A robot emotion expression method based on neural network emotion model

Abstract

The invention discloses a robot emotion expression method based on a neural network emotion model, which belongs to the field of robot and computer technology, and comprises the following steps: step S1, constructing an emotion pre-injection data set of an available sensor of a target robot and an input coding layer I1, an activation layer L1, and an inhibition layer L2 of a pulse neural network, and establishing connections and feedback connections between the layers; step S2, performing emotion injection pre-training for discrete environmental input on the initial pulse neural network, so that the pulse neural network can obtain a set emotion output for the discrete environmental input; step S3, according to the real-time interaction result between the robot and a person, using an STDP algorithm to continuously online adjust the connection weight of a designated connection layer of the pulse neural network, and positively strengthening a memory buffer pool corresponding to the interaction result and negatively attenuating a memory buffer pool opposite to the interaction result, thereby achieving online adjustment from emotion pre-injection to real interaction.

Description

Technical Field

[0001] The present invention belongs to the field of robot and computer technology, and in particular relates to a robot emotion expression method based on a neural network emotion model. Background Art

[0002] Neural networks are widely used in emotion recognition, emotion classification and emotion synthesis tasks in robot emotion models. However, traditional robot emotion models often have a fixed design structure, which leads to relatively rigid emotion expression. Although the robot emotion model based on neural network has improved the flexibility of robot emotion expression, due to the possible bias and sample imbalance of training data, the robot emotion model still faces certain difficulties and challenges in its deployment and actual use in real environments.

[0003] The spiking neural network is a new generation of artificial neural network model inspired by biology, which has a high degree of biological interpretability. In traditional artificial neural networks, the calculation of each layer of neurons is expressed as a specific nonlinear activation function. The spiking neural network uses a pulse sequence to represent information. Neurons perform a computational paradigm similar to that of biological models. After the calculation is completed, a series of pulses are emitted through the axon to different successor neurons at a specific time. This highly approximate encoding method at the biological level also includes multiple characteristics such as frequency, time, space, and phase. Therefore, the emotion model built using the spiking neural network is not only more advantageous in terms of biological interpretability, but also compared with the huge computational overhead brought by the traditional back-propagation algorithm, the STDP (Spike-Timing-Dependent Plasticity) learning algorithm based on the spiking neural network omits the calculation and back-propagation of the gradient, thereby greatly reducing the calculation amount of the model and thus reducing the computing energy consumption. On the other hand, the existing emotion model still cannot be adjusted online according to the real interaction results between robots and humans. Summary of the invention

[0004] In order to solve the shortcomings of the prior art, realize the flexibility of the robot emotion model expression, continuously adjust the existing emotion expression mode according to the interaction results between the robot and the human, and achieve the purpose of continuous emotion reinforcement and emotion transformation, the present invention adopts the following technical solutions:

[0005] A robot emotion expression method based on a neural network emotion model comprises the following steps:

[0006] Step S1: construct an emotion injection pre-training data set, and expand the data dimension in a targeted manner for data of different dimensions in a set of data sets; construct a neural network emotion model connection layer, including an input encoding layer, an activation layer, and an inhibition layer connected in sequence, and the inhibition layer is connected to the activation layer through lateral inhibition feedback;

[0007] Step S2: Train the neural network emotion model by injecting the emotion pre-training data set, and use the memory buffer pool to collect the emotion decoding outputs of the activation layer corresponding to different emotions. Each emotion corresponds to a buffer list in the memory buffer pool. During emotion injection and training, any expanded data input will correspond to a decoding output and an emotion label, and then the decoding output is placed in the buffer list corresponding to the emotion label. For any neuron in the activation layer of the decoding layer, the average pulse number of the neuron in different buffer lists is calculated, and the emotion label corresponding to the buffer list with the maximum average pulse number is selected as the emotion label of the neuron;

[0008] Step S3: According to the real-time interaction results between the robot and the outside world, the STDP algorithm is used to continuously adjust the connection weights of the specified connection layer of the pulse neural network online, and positively strengthen the memory buffer pool corresponding to the interaction results, and negatively attenuate the memory buffer pool opposite to the interaction results, so as to achieve online adjustment from emotion pre-injection to real interaction.

