DEEP NEURAL NETWORKS TRAINING FOR SPEECH AND PATTERN RECOGNITION

[0001] Acoustic modeling techniques that use context-dependent deep neural network hidden Markov models (CD-DNN-HMMs) for speech recognition or speech-to-text transcription can outperform acoustic modeling techniques that use conventional Gaussian-mixture based HMMs. Unlike Gaussian-mixture based HMMs, CD-DNN-HMMs use artificial neural networks with multiple hidden layers to directly model tied context-dependent states. However, the training of CD-DNN-HMMs for use in speech recognition is generally more time consuming that the training of Gaussian-mixture based HMMs. This larger amount of training time for CD-DNN-HMMs is a major obstacle to the widespread adoption and use of CD-DNN-HMMs for speech recognition.

[0002] The training of conventional Gaussian-mixture based HMMs for speech recognition may be optimized via parallelization. For example, the Baum-Welch training of Gaussian-mixture based HMMs may include statistics collection that is parallelized over hundreds or even thousands of servers. In such training, speech utterances may be processed independently across multiple servers. At the end of a batch of hundreds of millions of frames, partial statistics from the servers may be merged, and an updated model may be distributed to the servers. However, techniques for training Gaussian-mixture based HMMs are inapplicable to the training of CD-DNN-HMMs due to differences in model type, training procedures, and computation resource usage.

[0003] Described herein are enhanced techniques for training deep neural networks (DNN), herein referred to as DNNs, to speed up the training of the DNNs for use in performing pattern recognition and data analysis, such as speech recognition. In various embodiments, the DNNs may be context-dependent DNNs or context-independent DNNs. In some instances, such as for speech recognition, the context-dependent DNNs may be used in conjunction with hidden Markov Models (HMMs). In such instances, the combination of context-dependent DNNs and HMMs is known as context-dependent DNN-HMMs (CD-DNN-HMMs). Thus, the techniques described herein for training DNNs may be applied to train the CD-DNN-HMMs. The techniques described herein may include the use of a pipelined algorithm to parallelize the training of the DNNs across multiple multi-core processors, such as multiple general-purpose graphics processing units (GPGPUs). Accordingly, multiple layers of DNNs may be processed in parallel on the multiple multi-core processors. Further, the pipelined algorithm may be configured to process input data sample batches having a size that is defined to optimize a tradeoff between computation accuracy and execution efficiency.

[0004] The techniques may further include the use of model striping. In model striping, the top layer, which is the output layer of the DNNs, may be processed in parallel across multiple multi-core processors. Load balancing multiple layers of the DNNs across different multi-core processors may be another technique that is implemented. Other techniques may include the use of pipeline roundtrip reduction and parallelized data transfer with computation during the execution of the pipelined algorithm.

[0005] This Summary is provided to introduce a selection of concepts in a simplified form that is further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.

[0012] Described herein are enhanced techniques for training the deep neural networks (DNNs), herein referred to as DNNs. In various embodiments, the DNNs may be context-dependent DNNs or context-independent DNNs. In some instances, such as for speech recognition, the context-dependent DNNs may be used in conjunction with hidden Markov Models (HMMs). In such instances, the combination of context-dependent DNNs and HMMs is known as context-dependent DNN-HMMs (CD-DNN-HMMs). Thus, the techniques described herein for training DNNs may be applied to train the CD-DNN-HMMs. The training of the DNNs as described herein may be used to speed up the preparation of the DNNs for use in performing pattern recognition and data analysis, such as speech recognition. The techniques may include the use of a pipelined algorithm to parallelize the training of the DNNs across multiple multi-core processors, such as multiple general-purpose graphics processing units (GPGPUs). In at least one implementation, the multi-core processors may exchange data through an internal interface bus (e.g. PCIe), rather than a network. Accordingly, multiple layers of the DNNs may be processed in parallel on the multiple multi-core processors. Further, the pipelined algorithm may be configured to process input data sample batches having a size that is defined to optimize a tradeoff between computation accuracy and execution efficiency. In other words, the size may maximize both computation accuracy and execution efficiency of the pipelined algorithm 110.

[0013] The training techniques may further include the use of model striping. In model striping, the top layer, which is the output layer of the DNNs, may be processed in parallel across multiple multi-core processors. Load balancing multiple layers of the DNNs across different multi-core processors may be another training technique that is implemented. Other techniques may include the use of pipeline roundtrip reduction and parallelized data transfer with computation during the execution of the pipelined algorithm. The model striping of a top layer may be applied more frequently in the training of context-dependent DNNs because in context-independent DNNs the top layer size is typically much smaller than that in the context-dependent DNNs.

[0014] The techniques may reduce the amount of time used to train the DNNs for a particular purpose, such as for speech recognition. The decreased training time may lead to an increase in the implementation and usage of the DNNs in performing speech-to-text transcription or text-to-speech synthesis. Various examples of techniques for training of the DNNs for data analysis in accordance with various embodiments are described below with reference to FIGS. 1-5.

