JP2018097418A - Information processing apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform accurate prediction using data obtained at different sampling rates.SOLUTION: An acquisition part 102 acquires a plurality of sensor values xdetected at different periods, and a recurrent neural network processing part 103 calculates a forward propagation value Xby performing forward propagation processing on the acquired sensor values. A queue management part 102a stores the calculated forward propagation value for each sampling rate in a queue of a memory 104 distinguished at different sampling rates. Then, the queue management part 102a fetches the forward propagation value Xstored for each queue of the memory 104 for the recurrent neural network processing part 103, and the recurrent neural network processing part 103 calculates a predicted value Y (S).SELECTED DRAWING: Figure 1

Description

  The present invention relates to an information processing apparatus that performs prediction using a neural network.

  Patent Document 1 describes that a variation pattern of a process output value is predicted by using a value obtained by sampling a variable representing a state of a process at a predetermined time interval, and processing it with a neural network.

JP-A-5-204407

  The technique described in Patent Document 1 has a problem that when a plurality of different types of information appearing in real time are to be processed, the data cannot be properly predicted because the sampling rate differs for each data.

  In addition, the degree of delay differs for each piece of information as the information generation is delayed. Therefore, there is a problem that if the prediction is performed for each sampling rate, the necessary calculation time is different, and this delay is promoted.

  Therefore, in order to solve the above-described problems, an object of the present invention is to provide an information processing apparatus capable of performing accurate prediction using data obtained at different sampling rates.

  In order to solve the above-described problem, the information processing apparatus of the present invention calculates a forward propagation value by performing forward propagation processing on a sensor value detected by a sensor, and calculates by the forward propagation processing unit. Queue storage means for storing the forward propagation values that have been performed for each queue distinguished at different sampling rates, queue management means for retrieving the forward propagation values at one time stored for each queue of the queue storage means, Prediction value calculation means for calculating a predicted value of a state at the next time of the one time based on the forward propagation value at the one time taken out by the queue management means.

  Further, the information processing method of the present invention is a forward propagation processing step for calculating a forward propagation value by forward propagation processing of sensor values detected at different sampling rates by a sensor in the information processing method of the information processing apparatus; A queue storage step for storing the forward propagation value calculated in the forward propagation processing step in a queue storage unit for each queue distinguished at the different sampling rates; A queue management step for extracting a forward propagation value at a time, and a predicted value for calculating a predicted value of a state at the next time of the one time based on the forward propagation value at the one time taken out by the queue management step A calculation step.

  Further, the information processing program of the present invention includes a forward propagation processing means for calculating a forward propagation value by forward propagation processing of sensor values detected at different sampling rates by a sensor, and the forward propagation processing means. Queue storage means for storing the calculated forward propagation value for each queue distinguished at the different sampling rates, and queue management means for extracting the forward propagation value at one time stored for each queue of the queue storage means And a predicted value calculating means for calculating a predicted value of a state at the next time of the one time based on the forward propagation value at the one time taken out by the queue management means.

  According to the present invention, forward propagation values obtained by forward propagation processing of sensor values are stored for each queue distinguished at different sampling rates, and prediction processing is performed by forward propagation processing using the queues. Thus, accurate prediction processing using each sensor value acquired at different sampling rates can be performed.

  Further, the present invention is based on the updating means for updating the parameters of the neural network by performing the back propagation process by the neural network using the prediction value calculated by the prediction value calculating means, and the updated parameter. And determining means for determining whether or not the forward propagation value stored in the queue storage means is important.

  According to the present invention, the parameters of the neural network are updated to determine which sampling rate sensor value is important and which sensor value is important among the sensor values acquired at different sampling rates. be able to.

  According to the present invention, prediction for a plurality of inputs having different sampling rates can be performed accurately and quickly.

3 is a functional block diagram illustrating functions of the information processing apparatus 100. FIG. 2 is a hardware configuration diagram of the information processing apparatus 100. FIG. It is explanatory drawing of the information memorize | stored for every cue | queue formed for every sampling rate. 3 is a flowchart showing processing of the information processing apparatus 100. It is a schematic diagram which performs the recurrent neural network process using the processor arrange | positioned for every module.

