METHOD FOR IMPROVING CONVOLUTIONAL NEURAL NETWORK TO PERFORM COMPUTATIONS

Abstract

A method for improving a convolutional neural network (CNN) to perform computations is provided. The method includes the following steps: determining a number of a plurality of multipliers to be N and a number of a plurality of adders to be N according to a number of convolution kernels used by a plurality of convolution layers; and in response to an i-th convolutional layer of the convolutional neural network performing a convolution operation and N convolution kernels of the i-th convolutional layer being all in a size of K×1×1, using the N multipliers and the N adders to perform a multiplication operation once and an addition operation once for each of the N convolution kernels of the i-th convolutional layer in one cycle, such that N outputs of the N convolution kernels of the i-th convolutional layer are obtained after K cycles.

Claims

1. A method for improving a convolutional neural network to perform computations, the convolutional neural network including a plurality of convolutional layers, each of the plurality of convolutional layers using N convolution kernels, and the method comprising: determining a number of a plurality of multipliers to be N and a number of a plurality of adders to be N according to the N convolution kernels used by the plurality of convolution layers; and in response to an i-th convolutional layer of the convolutional neural network performing a convolution operation and the N convolution kernels of the i-th convolutional layer all having a size of K×1×1, using the N multipliers and the N adders to perform a multiplication operation once and an addition operation once for each of the N convolution kernels of the i-th convolutional layer in one cycle, such that N outputs of the N convolution kernels of the i-th convolutional layer are obtained after K cycles, wherein N is an integer greater than 1, i is an integer greater than or equal to 1, and K is an integer greater than 1.

2. The method according to claim 1, further comprising: in response to a j-th convolutional layer of the convolutional neural network performing the convolution operation and the N convolution kernels of the j-th convolutional layer all having a size of P×1×N, using the N multipliers and the N adders to perform N multiplication operations and N addition operations for a target convolution kernel of the N convolution kernels of the j-th convolutional layer in one cycle, such that an output of the target convolution kernel is obtained after P cycles, wherein j is an integer greater than or equal to 1, and P is an integer greater than 1.

3. The method according to claim 2, wherein the convolutional neural network further includes a plurality of fully connected layers, and the method further comprises: in response to a k-th fully connected layer of the convolutional neural network performing an operation and a total number of records of input data of the k-th fully connected layer being M*N, using the N multipliers and the N adders to complete conversion operations of N records of the input data in one cycle, such that an output of the k-th fully connected layer is obtained after M cycles, wherein k and M are integers greater than or equal to 1.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The described embodiments may be better understood by reference to the following description and the accompanying drawings, in which:

[0010] FIG. 1 is a flow chart of a method for improving a convolutional neural network (CNN) to enable usage of different sized convolution kernels in different convolution layers according to one embodiment of the present disclosure;

[0011] FIG. 2 is a schematic diagram of the CNN processing MFCC according to one embodiment of the present disclosure; and

[0012] FIGS. 3A to 3C are schematic diagrams of the method in FIG. 1 being applied to a first convolutional layer in FIG. 2.

DETAILED DESCRIPTION OF THE EXEMPLARY EMBODIMENTS

[0013] The present disclosure is more particularly described in the following examples that are intended as illustrative only since numerous modifications and variations therein will be apparent to those skilled in the art. Like numbers in the drawings indicate like components throughout the views. As used in the description herein and throughout the claims that follow, unless the context clearly dictates otherwise, the meaning of “a”, “an”, and “the” includes plural reference, and the meaning of “in” includes “in” and “on”. Titles or subtitles can be used herein for the convenience of a reader, which shall have no influence on the scope of the present disclosure.

[0014] The terms used herein generally have their ordinary meanings in the art. In the case of conflict, the present document, including any definitions given herein, will prevail. The same thing can be expressed in more than one way. Alternative language and synonyms can be used for any term(s) discussed herein, and no special significance is to be placed upon whether a term is elaborated or discussed herein. A recital of one or more synonyms does not exclude the use of other synonyms. The use of examples anywhere in this specification including examples of any terms is illustrative only, and in no way limits the scope and meaning of the present disclosure or of any exemplified term. Likewise, the present disclosure is not limited to various embodiments given herein. Numbering terms such as “first”, “second” or “third” can be used to describe various components, signals or the like, which are for distinguishing one component/signal from another one only, and are not intended to, nor should be construed to impose any substantive limitations on the components, signals or the like.

[0015] Referring to FIG. 1 and FIG. 2, FIG. 1 is a flow chart of a method for improving a convolutional neural network (CNN) to enable usage of different size convolution kernels in different convolution layers according to one embodiment of the present disclosure, and FIG. 2 is a schematic diagram showing the CNN processing MFCC according to one embodiment of the present disclosure. As mentioned above, the CNN includes a plurality of convolutional layers, and each convolutional layer can use N convolution kernels, where N is an integer greater than 1. For the convenience of the following description, the CNN of FIG. 2 that processes the MFCC is exemplified to include four convolutional layers in the present embodiment, and each convolutional layer uses 16 convolution kernels. However, the present disclosure does not limit input data of the CNN to be the MFCC, nor does the present disclosure limit a number of the convolutional layers included in the CNN and a number of the convolution kernels used in these convolutional layers. In general, the MFCC can be a parameter matrix with a size of 1×13×1, and 99 parameter matrices will be input to the CNN. That is, the input data in FIG. 2 can be a matrix with a size of 99×13×1, but the present disclosure is not limited thereto.

