Runtime optimization of neural networks at quasi-static inputs

Abstract

The invention relates to a method for applying an artificial neural network. Checking whether the first activation and the second activation of the at least one neuron are zero, and whether the second network input of the at least one neuron is closer to a threshold than the first network input of the at least one neuron, and if so, determining whether the second network input of the at least one neuron is closer to the threshold; extrapolating a time until the network input of the at least one neuron reaches a threshold value in dependence on a difference between a first network input of the at least one neuron and the threshold value, in dependence on a difference between a second network input of the at least one neuron and the threshold value, and in dependence on a time elapsed between the first measurement and the second measurement, and checking whether the third measurement takes place before the expiration of the extrapolated time, and if so, outputting a value zero as a third activation of the at least one neuron without applying an activation function of the at least one neuron.

Description

Technical Field

[0001] The invention relates to a method according to the preamble of claim 1 and to a computer program according to claim 8 . Background Art

[0002] Neural networks are often used in vehicle control units for demanding control tasks. The computing power required is relatively high. As a result, the hardware of the control unit becomes more expensive. Summary of the invention

[0003] The object of the present invention is to design the use of artificial neural networks in a more resource-efficient manner. This object is achieved by a method according to claim 1 and a computer program according to claim 8. Preferred developments are contained in the dependent claims and are derived from the following description.

[0004] The method according to the invention is used for the application of artificial neural networks that have been trained before. The artificial neural network consists of interconnected artificial neurons. The artificial neuron has one or more input terminals and an output terminal. The input values ​​present at the input terminals are weighted. This is usually achieved by multiplying the values ​​by weights respectively. With the help of a transfer function, the network input is known from the input values ​​weighted in this way. A summation function is usually used as a transfer function, which calculates the sum of the weighted input values. An activation function is applied to the network input. The activation function assigns an activation to the network input, which is forwarded through the output terminal. The output terminal of the neuron can form the output terminal of the artificial neural network, or can be connected to the input terminal of another neuron. The input terminal of the neural network is formed by the input terminal of the neuron. The neural network receives the input signal through the input terminal. Correspondingly, the output signal of the neural network is output through the output terminal.

[0005] The method according to the invention presupposes that the activation function of at least one neuron assigns the value zero to the network input of the neuron as a function of a threshold value. The activation function thus assumes the value zero for all network inputs below the threshold value or, alternatively, for all network inputs above the threshold value.

[0006] When implementing the method according to the invention, three iterations are performed, each of which measures one or more physical variables. Thus, a first measurement, a second measurement and a third measurement of one or more physical variables are performed.

[0007] A set of one or more input signals of the network is determined by measurement. A first set is determined by a first measurement, a second set is determined by a second measurement, and a third set is determined by a third measurement. The input signals are hereby respectively measured directly or calculated from one or more measured physical variables.

[0008] By means of a neural network, one or more output signals are respectively known in dependence on one or more input signals. A first set of one or more output signals is known in dependence on a first set of one or more input signals, a second set of one or more output signals is known in dependence on a second set of one or more input signals, and a third set of one or more output signals is known in dependence on a third set of one or more input signals.

[0009] The determination of the output signal in dependence on the first and second sets of one or more input signals is implemented in a conventional way. This means that the activation of each neuron of the network is determined by applying the respective activation function to the respective network input calculated by means of the respective transfer function. Therefore, the first activation of the neurons of the network is determined by applying the respective activation function to the respective first network input calculated by means of the respective transfer function, thereby determining the first set of one or more output signals. Similarly, the second activation of the neurons of the network is determined by applying the respective activation function to the respective second network input calculated by means of the respective transfer function, thereby determining the second set of one or more output signals. Therefore, for each neuron of the network, the first or second network input of the neuron is calculated by applying the transfer function of the neuron. The first or second network input of the neuron is passed to the transfer function of the neuron as a function value. The activation function assigns an activation to the first or second network input as the first or second output signal.

[0010] The activation function, like the transfer function, can be different for individual or all neurons of the network. Preferably, however, only one unique activation function is used that is the same for each neuron.

[0011] Depending on the one or more input signals, a third set of one or more output signals is also known based on the activation of the neurons of the network.Thus, a third activation of the neurons of the network is known respectively, thereby knowing a third set of one or more output signals depending on the third set of one or more input signals.

[0012] However, the invention provides for a different determination of the activation of the at least one neuron whose activation function assumes the value zero in dependence on a threshold value. The concept is based on the fact that the calculation of the activation function is omitted if it is foreseeable that the activation function assumes the value zero.