[0009] Furthermore, compared with the gradient directional propagation method of artificial neural networks, the emotion model uses the STDP algorithm unique to the pulse neural network to update the network weights. The update mechanism of the STDP algorithm is that the pulse sequence of adjacent neurons causes the connection weights to be strengthened or attenuated. While being more in line with the real characteristics of biological neurons, it also introduces a layer of randomness. Therefore, the first purpose of the data dimension expansion in step S1 is to use multiple input neurons to represent the emotion injection obtained by the same sensor, thereby compensating for the randomness brought by the introduction of the STDP algorithm, and thereby increasing the ability of the middle layer neurons to extract the abstract features of the input data; the second purpose of expanding the input is to introduce the priority set initially; for example Compared with the color of clothes, the expansion is injected The middle layer neurons corresponding to the emotional labels are more likely to trigger pulses, which, combined with lateral inhibition, makes the unexpanded The middle-layer neurons corresponding to the emotional labels are more difficult to emit pulses, which introduces artificial priority, expanding the more important and reliable sensor inputs into larger dimensions, while the less important and less reliable sensor inputs are expanded into smaller dimensions or not expanded at all.

[0010] Furthermore, in step S1, the activation layer L1 and the inhibition layer L2 are connected one-to-one, the connection weight is a positive number greater than one, and the weight does not participate in the update; the inhibition layer L2 and the activation layer L1 are connected in a lateral inhibition feedback connection, the diagonal of the connection weight matrix is ​​all 0, and the remaining weights are all less than negative one, and the weights do not participate in the update.

[0011] Furthermore, in step S2, the expanded input data is divided into different groups according to different predetermined emotions. In the same group, the expanded input data belongs to different sub-groups of the same group. Sub-groups are randomly selected to construct emotion pre-injected discrete data with the selected sub-group as the main input. In the same group, data not in the same sub-group is used as secondary input, and the remaining data is used as noise input, so that the data of the same emotion label group is discretized, thereby making it easier for the pulse neural network to capture the abstract features of the discrete sub-groups; the emotion labels corresponding to different groups are pre-specified, and the STDP algorithm is used in conjunction with the lateral inhibition mechanism to train the pulse neural network activation layer, and then the STDP algorithm is used to continuously pre-train the emotion model of the expanded input data and the predetermined emotion label, so that the emotion output of the emotion model for the expanded input data has a better predictive match with the predetermined emotion label.

[0012] Furthermore, in step S3, the pulse neural network continuously updates the connection weights of the designated layer using the STDP algorithm to memorize the abstract features of the input data for a long time; the lateral inhibition mechanism makes the neurons in the activation layer of the emotion model have competing pulse characteristics, and the earlier the neuron pulses in the activation layer, the more difficult it will be for other neurons to pulse. Therefore, the feature extraction of different emotion input data by the activation layer is a preemptive process, which will be more conducive to different neurons capturing different input features, increasing the diversity of input neurons in feature extraction, and thus expanding the input level, which is conducive to long-term memory.

[0013] Furthermore, in the step S3, the emotion label corresponding to the decoded output of the pulse neural network at the previous moment is positively reinforced, or the emotion label opposite to the decoded output is negatively suppressed, so as to achieve short-term memory and online adjustment of the abstract features of the input data; the emotion label of the robot's interaction with the outside world is pre-defined, and for the interaction in which the current input data is positive emotion, and the number of negative emotion pulses output by the decoded neurons at the previous moment is greater than the number of positive emotion pulses, then the output of the emotion model at the previous moment is negative emotion, which is inconsistent with the emotion label of the interaction, then the decoded output at the previous moment is placed in the buffer list of positive emotion, and the decoded output is inverted and placed in the buffer list corresponding to the negative emotion at the previous moment; if the output of the emotion model at the previous moment is positive emotion, which is consistent with the emotion label of the interaction result, then it is only necessary to place the decoded output at the previous moment in the buffer list of positive emotion.