  Embodiments of the present invention will be described with reference to the accompanying drawings. Where possible, the same parts are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 is a functional block diagram illustrating functions of the information processing apparatus 100 according to the present embodiment. The information processing apparatus 100 includes an acquisition unit 102, a queue management unit 102a (queue management unit), a recurrent neural network processing unit 103 (forward propagation processing unit, predicted value calculation unit, update unit), a memory 104 (queue storage unit), And an output unit 105.

  FIG. 2 is a hardware configuration diagram of the information processing apparatus 100. As shown in FIG. 2, the information processing apparatus 100 shown in FIG. 1 physically includes one or more CPUs 11, RAMs 12 and ROMs 13 as main storage devices, and input devices 14 such as keyboards and mice as input devices. The computer system includes an output device 15 such as a display, a communication module 16 that is a data transmission / reception device such as a network card, and an auxiliary storage device 17 such as a hard disk or a semiconductor memory. Each function in FIG. 1 operates the input device 14, the output device 15, and the communication module 16 under the control of the CPU 11 by loading predetermined computer software on the hardware such as the CPU 11 and the RAM 12 shown in FIG. 2. In addition, it is realized by reading and writing data in the RAM 12 and the auxiliary storage device 17. Hereinafter, each functional block will be described based on the functional blocks shown in FIG.

  The acquisition unit 102 is a part that acquires the sensor value detected by the sensor 101. The sensor 101 detects the state of the external environment that changes in time series, such as a microphone or a video camera. The acquisition unit 102 outputs the sensor value input from the sensor 101 to the recurrent neural network processing unit 103, and acquires the forward propagation value subjected to forward propagation processing. The acquired forward propagation value is output to the queue management unit 102a together with the acquired time.

  The queue management unit 102a determines which queue in the memory 104 stores the time and the forward propagation value output from the acquisition unit 102, and stores the forward propagation value in the queue. That is, the queue management unit 102a determines whether or not the time is before the minimum time (that is, corresponding to the sampling rate) to be held in the queue, which is obtained from the discretization width. If the output time is before the minimum time obtained from the discretization width, the forward propagation value is discarded and not stored in the memory 104. If the output time is after the minimum time obtained from the queue width, the forward propagation value is stored in the queue corresponding to that time.

  As a result, the queue management unit 102a can store the forward propagation values in the memory 104 at a plurality of different predetermined sampling rates. Note that the queue management unit 102a stores the forward propagation value and a function indicating the calculation process in the memory 104 in association with each other.

  In the process described below, the sensor value detected by the sensor 101a is stored in a queue that the queue management unit 102a distinguishes for each sampling rate. However, a sensor that detects a different type of sensor value. The sensor values from 101b and 101c may also be stored separately in queues. In that case, it is necessary that the acquisition unit 102 and the queue management unit 102a manage information indicating the sensor type, and a queue corresponding to the sensor type is prepared in the memory 104.

  A recurrent neural network (RNN) processing unit 103 performs forward propagation processing on sensor values of different sampling rates acquired by the acquisition unit 102 by a neural network prepared for each sensor source before processing by the queue management unit 102a. Then, the forward propagation value is calculated. For example, the following processing is performed.

Forward propagation value X _t = g (f (x _t , w _f ), w _g ) (1)

Here, t is time, x _t is a sensor value at time t, and w _f and w _g are parameters in a neural network prepared for each sensor source.

The recurrent neural network processing unit 103 outputs the calculated forward propagation value _Xt to the queue management unit 102a, and the queue management unit 102a stores it in each queue of the memory 104 prepared for each sampling rate as described above. .

Memory 104, to form a queue for each sampling rate, per queue, in association forward propagation value X _t at time t. This time t is the time when the forward propagation value X _t is calculated by the recurrent neural network processing unit 103. FIG. 3 shows a specific example thereof. As shown in FIG. 3, queues 1 to 3 are formed in the memory 104, and a forward propagation value and a forward propagation function indicating a calculation process thereof are stored in association with time for each queue. Yes. In this embodiment, the queue 1 stores the forward propagation value with the sampling rate being the densest, the queue 2 stores the forward propagation value acquired at a sampling rate that is half that of the queue 1, and the queue 3 further includes The forward propagation value acquired at half the sampling rate is stored. That is, the forward propagation value is stored by changing the sampling rate for each queue.