[0013] For this purpose, it is checked whether the first activation and the second activation of the at least one neuron are zero. In more detail, the first and second network inputs of the at least one neuron are checked with respect to a threshold value. It is checked whether the first and second network inputs of the at least one neuron are below the threshold value or above the threshold value, so that the activation function of the at least one neuron assigns the value zero to the first and second network inputs.

[0014] Additionally, it is checked whether the second network input of at least one neuron is closer to the threshold than the first network input. Thus, it is checked whether the absolute value of the difference between the second network input and the threshold is smaller than the absolute value of the difference between the first network input and the threshold.

[0015] If the first and second network inputs of at least one neuron are below or above a threshold and the second network input of at least one neuron is closer to the threshold than the first network input, this indicates that the network input of at least one neuron is close to the threshold and exceeds or falls below the threshold, so that the activation function of at least one neuron takes a value different from zero. The time until then is extrapolated. Thus, if the previously performed check has shown that the first and second network inputs of at least one neuron are below or above a threshold, so that the first activation and the second activation of at least one neuron are zero, and the second network input of at least one neuron is closer to the threshold than the first network input of at least one neuron, then the time until the network input of at least one neuron reaches the threshold is extrapolated.

[0016] The time until the network input of at least one neuron reaches the threshold starts at the time of the second measurement and ends as soon as the difference between the first network input and the second network input of at least one neuron reaches the threshold. The time depends on the difference between the first network input of at least one neuron and the threshold, the difference between the second network input of at least one neuron and the threshold and the time elapsed between the first measurement and the second measurement. This is taken into account when extrapolating the time.

[0017] Furthermore, it is checked whether the third measurement is performed before or after the expiration of the extrapolated time. If the third measurement is performed before the expiration of the extrapolated time, this indicates that the activation function of the at least one neuron assigns the value zero to the third network input of the at least one neuron. If so, the calculation of the weight of the at least one neuron, the application of the transfer function of the at least one neuron and the application of the activation function of the at least one neuron can be omitted to determine the third activation of the at least one neuron. Alternatively, the value zero is output as the third activation of the at least one neuron without applying the activation function of the at least one neuron.

[0018] Otherwise, a third activation of the at least one neuron is determined in a conventional manner. Thus, if the third measurement is performed after the expiration of the interpolated time and / or if the first activation and / or the second activation of the at least one neuron is different from zero and / or if the second network input of the at least one neuron is not closer to the threshold value than the first network input of the at least one neuron, a third activation of the at least one neuron is determined in a conventional manner by applying the activation function of the at least one neuron to the third network input calculated with the aid of the transfer function of the at least one neuron.

[0019] The method according to the invention is particularly advantageous when one or more of the measured physical variables changes only slowly. This is true for a plurality of physical variables measured in a vehicle. In the case of slowly changing variables, the risk of errors due to extrapolation is relatively low. It is also possible to extrapolate over longer time periods. By omitting the use of an activation function, a considerable amount of computing time can be saved.

[0020] In a preferred refinement, it is checked whether the first activation and the second activation of at least one neuron are zero, whether the first and second network inputs of at least one neuron are below a threshold or above a threshold, and whether the second network input of at least one neuron is further away from the threshold than the first network input of at least one neuron, i.e. whether the absolute value of the difference between the second network input and the threshold is greater than the absolute value of the difference between the first network input and the threshold. This indicates that the network input of at least one neuron is not close to the threshold, so that the activation of at least one neuron continues to take the value zero for a foreseeable time.

[0021] According to a refinement, in this context, a predefined time, i.e. a time specified before the implementation of the method, is used, until the expiration of which the transfer function of the at least one neuron is not used. Therefore, it is checked whether the third measurement was performed before the expiration of the predefined time. If the third measurement was performed before the expiration of the predefined time, then according to a refinement, the value zero is selected as the third activation of the at least one neuron without applying the activation function of the at least one neuron. After the expiration of the predefined time, the activation of the at least one neuron is preferably determined in a conventional manner by applying the activation function of the at least one neuron to the network input calculated with the aid of the transfer function of the at least one neuron.

[0022] In a preferred refinement, the time until the network input of the at least one neuron reaches the threshold value is determined by linear extrapolation. In particular, the time can be determined as the product of the time elapsed between the first measurement and the second measurement time, the difference between the second network input and the threshold value, and the reciprocal of the difference between the first network input and the second network input of the at least one neuron. Determining the time by linear extrapolation according to the refinement is particularly simple and fast to implement.