[0014] Furthermore, in step S3, the intensity scalar signal of the interaction (such as the strength of the touch signal) is extracted and passed to the STDP algorithm as a reward signal r, thereby affecting the pulse neural network to obtain different degrees of weight update for different input features, thereby achieving a direct connection between the human-computer interaction intensity signal and the pulse neural network weight memory. The specific formula is as follows:

[0015]

[0016] in, represents the connection weight between the input encoding layer and the activation layer L1 of the sentiment model, It represents the updated change of the connection weight given by the STDP algorithm, and i and j represent the horizontal and vertical coordinates of the weight matrix respectively.

[0017] Furthermore, in the step S3, the emotion labels corresponding to the activation layer neurons are recalculated and assigned; since the contents of different emotion buffer lists are updated, the average pulse times of different neurons in different emotion lists also need to be updated, and then the emotion labels corresponding to the activation layer neurons are also required to be recalculated in the step S2, and the buffer list is updated. For any decoding output that matches or does not match the emotion model, the above-mentioned adjustment method will continuously adjust the correct emotion buffer list and the mismatched emotion buffer list, thereby realizing the continuous online learning of the robot emotion model emotion pre-injection into the real interaction process.

[0018] Furthermore, the emotion model uses the pulse timing dependent plasticity STDP algorithm of the pulse neural network to update the network weights

[0019] Furthermore, both the activation layer and the inhibition layer use LIF neurons.

[0020] The advantages and beneficial effects of the present invention are:

[0021] The present invention provides a robot emotion expression method based on a neural network emotion model, which realizes that the output of the robot emotion model is highly flexible for inputs of different modalities. The emotion model can also continuously adjust the emotion output of the model according to the interaction results between the robot and the human, and realizes long-term and short-term memory of external environment characteristics through a pulse neural network and a memory buffer pool, thereby realizing a continuous and natural robot emotion reinforcement and emotion transformation process, thereby enriching the interaction mode between the robot and the human and enhancing the interaction experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of a method in an embodiment of the present invention.

[0023] Figure 2It is an overall flow chart of the emotion model in the embodiment of the present invention.

[0024] Figure 3 It is a flow chart of online adjustment of the emotion model memory buffer pool in an embodiment of the present invention.

[0025] Figure 4 It is a schematic diagram of the operation logic of the emotion buffer pool in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0027] like Figure 1 As shown, this example proposes a robot emotion expression method based on a neural network emotion model, including the following steps:

[0028] Step S1: construct a sentiment injection pre-training data set and expand the data dimension; construct a model network connection and feedback connection, specifically including the following steps:

[0029] Step S1.1: Construct an emotion injection pre-training dataset and expand the data dimension; specifically, the input of different sensors is represented as: In the embodiment of the present invention, The gesture input for the interactive object is then used to extract the user's body language. Input the clothing color of the interactive object to mark different users. is the external alcohol concentration sensor input. To make the example simple, the input is set to only 0 and 1 discrete inputs, such as [0,0,0] represents [thumbs down gesture, red clothes, no alcohol input signal], [1,1,1] represents [thumbs up gesture, blue clothes, alcohol input signal]; the input representation of the sentiment model after selecting the input to be expanded in two different forms can be written as:

[0030]

[0031] There are two purposes for expanding the input. First, compared with the gradient directional propagation method of artificial neural networks, the sentiment model uses the STDP algorithm unique to the pulse neural network to update the network weights. The update mechanism of the STDP algorithm: "The pulse sequence of adjacent neurons causes the connection weight to be strengthened or attenuated" is more in line with the real characteristics of biological neurons, while also introducing a layer of randomness. Therefore, the first purpose of using input expansion in the sentiment model is to use multiple input neurons to represent the same sensor input, thereby compensating for the randomness brought by the introduction of the STDP algorithm, and thereby increasing the ability of the middle layer neurons to extract abstract features of the input data.

[0032] The second purpose of expanding the input is to introduce the initial set of priorities. Compared with the color of clothes, the expanded input The middle layer neurons corresponding to the emotional label are more likely to trigger pulses, which, combined with lateral inhibition, leads to The middle layer neurons corresponding to the emotional labels are more difficult to emit pulses. Then artificial priorities are introduced, and more important and reliable sensor inputs are expanded in larger dimensions, while less important and less reliable sensor inputs are expanded in smaller dimensions or not expanded at all.