[0016] Since the convolution kernels of a first convolutional layer are all in a size of 10×1×1, the first convolutional layer of the conventional CNN completes one convolution operation by performing multiplication operations for 10 times and addition operations for 9 times on 10 elements of the input data and one of convolution kernels of the first convolutional layer, so as to obtain an output. In addition, since the convolution kernels of a second convolutional layer are all in a size of 10×1×16, the second convolutional layer of the existing CNN completes one convolution operation by performing the multiplication operations for 160 times and the addition operations for 159 times on 10*16 elements of the input data and one of the convolution kernels of the second convolutional layer, so as to obtain an output. Similarly, since the convolution kernels of a third and a fourth convolutional layer are all in a size of 6×1×16, in the existing CNN, the 3rd convolutional layer completes one convolution operation by performing the multiplication operations for 96 times and the addition operations for 95 times on 6*16 elements and one of the convolution kernels of the 3rd convolutional layer, and the 4th convolutional layer completes one convolution operation by performing the multiplication operations for 96 times and the addition operations for 95 times on 6*16 elements and one of the convolution kernels of the 4th convolutional layer, so as to obtain an output.

[0017] It can be observed that the existing CNN requires 10 multipliers for the first convolutional layer, 160 multipliers for the second convolutional layer, and 96 multipliers for each of the third and the fourth convolutional layer. Therefore, integration of circuits is difficult to achieve. In addition, each convolutional layer needs independent control and access circuits. Especially for data storage, a number of the elements that need to be read for each operation is different, such that storage control and intermediate buffering mechanisms are complicated. In response to the above-referenced technical inadequacies, in step S110 of FIG. 1, according to the number of convolution kernels used by the convolution layers, a number of the multipliers and a number of adders are determined to be N in this embodiment of the present disclosure. Next, in step S120, when an i-th convolutional layer of the convolutional neural network performs the convolution operation and the N convolution kernels of the i-th convolutional layer are all in a size of K×1×1, the N multipliers and the N adders are used to perform the multiplication operation once and the addition operation once for the N convolution kernels of the i-th convolutional layer in one cycle in the embodiment of the present disclosure, such that N outputs of the N convolution kernels of the i-th convolutional layer are obtained after K cycles. Here, i is an integer greater than or equal to 1, and K is an integer greater than 1.

[0018] In other words, i, N, and K can respectively be 1, 16 and 10 in this embodiment, but the present disclosure is not limited thereto. Therefore, the first convolutional layer of FIG. 2 of the present disclosure performs the multiplication operation once and the addition operation once on one element and each convolution kernel of the first convolution layer. Reference can be made to FIGS. 3A to 3C, which are schematic diagrams showing that the method in FIG. 1 is applied to a first convolutional layer in FIG. 2. As shown in FIG. 3A, in a first cycle of the present disclosure, an element A.sub.1,1 of the input data is multiplied with an element B.sub.1,1 of a first convolution kernel CK.sub.1,1 of the first convolution layer, and is then added with an operation result of a previous stage to obtain an operation result C.sub.1,1 of this stage. The element A.sub.1,1 of the input data is also multiplied with an element B.sub.2,1 of a second convolution kernel CK.sub.1,2 of the first convolution layer, and is then added with the operation result of the previous stage to get an operation result C.sub.2,1 of this stage, and so forth. In this embodiment of the present disclosure, the element A.sub.1,1 is multiplied with an element B16,1 of a sixteenth convolution kernel CK.sub.1,16 of the first convolution layer, and is then added with the operation result of the previous stage to obtain an operation result C16,1 of this stage. Since there is no operation result of the previous stage at this time, in this embodiment of the present disclosure, the element A.sub.1,1 and the elements B.sub.1,1 to B16,1 are multiplied correspondingly, and then added correspondingly with 0 to obtain the operation results C.sub.1,1 to C16,1.