[0015] FIG. 1 is a block diagram that illustrates an example scheme 100 for implementing a training engine that uses a pipelined algorithm to train DNNs. The example scheme 100 may include a training engine 102 and a data analysis engine 104. The training engine 102 may be running on a computing device 106. The computing device 106 may include multiple multi-core processors 108(1)-108(N). In some embodiments, each of the multi-core processors 108(1)-108(N) may be a general-purpose graphics processing unit (GPGPU). In other embodiments, each of the multi-core processors 108(1)-108(N) may be a field-programmable gate array (FPGA), or another type of customizable multi-core processor. The multi-core processors 108(1)-18(N) may exchange data through an internal bus, such as a peripheral component interconnect express (PCIe) bus. However, in additional embodiments, one or more of the multi-core processors 108 may be substituted with a single-core processor in some instances.

[0016] The training engine 102 may use a pipelined algorithm 110 to train DNNs 112 for performing data analysis, such as for use in speech recognition. The DNNs 112 may be a multi-layer perceptron (MLP). As such, the DNNs 112 may include a bottom input layer 114(1) and a top layer 114(N), as well as multiple hidden layers, such as the multiple layers 114(2)-114(4), in which N denotes any integer. In some embodiments, the context dependent DNNs 112 may include a total of eight layers. Accordingly, in such embodiments, the value of N is “8” as there are eight layers. In various embodiments, the DNNs 112 may be context-dependent DNNs or context-independent DNNs. Training data 116 may be used by the pipelined algorithm 110 as training data to train the DNNs 112. The training data 116 may be a speech corpus that includes a collection of sample speech from human speakers. For example, the speech corpus may include North American English speech samples collected from speakers of North American English in the United States and Canada. However, in other implementations, the training data 116 may include sample speech in other respective languages (e.g., Chinese, Japanese, French, etc.), depending on the desired language of the speech to be recognized, or other kinds of training data for entirely different applications like handwriting or images.

[0017] The computations performed by the pipelined algorithm 110 may be parallelized across the multi-core processors 108(1)-108(N). For example, during pipelined back-propagation, a computation on input data performed by the multi-core processor 108(1) may produce a first computation result. The first computation result may be pipelined to the multi-core processor 108(2) for further computation to generate a second computation result. However, concurrent with the generation of the second computation result, the multi-core processor 108(1) may be processing additional input data to generate a third computation result. Such concurrent computations by the multi-core processors 108(1)-108(N) may result in a pipelining of computations that train the DNNs 112, and, accordingly, to a reduction of computation time due to the resulting parallelism of computation.

[0018] In various embodiments, the computations performed by the pipelined algorithm 110 may be enhanced using one or more techniques, such as batch size optimization 118, layer grouping 120, model striping 122, pipeline roundtrip reduction 124, and data transfer parallelization 126. Since the training data 116 is processed by the pipelined algorithm as batches of input samples 128, batch size optimization 118 may include configuring the size of each input sample batch to optimize a tradeoff between computation accuracy and execution efficiency.

[0019] Further, the layers 114(1)-114(N) in the DNNs 112 may have varying sizes due to differences in the number of units in each layer. Thus, layer grouping 120 is the grouping of at least two layers of the layers 114(1)-114(N) for executing by a single multi-core processor for the purpose of load balancing. For example, a largest layer in the DNNs 112 may have a size that is ten times larger than that of the one or more smallest layers. Accordingly, it may be more efficient to devote a particular multi-core processor to process the largest layer, while processing two or more of the smallest layers on another multi-core processor. Such grouping may further eliminate some pipeline roundtrip delays and improve efficiency.

[0020] Model striping 122 is the parallelized processing of a particular layer of the DNNs 112 across multiple multi-core processors, such as a plurality of the processors of the multi-core processors 108(1)-108(N). For example, the top layer 114(N) of the DNNs 112 may have a size that is ten times larger than that of the next largest layer in the DNNs 112. Accordingly, the processing of the top layer 114(N) may be paralleled across multiple multi-core processors. In this way, the model striping 122 of the top layer 114(N) may reduce the execution time of the pipelined algorithm 110 for training the DNNs 112.

[0021] A computation iteration of the pipelined algorithm 110 may execute the following steps in sequence: forward propagation of input data, error back propagation, and model update. However, pipeline roundtrip reduction 124 is based on the execution of a model update step before a forward propagation step during one or more computation iterations. The execution of the model update step before the forward propagation step may reduce pipeline roundtrip delay, which may in turn reduce the overall execution time of the pipelined algorithm 110 in completing the training. Additionally, data transfer parallelization 126 may include parallelizing the streaming of the output data from a computation iteration of the pipelined algorithm 110 with other steps in the computation iteration. In instances in which streaming time is shorter than computation time, such parallelization may reduce or eliminate time delay due to the exchange of data between multi-core processors during execution of the pipelined algorithm 110.

