The one or more processors of claim 1 , wherein the one or more processors are also to:

import the first neural network model; and

export the updated neural network model.

7 . The one or more processors of claim 1 , wherein adjusting the one or more subsequent layers of the first neural network model includes:

identifying a first type of a first layer of the first neural network model;

identifying a second type of a second layer of the first neural network model, wherein the second layer is subsequent to the first layer; and

adjusting the second layer based, at least in part, on the identified first type and second type.

8 . The one or more processors of claim 7 , wherein the first type is fully-connected.

9 . A non-transitory machine-readable medium having stored thereon a set of instructions, which if performed by one or more processors, cause the one or more processors to perform at least one of object detection or language model inferencing based, at least in part, on:

updating a first neural network model in a data storage device communicatively coupled with the one or more processors by at least:

removing one or more neurons of the first neural network model in the data storage device communicatively coupled with the one or more processors; and

adjusting one or more subsequent layers of the first neural network model to compensate for the removed one or more neurons based, at least in part, on one or more operations to be performed in at least one layer of the one or more subsequent layers; and

using the updated neural network model to perform the at least one of object detection or language model inferencing using one or more data inputs.

10 . The non-transitory machine-readable medium of claim 9 , wherein adjusting the one or more subsequent layers of the first neural network model includes propagating one or more bias terms of the removed one or more neurons to a subsequent layer of the one or more subsequent layers.

11 . The non-transitory machine-readable medium of claim 10 , wherein propagating the one or more bias terms includes calculating one or more activation functions of the one or more bias terms.

12 . The non-transitory machine-readable medium of claim 9 , wherein removing the one or more neurons includes identifying zero-weighted neurons in the first neural network model, and removing the zero-weighted neurons.

13 . The non-transitory machine-readable medium of claim 9 , wherein adjusting the one or more subsequent layers of the first neural network model includes determining that removing the one or more neurons would result in an element-wise operation performed between layers having operands of different length, and inserting one or more scatter operations to generate operands of equal length for the element-wise operation.

14 . The non-transitory machine-readable medium of claim 9 , wherein the instructions, which if performed by the one or more processors, further cause the one or more processors to:

import the first neural network model; and

export the updated neural network model.

15 . The non-transitory machine-readable medium of claim 14 , wherein the first neural network model is a first neural network model stored according to Open Neural Network Exchange format and the updated neural network model is a second neural network model stored according to Open Neural Network Exchange format.

16 . A method to perform at least one of object detection or language model inferencing, comprising:

updating a first neural network model in a data storage device communicatively coupled with one or more processors by at least:

removing one or more neurons of the first neural network model and adjusting one or more subsequent layers of the first neural network model to compensate for the removed one or more neurons based, at least in part, on one or more operations to be performed in at least one layer of the one or more subsequent layers; and

using the updated neural network model to perform the at least one of object detection or language model inferencing using one or more data inputs.

17 . The method of claim 16 , wherein adjusting the one or more subsequent layers of the first neural network model includes propagating one or more bias terms of the removed one or more neurons to a subsequent layer of the one or more subsequent layers.

18 . The method of claim 17 , wherein propagating the one or more bias terms includes calculating one or more activation function results of the one or more bias terms and calculating one or more bias terms of the subsequent layer based, at least in part, on the calculated one or more activation function results.

19 . The method of claim 16 , wherein removing the one or more neurons includes identifying zero-weighted neurons in the first neural network model, and removing the zero-weighted neurons.

20 . The method of claim 16 , wherein adjusting the one or more subsequent layers of the first neural network model includes determining that removing the one or more neurons would result in an element-wise operation performed between layers having operands of different length, and inserting one or more scatter operations to generate operands of equal length for the element-wise operation.

21 . The method of claim 16 , further comprising:

importing the first neural network model; and

exporting the updated neural network model.

22 . A system, comprising:

one or more processors to perform at least one of object detection or language model inferencing based, at least in part, on:

updating a first neural network model in a data storage device communicatively coupled with the one or more processors by at least:

removing one or more neurons of the first neural network model in the data storage device communicatively coupled with the one or more processors; and

adjusting one or more subsequent layers of the first neural network model to compensate for the removed one or more neurons based, at least in part, on one or more operations to be performed in at least one layer of the one or more subsequent layers; and

using the updated neural network model to perform the at least one of object detection or language models inferencing using one or more data inputs; and

one or more memories to store the updated neural network model.

23 . The system of claim 22 , wherein adjusting the one or more subsequent layers of the first neural network model includes propagating one or more bias terms of the removed one or more neurons to a subsequent layer of the one or more subsequent layers.

24 . The system of claim 23 , wherein propagating the one or more bias terms includes calculating one or more activation function results of the one or more bias terms.

25 . The system of claim 22 , wherein removing the one or more neurons includes identifying zero-weighted neurons in the first neural network model, and removing the zero-weighted neurons.

26 . The system of claim 22 , wherein adjusting the one or more subsequent layers of the first neural network model includes determining that removing the one or more neurons would result in an element-wise operation performed between layers having operands of different length, and inserting one or more scatter operations to generate operands of equal length for the element-wise operation.

27 . The system of claim 22 , wherein the one or more processors are further to:

import the first neural network model; and

export the updated neural network model.