[0023] In a preferred refinement, one or more of the input signals of the network represent physical variables of the vehicle. For example, variables such as temperature, force, torque, speed, position and engaged gear or transmission ratio come into consideration. Preferably, these are physical variables of the transmission and / or the motor of the vehicle.

[0024] In a preferred improvement, the input signals of the network are provided at least in part by sensors of the vehicle. Thus, according to the improvement, one or more input signals of the network are signals of sensors of the vehicle. Preferably, the sensors are arranged on the transmission and / or the motor of the vehicle. In particular, the above-mentioned physical variables can be measured by means of sensors.

[0025] Similarly, in a preferred refinement, one or more of the output signals of the network represent physical variables of the vehicle, such as temperature, force, torque, position, speed, rotational speed and / or an engaged gear or transmission ratio. If the output signal represents an engaged gear or transmission ratio, a transmission control can be implemented with the aid of the method according to the invention.

[0026] The actuator can also be controlled by the neural network. In a corresponding preferred improvement, one or more of the output signals of the network are control signals for the actuator of the vehicle. In particular, the gear adjustment device of a transmission capable of automatic gear shifting can be controlled.

[0027] The computer program according to the invention implements the method according to the invention or a preferred improvement of the method. Therefore, the computer program includes instructions for executing the method according to the invention or a preferred improvement. This means that when the computer program is executed by a computer, the computer program causes the computer to execute the method according to the invention or a preferred improvement. The computer program can be stored on a storage medium or encoded in one or more transmittable or transmitted signals. In particular, the computer program can exist as a computer program product, i.e. as a tradable unit or a unit for transferring ownership of the computer program. The computer mentioned is preferably a vehicle controller, such as a motor controller or a transmission controller.

Claims

1. A method for applying an artificial neural network; wherein: The activation function of at least one neuron assigns the value zero to the network input of the neuron in a threshold-dependent manner; wherein, A first set of one or more input signals of the network is determined by first measuring one or more physical parameters; wherein, A first activation of the neurons of the network is respectively known by applying a respective activation function to a respective first network input calculated by means of a respective transfer function, thereby knowing a first set of one or more output signals in dependence on the first set of one or more input signals; wherein, A second set of the one or more input signals is determined by a second measurement of the one or more physical parameters; wherein, A second activation of the neurons of the network is respectively known by applying the respective activation function to the respective second network input calculated by means of the respective transfer function, thereby knowing the second set of the one or more output signals in dependence on the second set of the one or more input signals; wherein, A third set of the one or more input signals is determined by a third measurement of the one or more physical quantities; and wherein, A third activation of neurons of the network is respectively known, whereby a third set of the one or more output signals is known depending on the third set of the one or more input signals; characterized in that checking whether the first activation and the second activation of the at least one neuron are zero and whether the second network input of the at least one neuron is closer to a threshold than the first network input of the at least one neuron, and if so, extrapolating the time until the network input of the at least one neuron reaches the threshold value dependent on a difference between a first network input of the at least one neuron and the threshold value, dependent on a difference between a second network input of the at least one neuron and the threshold value and dependent on a time elapsed between the first measurement and the second measurement, and checking whether the third measurement is made before the expiration of the extrapolated time, and if so, A value of zero is output as a third activation of the at least one neuron without applying an activation function of the at least one neuron.

2. The method according to the preceding claim, characterized in that checking whether the first activation and the second activation of the at least one neuron are zero, whether the second network input of the at least one neuron is further away from the threshold than the first network input of the at least one neuron and whether the third measurement is performed before the expiration of a predefined time, and if so, A value of zero is selected as a third activation of the at least one neuron without applying an activation function of the at least one neuron.

3. The method according to any one of the preceding claims, characterized in that The time until the network input of the at least one neuron reaches the threshold is found by linear extrapolation.

4. The method according to any one of the preceding claims, characterized in that One or more of the input signals of the network represent values ​​of physical variables of the vehicle.

5. The method according to any one of the preceding claims, characterized in that One or more of the input signals of the network are signals of sensors of / of the vehicle.

6. The method according to any one of the preceding claims, characterized in that One or more of the output signals represent a value of a physical variable of / of the vehicle.

7. The method according to any one of the preceding claims, characterized in that One or more of the output signals are control signals for actuators of the vehicle / the vehicle.

8. Computer program having instructions for executing the method according to any of the preceding claims.