[0022] Thus, by using the pipelined algorithm 110 and the training data 116, the training engine 102 may produce trained DNNs 130 from the DNNs 112. In turn, the data analysis engine 104 may use the trained DNNs 130 to produce output data 132 from the input data 134. In some implementations, the data analysis engine 104 may be a speech-to-text engine that uses the trained DNNs 130 in the form of trained context-dependent DNN-HMMs. The speech-to-text engine may use the trained context-dependent DNN-HMMs to produce output data 132 in the form of output text from the input data 134 that is in the form of input speech. The data analysis engine 104 may be executed on the computing device 106 or a computing device that is similar to the computing device 106. Moreover, the data analysis engine 104 may receive live input data 134 from a microphone and audio processing components of the computing device. Alternatively, the data analysis engine 104 may receive input data 134 from a media file or stream, for example for the purpose of audio-indexing of the spoken content in said media file/stream. In other embodiments, the data analysis engine 104 may also be a text-to-speech engine that uses the trained context-dependent DNNs to synthesize output speech based on input text.

[0023] In alternative embodiments, the pipelined algorithm 110, as enhanced with one or more of the techniques 118-126, may be implemented to produce trained context-independent DNNs 130 under other scenarios that exhibit similar characteristics. In this way, context-independent forms of the DNNs 112 may be trained with appropriate training data for a variety of other data analysis purposes. The characteristics may include a larger set of training data that results in prolonged processing time (e.g., greater than 50 million samples), the implementation of DNNs in which the output of each network of the DNNs exceeds a threshold (e.g., greater than four thousand data outputs), and/or so forth. The data analysis purposes may include using trained context-independent DNNs for activities such as image recognition, computer vision, video tracking, and/or so forth.

[0024] FIG. 2 is an illustrative diagram that shows example components of the computing device 106 that implements the training engine 102 to train the DNNs 112. In various embodiments, the DNNs 112 may be context-dependent DNNs or context-independent DNNs. The computing device 106 may include one or more main processors 202, a plurality of multi-core processors 108(1)-108(N), an interface bus 204, a user interface 206, a network interface 208, and memory 210. In various embodiments, the computing device 106 may be a server, a desktop computer, another type of electronic device, or a combination of the above that is capable of hosting multiple multi-core processors to process data.

[0025] Each of the main processors 202 may be a single-core processor or a multi-core processor. The main processors 202 may be responsible for executing an operating system that is installed on the computing device 106. In some embodiments, each of the multi-core processors 108(1)-108(N) may be a general-purpose graphics processing unit (GPGPU). In other embodiments, each of the multi-core processors 108(1)-108(N) may be a field-programmable gate array (FPGA), or another type of customizable processor. In still other embodiments, the multi-core processors 108(1)-108(N) may be a heterogeneous mixture of multiple types of processors. For example, the multi-core processors 108(1)-108(N) may be a combination of one or more GPGPUs and one or more FPGAs. The multi-core processors 108(1)-18(N) may exchange data through the interface bus 204. For example, the interface bus 204 may be a peripheral component interconnect express (PCIe) bus. In some implementations, the main processors 202 and the multi-core process 108(1)-108(N) may be the same processors.

[0026] The user interface 206 may include a data output device (e.g., visual display, audio speakers), and one or more data input devices. The data input devices may include, but are not limited to, combinations of one or more of keypads, keyboards, mouse devices, touch screens that accept gestures, microphones, voice or speech recognition devices, and any other suitable devices or other electronic/software selection methods.

[0027] The network interface 208 may include wireless and/or wireless communication interface components that enable the computing device 106 to transmit and receive data via a network. In various embodiments, the wireless interface component may include, but is not limited to cellular, Wi-Fi, Ultra-wideband (UWB), Bluetooth, satellite transmissions, and/or so forth. The wired interface component may include a direct I/O interface, such as an Ethernet interface, a serial interface, a Universal Serial Bus (USB) interface, and/or so forth. As such, the computing device 106 may have network capabilities. For example, the computing device 106 may exchange data with other electronic devices (e.g., laptops computers, servers, etc.) via one or more networks, such as the Internet.

[0028] The memory 210 may be implemented using computer-readable media, such as computer storage media. Computer-readable media includes, at least, two types of computer-readable media, namely computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that may be used to store information for access by a computing device. In contrast, communication media may embody computer readable instructions, data structures, program modules, or other data in a modulated data signal, such as a carrier wave, or other transmission mechanism. As defined herein, computer storage media does not include communication media.