[0019] Similarly, as shown in FIG. 3B, in a second cycle of the present disclosure, an element A.sub.1,2 of the input data is multiplied with an element B.sub.1,2 of the first convolution kernel CK.sub.1,1 of the first convolution layer, and is then added with the operation result C.sub.1,1 of a previous stage to get an operation result C.sub.1,2 of this stage. The element A.sub.1,2 of the input data is also multiplied with an element B.sub.2,2 of the second convolution kernel CK.sub.1,2 of the first convolution layer, and is then added with the operation result C.sub.2,1 of the previous stage to get an operation result C.sub.2,2 of this stage, and so forth. In this embodiment of the present disclosure, the element A.sub.1,2 is also multiplied with an element B16,2 of the sixteenth convolution kernel CK.sub.1,16 of the first convolution layer, and is then added with the operation result C16,1 of the previous stage to obtain an operation result C16,1 of this stage. Therefore, as shown in FIG. 3C, in a tenth cycle of the present disclosure, an operation result Cr,10 can be obtained, which is equal to:

A.sub.1,1*B.sub.r,1+A.sub.1,2*B.sub.r,2+A.sub.1,3*B.sub.r,3+A.sub.1,4*B.sub.r,4+A.sub.1,5*B.sub.r,5+A.sub.1,6*B.sub.r,6+A.sub.1,7*B.sub.r,7+A.sub.1,8*B.sub.r,8+A.sub.1,9*B.sub.r,9+A.sub.1,10*B.sub.r,10;

where r is an integer from 1 to 16. That is, outputs of 16 convolution kernels can be obtained at the same time.

[0020] Taking FIG. 2 as an example, it can be observed that 16 multipliers and 16 adders are utilized in the present disclosure to perform the multiplication operation once and the addition operation once for each of the 16 convolution kernels of the first convolution layer in one cycle, such that the 16 outputs of the 16 convolution kernels of the first convolution layer are obtained after 10 cycles. Therefore, storage control is no longer complicated, and only one independent region for storage is needed for intermediate buffering. In addition, in step S130 of FIG. 1 of the present disclosure, when a j-th convolutional layer of the convolutional neural network performs a convolution operation and the N convolution kernels of the j-th convolutional layer are all in a size of P×1×N, the N multipliers and the N adders are used to perform N multiplication operations and N addition operations for a target convolution kernel of the N convolution kernels of the j-th convolutional layer in one cycle, such that an output of the target convolution kernel is obtained after P cycles. Here, j is an integer greater than or equal to 1, and P is an integer greater than 1.

[0021] In other words, j and P can respectively be 2 and 10 in this embodiment, but the present disclosure is not limited thereto. Therefore, the second convolutional layer of FIG. 2 of the present disclosure performs the multiplication operations for 16 times and the addition operations for 16 times on 16 elements and one of the convolution kernels (i.e., the target convolution kernel) of the second convolution layer, such that an output of the target convolution kernel is obtained after 10 cycles. Similarly, j and P can respectively be 3 and 6 in this embodiment. Therefore, the third convolutional layer of FIG. 2 of the present disclosure performs the multiplication operations for 16 times and the addition operations for 16 times on 16 elements and one of the convolution kernels (i.e., the target convolution kernel) of the third convolution layer, such that an output of the target convolution kernel is obtained after 6 cycles. Alternatively, j and P can respectively be 4 and 6 in this embodiment. Therefore, the fourth convolutional layer of FIG. 2 of the present disclosure performs the multiplication operations for 16 times and the addition operations for 16 times on 16 elements and one of the convolution kernels (i.e., the target convolution kernel) of the fourth convolution layer, such that an output of the target convolution kernel is obtained after 6 cycles.

[0022] It should be understood that the present disclosure does not limit an execution order and execution times of step S120 and step S130. In addition, the CNN can also include a plurality of fully connected layers for classification. However, since an operating principle of the fully connected layer is already known to those skilled in the art, the details thereof are omitted herein. In short, in step S140 of FIG. 1 of the present disclosure, when a k-th fully connected layer of the CNN performs an operation and a total number of records of input data of the k-th fully connected layer is M*N, the N multipliers and the N adders are used to complete conversion operations of N records of the input data in one cycle, such that an output of the k-th fully connected layer is obtained after M cycles. Here, k and M are integers greater than or equal to 1.

[0023] As shown in FIG. 2, k and M can be 1 and 13, respectively. Therefore, when a first fully connected layer of FIG. 2 performs the operation, these 16 multipliers and 16 adders are used to complete the conversion operations of 16 records of input data in one cycle, such that an output of the first fully connected layer is obtained after 13 cycles. Similarly, k and M can both be 2. Therefore, when a second fully connected layer of FIG. 2 performs the operation, these 16 multipliers and 16 adders are used to complete the conversion operations of 16 records of input data in one cycle, such that an output of the second fully connected layer is obtained after 2 cycles.

[0024] In conclusion, compared with the existing CNN, the present disclosure provides a method for improving a CNN to perform computations, such that complicated storage control and intermediate buffering mechanisms are not required, and an area and power consumption used for such implementation are relatively small.

[0025] The foregoing description of the exemplary embodiments of the disclosure has been presented only for the purposes of illustration and description and is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Many modifications and variations are possible in light of the above teaching.

[0026] The embodiments were chosen and described in order to explain the principles of the disclosure and their practical application so as to enable others skilled in the art to utilize the disclosure and various embodiments and with various modifications as are suited to the particular use contemplated. Alternative embodiments will become apparent to those skilled in the art to which the present disclosure pertains without departing from its spirit and scope.