The queue management unit 102 a takes out the forward propagation value X _{t for} which the allowable delay time T (greater than the discretization width of the sampling rate) has elapsed from each queue of the memory 104 and causes the recurrent neural network processing unit 103 to process it. Recurrent neural network processing unit 103 calculates the state value _{S t} at time t and the forward propagation values _{X 1} and forward propagation process using the state values _{S t,} the predicted value at time _{t + 1 Y t = Y (} S _t ) is calculated. For example, the following processing is performed.

S _t = F (w ₁ , w ₂ , w ₃ , S _t−1 , X ⁽¹⁾ _t , X ⁽²⁾ _t , X ⁽³⁾ _t ) (2)

Y _t = Y (S _t ) (3)
Here, w ₁ , w ₂ , and w ₃ are parameters for each queue in the recurrent neural network. F is a function for calculating the state at time t, and is for a predetermined forward propagation process. X ⁽¹⁾ _t is the forward propagation value stored at time t in queue 1, X ⁽²⁾ _t is the forward propagation value stored at time t in queue 2, and X ⁽³⁾ _t is the queue 1 is a forward propagation value stored at time t. Further, Y _t is a function for calculating the predicted value based on the state value S _t at time t.

  Equation (2) is for calculating the state value S when the forward propagation value is stored in the queue at time t. Depending on the sampling rate, queue 2 or queue 3 may not have a forward propagation value corresponding to time t. In that case, it can also be expressed as in equations (4) and (5).

S _{t + 1} = F (w ₁ , S _t , X ⁽¹⁾ _t ) (4)

S _{t + 2} = F (w ₁ , w ₂ , S _{t + 1} , X ⁽¹⁾ _t , X ⁽²⁾ _t ) (5)

The recurrent neural network processing unit 103 further performs learning processing by performing error back propagation processing of the recurrent neural network using the predicted value Y (S _t-1 ). For example, the following processing is performed.

An observation value y at time t is acquired. This is obtained by the sensor 101. Then, the recurrent neural network processing unit 103 calculates a loss (for example, a square error) from the observed value y at the time t and the predicted value Y (S _t-1 ), and the gradient with respect to the parameters w ₁ , w ₂ , and w ₃ . And the parameters w ₁ , w ₂ , and w ₃ are updated by a gradient descent method or the like.

Furthermore, recurrent neural network processing unit 103, to the function shown in equation (2), by backpropagation of ^{_{S t (BPTT), X (}} 1) t, X (2) t, X (3 ⁾ calculates _t, further, by backpropagation to X _t, and calculates the gradient of the parameter w _f and w _g in recurrent neural network, and updates the parameters w _f and w _g the like gradient descent.

By updating the parameters w ₁ , w ₂ , w ₃ , w _f , and w _g of the recurrent neural network in this way, an unimportant cue (sampling rate sensor value) and a weight for the input of the sensor value Becomes smaller, and an accurate predicted value can be calculated. When handling sensor values of a plurality of types of sensors, it is necessary to use the recurrent neural network processing parameter w for each sensor.

The output unit 105 outputs the predicted value Y (S _{t + 1} ) calculated by the recurrent neural network processing unit 103.

  Next, a processing flow of the information processing apparatus 100 configured as described above will be described. FIG. 4 is a flowchart of the information processing apparatus 100. The sensor value detected by the sensor 101 is acquired by the acquisition unit 102 (S101), and forward propagation processing is performed by the recurrent neural network processing unit 103 according to queue management by the queue management unit 102a (S102).

  The forward propagation value obtained by the forward propagation process is stored in a queue formed for each sampling rate in the memory 104 (S103).

The queue management unit 102a, is taken forward propagation value X _t has elapsed allowable delay time T from the memory 104, the forward propagation value X _t is forward propagation process by recurrent neural network processing unit 103, the state value S _t The predicted value Y (S _t ) is calculated (S104).