[0029] The memory 210 of the computing device 106 may store the modules of the training engine 102. The modules may include an operating system 212, an algorithm execution module 214, a batch generation module 216, an algorithm configuration module 218, a load balance module 220, and a model striping module 222. Additionally, a data store 224 may reside in the memory 210.

[0030] The operating system 212 may include components that enable the computing device 106 to receive data via various inputs (e.g., user controls, network interfaces, and/or memory devices), and process the data using the main processors 202 to generate output. The operating system 212 may further include one or more components that present the output (e.g., display an image on an electronic display, store data in memory, transmit data to another electronic device, etc.). The operating system 212 may enable a user to interact with modules of the training engine 102 using the user interface 206. Additionally, the operating system 212 may include other components that perform various other functions generally associated with an operating system.

[0031] The algorithm execution module 214 may use the pipelined algorithm 110 to train the DNNs 112 based on the training data 116, which may be a speech corpus. In instances in which the DNNs 112 are trained for speech analysis purposes, the DNNs 112 may be context-dependent DNNs that are used in conjunction with HMMs. However, in other instances in which the DNNs are trained for other types of data analysis, the DNNs may be context-independent DNNs. The DNNs 112 may be a MLP that models the posterior probability P_s|o(s|o) of a class s, given an observation vector o, as a stack of (L+1) layers of log-linear models. The first L layers, l=0 . . . L−1, model posterior probabilities of hidden binary vectors h^l given input vectors v^l, while the top layer L models the desired class posterior as:

$\begin{array}{cc}{P}_{hv}^l\left(h^lv^l\right)=\prod _{j=1}^{N^l}\frac{e^{z_j^l\left(v^l\right)·h_j^l}}{e^{z_j^l\left(v^l\right)·1}+e^{z_j^l\left(v^l\right)·0}},0\le l<L& \left(1\right)\\ P_{sv}^L\left(sv^L\right)=\frac{{}^{z_s^L\left(v^L\right)}}{\sum _{s^{\prime }}{}^{z_{s^{\prime }}^L\left(v^L\right)}}={\mathrm{softmax}}_s\left(z^L\left(v^L\right)\right)& \left(2\right)\\ z^l\left(v^l\right)={\left(W^l\right)}^Tv^l+a^l;v^l\stackrel{\mathrm{def}}{=}E^{l-1}\left\{h^{l-1}\right\}& \left(3\right)\end{array}$

with weight matrices W^l and bias vectors a^l, where h_j^l and z_j^l(v^l) are the j-th component of h^l and z^l (v^l), respectively. Full out-summation over all hidden variables, which is infeasible, may be approximated by a “mean-field approximation” where the input v^l to each hidden layer is taken as the expectation of the output vector h^l of the previous layer. Further, for use with the DNNs 112, state posteriors P_s|o(s|o) may be converted to scaled likelihoods by dividing by their prior.

[0032] Accordingly, the algorithm execution module 214 may train the DNNs 112 according to the following cross entropy criterion:

D=Σ_t=1^Tcorpus log P_s|o(s(t)|o(t)),  (4)

by using stochastic gradient descent:

$\begin{array}{cc}\left(W^l,a^l\right)\leftarrow \left(W^l,a^l\right)+\varepsilon \frac{\partial D}{\partial \left(W^l,a^l\right)},0\le l\le L,& \left(5\right)\end{array}$

with learning rate ε and gradients as follows:

$\begin{array}{cc}\frac{\partial D}{\partial W^l}=\sum _tv^l\left(t\right)\left({\omega }^l\left(t\right)e^l\left(t\right)\right);\frac{\partial D}{\partial a^l}=\sum _t{\omega }^l\left(t\right)e^l& \left(6\right)\\ e^L\left(t\right)={\left(\log \mathrm{softmax}\right)}^{\prime }\left(z^L\left(v^L\left(t\right)\right)\right)& \left(7\right)\\ e^{l-1}\left(t\right)=W^l·{\epsilon }^l\left(t\right)·e^l\left(t\right)\mathrm{for}0\le l<L& \left(8\right)\\ {\omega }^l\left(t\right)=\{\begin{array}{cc}\mathrm{diag}({\sigma }^{\prime }\left(z^l\left(v^l\left(t\right)\right)\right)& \mathrm{for}0\le l<L\\ 1& \mathrm{else}\end{array}& \left(9\right)\end{array}$

and with error signals e^l(t), the component-wise derivatives σ_j, (z)=σ_j(z)·(1−σ_j(z)), and (log softmax)_j, (z)=δ_s(t),j−softmax_j(z), and Kronecker delta δ.