[0033] Step S1.2: construct the emotion pre-injection dataset of the target robot’s available sensors and the input encoding layer I1, activation layer L1, inhibition layer L2 of the spiking neural network, and establish connections and feedback connections between the layers;

[0034] Specifically, the activation layer L1 is set to have the same dimension as the suppression layer L2, the input coding layer I1 and the activation layer L1 are fully connected, and the connection weights are updated using the STDP algorithm;

[0035] The activation layer L1 and the inhibition layer L2 are connected one to one, and the connection weight is a positive number greater than one. The weight does not participate in the update; when the dimension is equal to 3, the connection weight is as follows:

[0036]

[0037] The feedback connection between the inhibition layer L2 and the activation layer L1 is a lateral inhibition connection, that is, the diagonal of the weight matrix is ​​all 0, and the remaining weights are all less than negative one, and the weights do not participate in the update; when the dimension is equal to 3, the connection weights are as follows:

[0038]

[0039] Both the activation layer L1 and the inhibition layer L2 use LIF neurons; the overall structure of the emotion model is as follows Figure 2As shown, dropout means randomly discarding some connections with the activation layer neurons, thereby weakening the strength of lateral inhibition and providing opportunities for activation layer pulses of other characteristic patterns. The 256 of the encoding input layer represents the input dimension actually used in the embodiment of the present invention, and the 100 of the activation layer and the inhibition layer represent that the dimensions of the activation layer L1 and the inhibition layer L2 are the same, both 100.

[0040] Step S2: Based on the construction of the emotion model connection and the emotion pre-injection dataset, the STDP algorithm is used to train the specified connection layer, the output encoding is saved to the memory buffer pool, and then the emotion labels of the decoding layer neurons are counted.

[0041] Step S2.1: Pre-training the initial spiking neural network for emotion injection for discrete environmental inputs, so that the spiking neural network can obtain a set emotion output for the discrete environmental inputs;

[0042] Specifically, the expanded input data is divided into different groups T according to different predetermined emotions. When the robot is positive, When the robot is in a negative mood, the specific grouping method of the emotion data set should be set according to the actual situation. For example, the expanded data input , Belong to the same group Then randomly select subgroups to As an example, construct the emotion pre-injected discrete data with the selected sub-group as the main input, the data of the same group T but not in the same sub-group as the secondary input, and the rest of the data as the noise input, so that the data of the same emotion label group is discretized, and then the pulse neural network can more easily capture the abstract features of the discrete sub-group. Pre-specify the emotion labels corresponding to different groups, and use the STDP algorithm with the lateral inhibition mechanism to train the connection layer L1 of the pulse neural network. Therefore, you can select As the main input, As a secondary input, As the noise input, the expanded input data is:

[0043]

[0044] Then, the STDP algorithm is used to continuously pre-train the sentiment model of the expanded input data and the predetermined sentiment label, so that the sentiment output of the sentiment model for the expanded input data has a better prediction match with the predetermined sentiment label.

[0045] Step S2.2: While the emotion model is pre-injected, the emotion buffer pools corresponding to different emotion labels are collected, and the average number of pulses of each neuron is counted, and the emotion labels corresponding to the neurons in the L1 layer are assigned.

[0046] Specifically, use the memory buffer pool B to collect the most recent The L1 layer of the emotion model at each moment decodes the emotion output. Corresponding to a buffer list in memory buffer pool B During emotion injection and training, any expanded data input will correspond to a pulse decoding output and an emotion label, and then the pulse decoding output will be placed in the buffer list corresponding to the emotion label. The emotional output method is: for any neuron in the decoding layer, calculate different buffer lists The average number of pulses of the neuron in the buffer list, and select the maximum average number of pulses The corresponding emotion label is used as the emotion label of the neuron. In the embodiment of the present invention, the dimension of the decoding neuron is , so if the buffer list corresponding to positive sentiment The average number of the four neurons is , the buffer list corresponding to negative emotions The average number of the four neurons is , then neurons 0 and 1 are active neurons; neurons 2 and 3 are passive neurons.