Then, based on the difference (loss) between the observed value y at time t and the predicted value Y (S _t ), the parameters w ₁ , w ₂ , and w _{3 of} the recurrent neural network are updated (S 105).

Meanwhile, the state value S _t at time t is backpropagation, the error of X _t is calculated (S106), that the error of the further X _t is backpropagation, and parameters w _f of recurrent neural network w _g is updated (S107). A learning process by a recurrent neural network is performed by the processes of S105 to S107. These processes are learning processes in a known recurrent neural network.

In S105 and S107, the parameters w ₁ , w ₂ , w ₃ , w _f , and w _g in the recurrent neural network are updated. In general, for these parameters, the loss due to the actually measured value and the predicted value in each observation is sufficiently reduced and converged, so that an accurate predicted value can be calculated. On the other hand, when these parameters are 0 or a value close to 0, it can be determined that the forward propagation value and the entire forward propagation value stored in the queue are not important. Therefore, in the process at time t, the determination unit (not shown) determines whether or not such a parameter is 0 or close to 0, thereby determining the sensor value x _{t at} time _t or the sampling rate thereof. You may make it judge importance.

Next, operational effects of the information processing apparatus 100 according to the present embodiment will be described. In the information processing apparatus 100 of the present embodiment, the acquisition unit 102 acquires a plurality of sensor values x _t detected at the sensor 101a and the like, recurrent neural network processing unit 103 sequentially propagates the acquired sensor value x _t The forward propagation value _Xt is calculated by processing. Queue manager 102a stores the calculated forward propagation value X _t, the memory 104 distinguished queue is formed at different sampling rates. The queue management unit 102a takes out a forward propagation value X _t stored in each queue of the memory 104, recurrent neural network processing unit 103 calculates the state value S _t indicating the status of the entire sensor value at time t To do. Then, the recurrent neural network processing unit 103 calculates a predicted value Y (S _t ) at time t + 1 based on the state value S _t .

  Thus, even when sensor values obtained at different sampling rates are obtained, the prediction processing can be performed collectively.

  The significance of acquiring sensor values obtained at different sampling rates is as follows. Even if the sensors serving as the information sources are the same, it is generally different from which sampling rate information is important. For example, when collecting human voice from a microphone, the finest granular information is simply waveform information (sound pressure information) around 1 kHz. The voice (phoneme) is not known from the waveform change itself at this level. Voice information is generated at a coarser granularity around 10 Hz. A word is generated with a granularity of about 1 Hz, which is a series of phonemes, and takes a few seconds per sentence at the sentence level. Similarly, the sampling rate of information to be captured varies depending on what information is important in sensor information and camera information.

In the information processing apparatus 100, the recurrent neural network processing unit 103 performs processing by the recurrent neural network using the calculated predicted value Y (S _t ), thereby updating the parameters of the recurrent neural network. Thereby, the learning process by a recurrent neural network can be performed, and an accurate predicted value can be calculated.

  Further, in the information processing apparatus 100, it is determined whether or not the forward propagation value stored in each queue of the memory 104 is important based on the updated parameter. As a result, it is possible to determine which queue is important and which first forward propagation value is important, and as a result, the sensor value obtained at which sampling rate is important. Can be judged.

  Here, a modified example will be described. In the above-described embodiment, the single processor explained the forward propagation process by the neural network and the forward propagation process by the recurrent neural network. For example, a processor arranged for each sensor or a processor for performing a recurrent neural network process may perform each process without using a single processor. More specifically, each processor as shown in FIG. 5 performs the following processing.

As shown in FIG. 5, to a processor p _S for processing the input sensor s, a processor p _T for processing the input sensor t, and there is a processor p _R for processing recurrent neural network. The processor p _S , the processor p _T , and the processor p _R have a function corresponding to the recurrent neural network processing unit 103 described above.