[0033] The training of the DNNs 112 may be achieved by pipelining computations of back-propagation in a parallelized fashion (i.e., simultaneously executing multiple computations) using the multiple multi-core processors 108(1)-108(N). This pipelined back-propagation is an approximation due to the distributions of the layers 114(1)-114(N) of the DNNs 112 across the multiple multi-core processors 108(1)-108(N) to form a pipeline. In the pipeline, data flows sequentially from each multi-core processor to the next multi-core processor, in which all of the multi-core processors 108(1)-108(N) work simultaneously on the received data. However, such data flows may result in pipeline roundtrip delays that affect computation accuracy.

[0034] Thus, because of the highly non-linear nature of the training, reasonable convergence (i.e., training completion) may be achieved by performing the stochastic gradient descent, as described above in expression (5), using discretely sized batches of randomly sampled frames 128 from the training data 116. The size of the batches may be limited by the parallelized computation nature of the pipelined algorithm 110. For instance, model updates to the DNNs 112, which involve the exchange of data between multi-core processors, are used for the computation iterations of the algorithm 110. However, model updates across multiple multi-core processors may use a high amount bandwidth during the execution of the pipelined algorithm 110. In one example, the DNNs 112 (with seven hidden layers) may include 10⁸ parameters. In such an example, the processing of a reasonably sized batch of sample frames with respect to the DNNs 112 may translate into the gathering and redistribution of 400 megabyte (MB) worth of gradients and another 400 MB of model parameters by each of the multi-core processors 108(1)-108(N).

[0035] In other words, the size of each batch that is used to train the DNNs may be constrained by two factors. The upper constraint for the batch size is the frequency of model updates. Larger batch size for the batches of sample frames 128 may result in less model updates. However, increasing the batch size may result in the loss of computation accuracy, especially during early computation iterations of the pipelined propagation algorithm 110. Such loss of computation accuracy may result in prolonged execution time for the pipelined algorithm 110 to reach convergence, i.e., completes the training of the DNNs 112. In extreme cases, the prolonged execution time may even result in a failure of the pipelined algorithm 110 to reach convergence, i.e., failure to complete the training. The lower constraint for the batch size is the efficiency in the utilization of the multi-core processors. The efficiency in the use of the computation cycles performed by the multi-core processors may decrease as a batch size for the sample frame batches 128 is reduced. Thus, excess reduction in batch size may also lead to inefficiencies that prolong the execution time for the pipelined algorithm 110 to reach convergence.

[0036] In practical terms, the batch generation module 216 may configure the batch size for the sample frame batches 128 based on rates of data transfers between the multi-core processors and a number of operations per second that each of the multi-core processors 108(1)-108(N) is capable of executing. For example, given an array of 2-4 GPGPUs that are capable of 2-4 tera floating point operations per second (TFLOPS), and transfer rates of 6 gigabytes (GB)/s between the GPGPUs, the batch size may be in the range of 256 to 1024 sample frames per sample batch. Accordingly, batch size optimization 118 may produce the lowest amount of computation accuracy loss and the smallest amount of efficiency loss.

[0037] Generally speaking, the batch generation module 216 may configure a larger batch size when the rates of data transfers for the multi-core processors 108(1)-108(N) are relatively superior to the execution speeds of the multi-core processors 108(1)-108(N). Conversely, the batch generation module 216 may configure a smaller batch size when the execution speeds of the multi-core processors 108(1)-108(N) are relatively superior to the rates of data transfers between the multi-core processors 108(1)-108(N). The batch generation module 216 may partition the training data 116 into randomly sampled frame batches 128 based on the configured batch size.

[0038] The algorithm configuration module 218 may modify the pipelined algorithm 110 with the pipeline roundtrip reduction 124 and the data transfer parallelization 126 techniques. The algorithm configuration module 218 may perform each modification based on a corresponding user configuration input. Each of the computation iterations performed by the pipelined algorithm 110 may execute the following steps in sequence: forward propagation of input data, error back propagation, and model update. The forward propagation of the input data is described by the equations (1)-(3), the error back propagation is described by the equation (8), and the model update is described by the expression (5).

[0039] However, the pipeline roundtrip reduction 124 modification may cause the pipelined algorithm 110 to perform a model update step before a forward propagation step during one or more of the computation iterations. Each application of pipeline roundtrip reduction 124 may reduce the pipeline roundtrip delay by one batch, which may in turn reduce approximation performed by the pipelined algorithm 110. In turn, the reduction in approximation may increase the execution efficiency of the pipelined algorithm 110.

[0040] Further, the data transfer parallelization 126 technique involves the parallelization of data transfer with computation. A first part of the data transfer parallelization 126 may occur after the performance of an error back propagation step. In this part, output data from a computation at a multi-core processor that processes an upper layer may be streamed to another multi-core processor that processes a lower layer of the DNNs 112. Such streaming may be performed in parallel or partially in parallel with a model update step and/or an input data forward propagation step, as the model update step and the forward propagation step use data that is different from the output data.

[0041] A second part of the data transfer parallelization 126 may occur after the performance of the input data forward propagation step. In this part, output data from a computation at the multi-core processor that processes the lower layer may be streamed to the multi-core processor that processes the upper layer in the DNNs 112. Such streaming may be performed in parallel or partially in parallel with the computation of an error for another error back propagation step. Thus, since streaming time is generally short than compute time, the use of the data transfer parallelization 126 may reduce or eliminate any time delay resulting from the exchange of data between multiple multi-core processors.

[0042] The load balance module 220 may determine whether to perform layer grouping 120. In order to perform layer grouping 120, the load balance module 220 may distribute the layers 114(1)-114(N) of the DNNs 112 between the multi-core processors 108(1)-108(N) in groups for processing by the pipelined algorithm 110. For instance, the layers 114(1)-114(N) may have varying sizes and thus takes varying runtimes to process. Thus, a large layer (such as the top layer 114(N)) may act as a bottleneck that prevents the multi-core processors from processing the other smaller layers during the execution of the pipelined algorithm 110. In other instances in which there are more layers to be processed than multi-core processors, the layer grouping 120 may be used to allocate the layers into groups for processing by the multi-core processors.

[0043] For example, the load balance module 220 may assign each of four groups of multiple layers from the layers 114(1)-114(N) to a corresponding multi-core processor, such that the amount of data processed by each of the four multi-core processors for its respective assigned layers is equalized or as equalized as possible. In another example, the load balance module 220 may assign a largest layer to be processed by a first multi-core processor, while assigning the remaining multiple layers to be processed by a second multi-core processor. However, the load balance module 220 may assign each of one or more sets of multiple layers of the layers 114(1)-114(N) to a corresponding multi-core processor in any combination, in order to balance computations performed for the pipelined algorithm 110 between the multi-core processors 108(1)-108(N). Thus, the load balance module 220 may perform layer grouping 120 based on the number of the multi-core processors 108(1)-108(N) and the size of each layer in the DNNs 112. The grouping of multiple layers into a single set for processing may also eliminate pipeline roundtrip delay, thereby increasing the execution efficiency of the pipelined algorithm 110.

[0044] The model striping module 222 may determine whether the algorithm execution module 214 is to be configured to perform model striping 122 with respect to the top layer 114(N) of the DNNs 112. Model striping 122 is the parallelization of the processing of a particular layer of the DNNs 112 across multiple multi-core processors, such as the multi-core processors 108(1)-108(N). In various embodiments, the model striping 122 may be applied more frequently to the training of context-dependent DNNs because in context-independent DNNs the top layer size is typically much smaller than that in the context-dependent DNNs. By implementing model striping with respect to the top layer 114(N), the input v^l of the top layer 114(N) may be distributed across the multi-core processors 108(1)-108(N) in forward propagation, in which each of the multi-core processors 108(1)-108(N) may compute a slice of the output vector E^l{H^l}. The slices may then be distributed to the other multi-core processors of the multi-core processors 108(1)-108(N) for computing the next layer. In back-propagation, error vectors are parallelized as slices, but the resulting matrix products from each slice are partial sums that are further summed up. As a result, in both forward propagation and back propagation, each vector is transferred K−1 times, in which K is the number of the multi-core processors 108(1)-108(N).

[0045] In various embodiments, the decision to implement model striping 122 for the top layer 114(N) may be based on a comparison of the size of the top layer 114(N) to a size of at least one other layer in the DNNs 112. The model striping module 222 may detect the size of each layer of the DNNs 112. Accordingly, in one instance, the model striping module 222 may direct the implementation of the model striping 122 when a ratio between a size of the top layer 114(N) and a size of the next largest layer in the DNNs 112 exceeds a predetermined ratio value threshold. For example, model striping 122 may be implemented for the top layer 114(N) when the size of the top layer 114(N) is over ten times larger than the next largest layer in the DNNs 112.

[0046] In another instance, the model striping module 222 may direct the implementation of the model striping 122 when a ratio between a size of the top layer 114(N) and a total size of the remaining layers in the DNNs 112 exceeds a predetermined ratio value threshold. For example, model striping 122 may be implemented for the top layer 114(N) when the size of the top layer 114(N) is over four times larger than the total size of the remaining layers in the DNNs 112.

[0047] In additional instances, the model striping module 222 may use other predetermined ratio value thresholds to determine whether to implement model striping 122 for the top layer 114(N). For example, the model striping module 222 may compare the size of the top layer 114(N) to an average size of the hidden layers, such as the hidden layers 114(2)-114(4), to produce a ratio value, a size of the smallest layer (e.g., input layer 114(1)) of the DNNs 112 to produce a ratio value or a total size of the hidden layers 114(2)-114(4) produce a ratio value, etc. Accordingly, depending on the particular ratio value used, the model striping module 222 may implement model striping 122 for the top layer 114(N) when the particular ratio value exceeds a corresponding ratio value threshold. The implementation of model striping 122 on the top layer 114(N) may alleviate bottlenecks in the execution of the pipelined algorithm 110 caused by an excessively large top layer 114(N).