Processor p _S performs forward propagation process composed neural network from the input of the sensor s, and outputs the result r _S. Processor _{p S} is the result impart unique ID to _{r S,} is sent to the processor _{p R} with communication means results _{r S} along with its ID. Processor p _R holds put ID and string information of the network at this time. Processor p _T also performs the same processing. Processor p _R receives information from the processor p _T and the processor p _S, and stores the information in a queue in response to the occurrence time of the data in the same manner as described above, it performs the forward propagation process recurrent network. Incidentally, the time of occurrence of data is output from the processor p _T and the processor p _S.

Unique back propagation processing processor p _R is but begins by calculating the loss, When you reach the propagation to the point of receiving input from the processor p _S a processor p _T processor p _R is recommended backpropagation, received on input The gradient information for the information propagated along with the ID is transferred to the respective processors p _S and p _T. When each of the processors p _S and p _T receives the unique ID and the information for back propagation, the processors p _S and p _T perform back propagation using the corresponding forward propagation calculation graph and update the parameters (such as w described above). If this process is repeated, the information in the backpropagation calculation graph will gradually increase, and it will take time for backpropagation, so the backpropagation calculation will be based on the determined time or the determined maximum number. Discards the graph information.

  Further, the reaction rate can be adjusted by using a plurality of time series processes having different allowable delay times and discretization widths. The recurrent neural network that performs the determination by time series processing performs processing after a predetermined allowable delay time. For this reason, if it takes more time to sense data accidentally than this allowable delay time, or if it takes time for forward propagation processing, the information obtained here will not be used at all. Increasing the allowable delay time makes it easier to prevent this, but slows the reaction rate. If the discretization width is shortened, information with a high frequency can be handled, while the number of stages of the recurrent neural network becomes deep and it becomes difficult to learn long dependency relationships. Therefore, by preparing multiple different queues according to the allowable delay time and discretization width that are these parameters, the objective function is often used for each of reaction delay, information loss, information frequency, and long time dependence. You can select the parameters to explain.

DESCRIPTION OF SYMBOLS 100 ... Information processing apparatus, 101 ... Sensor, 102 ... Acquisition part, 103 ... Recurrent neural network processing part, 104 ... Memory, 105 ... Output part.

Claims (5)

Forward propagation processing means for calculating forward propagation values by performing forward propagation processing of sensor values detected at different sampling rates by a sensor;
Queue storage means for storing the forward propagation value calculated by the forward propagation processing means for each queue distinguished at the different sampling rates;
Queue management means for retrieving a forward propagation value at one time stored for each queue of the queue storage means;
An information processing apparatus comprising: a predicted value calculation unit that calculates a predicted value of a state at a time next to the one time based on the forward propagation value at the one time taken out by the queue management unit. The queue storage means includes
A forward propagation value calculated based on each sensor value detected at a different sampling rate by a plurality of sensors is stored for each queue distinguished for each sensor type and sampling rate. Item 4. The information processing apparatus according to Item 1. Update means for updating the parameters of the neural network by performing back propagation processing by the neural network using the predicted value calculated by the predicted value calculation means;
Determining means for determining whether or not the forward propagation value stored in the queue storage means is important based on the updated parameter;
The information processing apparatus according to claim 1, further comprising: In the information processing method of the information processing apparatus,
A forward propagation processing step of calculating a forward propagation value by performing forward propagation processing of sensor values detected at different sampling rates by the sensor;
A queue storage step of storing the forward propagation value calculated by the forward propagation processing step in a queue storage means for each queue distinguished at the different sampling rate;
A queue management step for extracting a forward propagation value at one time stored for each queue of the queue storage means;
An information processing method comprising: a predicted value calculation step of calculating a predicted value of a state at a time next to the one time based on the forward propagation value at the one time taken out by the queue management step. Computer
Forward propagation processing means for calculating forward propagation values by performing forward propagation processing of sensor values detected at different sampling rates by a sensor;
Queue storage means for storing the forward propagation value calculated by the forward propagation processing means for each queue distinguished at the different sampling rates;
Queue management means for retrieving a forward propagation value at one time stored for each queue of the queue storage means;
A predicted value calculating means for calculating a predicted value of a state at the next time of the one time based on the forward propagation value at the one time taken out by the queue managing means;
Information processing program to be executed as

Images