[0048] The data store 224 may store data that are used by the various modules. In at least one embodiment, the data store may store the training data 116, the DNNs 112, temporary variables and other data used for training the DNNs 112, and the DNNs 130. Each set of data may be stored in a data table, a data array, and/or other data storage structures.

[0049] While the multi-core processors 108(1)-108(N) are described as residing on the computing device 106 and connected by the interface bus 204 in the above the embodiments, the multi-core processors 108(1)-108(N) may also reside on different computing devices in other embodiments. In some alternative embodiments, each of the multi-core processors 108(1)-108(N) may reside on a corresponding computing device, and may exchange data through a network via a network interface. The network may be a local area network (LAN), a wide area network (WAN), or a combination of both, such as the Internet. In other alternative embodiments, at least two of the multi-core processors 108(1)-108(N) may reside on different computing devices. In such embodiments, multi-core processors on the same computing device may use an interface bus of the computing device to exchange data, while multi-core processors on different computing devices may exchange data via the network.

[0050] FIGS. 3-5 describe various example processes for using a pipelined algorithm to train DNNs for performing data analysis, such as for the purpose of speech recognition. The order in which the operations are described in each example process is not intended to be construed as a limitation, and any number of the described operations may be combined in any order and/or in parallel to implement each process. Moreover, the operations in each of the FIGS. 3-5 may be implemented in hardware, software, and a combination thereof. In the context of software, the operations represent computer-executable instructions that, when executed by one or more processors, cause one or more processors to perform the recited operations. Generally, computer-executable instructions include routines, programs, objects, components, data structures, and so forth that cause the particular functions to be performed or particular abstract data types to be implemented.

[0051] FIG. 3 is a flow diagram that illustrates an example process 300 for modifying the execution order of propagation steps in the execution of a pipelined algorithm to reduce delay due to the exchange of data between multi-core processors.

[0052] At block 302, the training engine 102 may assign the pipelined algorithm 110 to train the DNNs 112 for performing data analysis. The training of the DNNs 112 may be achieved by pipelining computations of back-propagation in a parallelized fashion (i.e., simultaneously executing multiple computations) using the multiple multi-core processors 108(1)-108(N).

[0053] At decision block 304, the training engine 102 may determine whether to implement the pipeline roundtrip reduction 124. The pipeline roundtrip reduction 124 includes configuring the pipelined algorithm 110 to perform a model update step before a forward propagation step during one or more of the computation iterations. In various embodiments, the training engine 102 may determine whether to implement the pipeline roundtrip reduction 124 based on a configuration input from a user.

[0054] Accordingly, if the training engine 102 implements the pipeline roundtrip reduction 124 (“yes” at decision block 304), then the process 300 may proceed to block 306.

[0055] At block 306, the training engine 102 may modify the pipelined algorithm 110 to implement the pipeline roundtrip reduction 124 during execution. The pipeline roundtrip reduction may reduce the pipeline roundtrip delay by one batch for each of the computation iterations performed by the pipelined algorithm 110, which may in turn reduce approximation performed by the pipelined algorithm 110.

[0056] At decision block 308, the training engine 102 may determine whether to implement the data transfer parallelization 126. The data transfer parallelization 126 may include parallelizing the streaming of the output data from a computation iteration of the pipelined algorithm 110 with other steps in the computation iteration. In various embodiments, the training engine 102 may determine whether to implement the data transfer parallelization 126 based on a configuration input from a user. Accordingly, if the training engine 102 implements the data transfer parallelization 126 (“yes” at decision block 308), then the process 300 may proceed to block 310.

[0057] At block 310, the training engine 102 may modify the pipelined algorithm 110 to implement the data transfer parallelization during execution. Accordingly, since streaming time is generally short than compute time, the configuration of pipelined algorithm 110 to implement the data transfer parallelization 126 may reduce or eliminate any time delay resulting from the exchange of data between multiple multi-core processors during execution. At block 312, the training engine 102 may provide the pipelined algorithm 110 to train the DNNs 112 for performing data analysis, such as speech recognition.

[0058] However, returning to decision block 304, if the training engine 102 does not implement the pipeline roundtrip reduction 124 (“no” at decision block 304), then the process 300 may proceed directly to decision block 308. Further, returning to decision block 308, if the training engine 102 does not implement the data transfer parallelization 126 (“no” at decision block 308), then the process 300 may proceed directly to block 312. At block 312, the training engine 102 may provide the pipelined algorithm 110 to train the DNNs.

[0059] FIG. 4 is a flow diagram that illustrates an example process 400 for configuring the size of batches 128 obtained from the training data to reduce training runtime of the DNNs. At block 402, the training engine 102 may determine a batch size for partitioning training data 116 (e.g., a speech corpus) that is used to train the DNNs. In at least one embodiment, the training engine 102 may configure the batch size based on rates of data transfers between the multi-core processors and execution speed, i.e., a number of operations per second, of each of the multi-core processors 108(1)-108(N).

[0060] For example, the training engine 102 may designate a larger batch size when the rates of data transfers of the multi-core processors 108(1)-108(N) are relatively superior to the execution speeds of the multi-core processors 108(1)-108(N). Conversely, training engine 102 may designate a smaller batch size when the execution speeds of the multi-core processors 108(1)-108(N) are relatively superior to the rates of data transfers between the multi-core processors 108(1)-108(N).

[0061] At block 404, the training engine 102 may partition the training data into batches 128 according to the batch size, in which each batch is designed to optimize the tradeoff between computation accuracy and execution efficiency. In various embodiments, each batch may include randomly sampled frames from the training data 116.

[0062] At block 406, the training engine 102 may execute the pipelined algorithm 110 to train the DNNs 112 using the batches 128 derived from the training data 116. In various embodiments, the pipelined algorithm 110 may have been modified according to the process 300 described in FIG. 3.

[0063] FIG. 5 is a flow diagram that illustrates an example process 500 for load balancing parallelized computations between multiple multi-core processors to reduce training runtime of the DNNs. In at least one embodiment, the process 500 may further describe block 406 of the process 400.

[0064] At block 502, the training engine 102 may allocate the batches 128 of sample frames from the training data 116 (e.g., a speech corpus) for training the DNNs 112. The training may be performed using the pipelined algorithm 110.

[0065] At decision block 504, the training engine 102 may determine whether to implement the layer grouping 120 to group at least two layers in the DNNs 112 together for processing on a single multi-core processor by the pipelined algorithm 110. In some embodiments, the at least two layers may be at least two consecutive layers. The layer grouping 120 may be performed to load balance computations performed for the pipelined algorithm 110 between the multi-core processors 108(1)-108(N). Thus, the number of groupings and the layers in each grouping may be based on the number of the multi-core processors 108(1)-108(N) and the size of each layer in the DNNs 112. Thus, if the training engine 102 implements the layer grouping 120 (“yes” at decision block 504), then the process 500 may proceed to block 506.

[0066] At block 506, the training engine 102 may group at least two layers of the DNNs 112 for processing on a single multi-core processor. In various embodiments, the training engine 102 may group the layers in the DNNs 112 into multiple sets of two or more layers, in which each of the multiple sets may be processed by a corresponding multi-core processor.

[0067] At decision block 508, the training engine 102 may determine whether to implement the model striping 122 to distribute the processing of the top layer 114(N) of the DNNs 112 across multiple multi-core processors. In various embodiments, the training engine 102 may make such a determination based on a ratio of the size of the top layer 114(N) to the size of at least one other layer in the DNNs 112. Thus, if the training engine 102 implements the model striping 122 for the top layer 114(N) (“yes” at decision block 508), then the process 500 may proceed to block 510.

[0068] At block 510, the training engine 102 may distribute the top layer 114(N) of the DNNs 112 across the multi-core processors 108(1)-108(N) for parallelized processing by the pipelined algorithm 110. However, in other embodiments, the training engine 102 may distribute the top layer 114(N) across a set of multiple multi-core processors other than the multi-core processors 108(1)-108(N), or distribute the top layer 114(N) to a plurality of but less than all of the multi-core processors 108(1)-108(N).

[0069] At block 512, the training engine 102 may pipeline an execution of the algorithm 110 on a set of multi-core processors to train the DNNs 112 based on the batches 128 of the training data 116. The set of multi-core processors may include the multi-core processors 108(1)-108(N). The pipelined algorithm 110 that is executed may be configured through the process 300 shown in FIG. 3 The training may produce the trained DNNs 112 that are used for a performing data analysis, such as speech recognition.

[0070] However, returning to decision block 504, if the training engine 102 does not implement the layer grouping 120 (“no” at decision block 504), then the process 500 may proceed directly to decision block 508. Further, at decision block 508, if the training engine 102 does not implement the model striping 122 for the top layer 114(N) (“no” at decision block 508), then the process 500 may proceed directly to block 512. Once again, at block 512, the training engine 102 may pipeline an execution of the algorithm 110 on a set of multi-core processors to train the DNNs 112 based on the batches 128 of training data 116.

[0071] The training techniques described herein may reduce the amount of time used to train DNNs for a particular purpose, such as for speech recognition. The decreased training time may lead to an increase in the implementation and usage of the DNNs in performing speech-to-text transcription or text-to-speech synthesis.

[0072] In closing, although the various embodiments have been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended representations is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as exemplary forms of implementing the claimed subject matter.