[0047] Step S3: Based on the emotion model in step S2 and the specific emotion labels assigned to each neuron, the STDP algorithm is used to continuously adjust the connection weights of the designated connection layer of the spiking neural network online according to the real-time interaction results between the robot and the human, and the memory buffer pool corresponding to the interaction results is positively strengthened and the memory buffer pool opposite to the interaction results is negatively attenuated, thereby achieving online adjustment from emotion pre-injection to real interaction, such as Figure 3 As shown, the following steps are included:

[0048] Step S3.1: The connection weights of the designated layer of the spiking neural network are continuously updated using the STDP algorithm to achieve long-term memory of the abstract features of the input data.

[0049] The introduction of the lateral inhibition mechanism makes the neurons in the activation layer L1 of the emotion model have the characteristics of competing pulses. The earlier the neuron pulses in the activation layer L1, the more difficult it is for other neurons to pulse. Therefore, the feature extraction of different sensor input data by the activation layer is often a preemptive process. In addition, since the feedback connection between the activation layer L1 and the inhibition layer L2 still requires a time step to transmit lateral inhibition, the preemption process in the activation layer L1 still leaves a time step of vacancy;

[0050] Step S3.2: According to the interaction results between the robot and the human, the robot's emotional output is adjusted online.

[0051] The respective sentiment buffer pools for the case of sentiment labels with only positive and negative sentiment, such as Figure 4 As shown in the figure, the main operation and design idea of ​​the emotion buffer pool is to achieve short-term memory and online regulation of the abstract features of the input data by positively reinforcing the emotion label corresponding to the decoding output of the pulse neural network at the previous moment, or negatively inhibiting the emotion label opposite to the decoding output.

[0052] Specifically, the emotional labels of the robot-human interaction are predefined. Take the human touch input in the interaction between the quadruped robot and the human at the current moment as an example. The touch input is defined as the interaction of positive emotions. Assume that the output of the neuron in the decoding layer at the previous moment is , and at the last moment, neurons 0 and 1 were positive neurons, and neurons 2 and 3 were negative neurons; the total number of negative emotion pulses was 7, which was greater than the total number of positive emotion pulses 3, so the output emotion of the emotion model at the last moment was negative emotion, which was inconsistent with the emotion label of the interaction result. Therefore, it is necessary to put the decoded output at the last moment in the buffer list of positive emotion, and invert the decoded output and put it in the buffer list corresponding to the negative emotion at the last moment; if the output emotion of the emotion model at the last moment was positive emotion, which was consistent with the emotion label of the interaction result, then it is only necessary to put the decoded output at the last moment in the buffer list of positive emotion.

[0053] On the other hand, the extracted interaction intensity scalar signal (such as the strength of the touch signal) is used as the reward signal r to pass to the STDP algorithm, which in turn affects the pulse neural network to obtain different degrees of weight updates for different input features, thereby achieving a direct connection between the human-computer interaction intensity signal and the pulse neural network weight memory. The specific formula is as follows:

[0054]

[0055] in, represents the connection weight between the input encoding layer and the activation layer L1 of the sentiment model, It represents the updated change of the connection weight given by the STDP algorithm, and i and j represent the horizontal and vertical coordinates of the weight matrix respectively.

[0056] Step S3.3: Re-count and assign the emotion labels corresponding to the neurons in the L1 layer.

[0057] Because the contents of different emotion buffer lists are updated, the average pulse counts of different neurons in different emotion lists also need to be updated, and then the emotion labels corresponding to the neurons in the L1 layer also need to be re-counted and updated in the same way as in the S2.2 stage. Therefore, for any decoding output that matches or does not match the emotion model, the above-mentioned adjustment method will continuously adjust the correct emotion buffer list and the mismatched emotion buffer list, thereby realizing the continuous online learning of the robot emotion model pre-injection into the real interaction process.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A robot emotion expression method based on a neural network emotion model, characterized in that The steps include: Step S1: construct an emotion injection pre-training data set, and expand the data dimension in a targeted manner for data of different dimensions in a set of data sets; construct a neural network emotion model connection layer, including an input encoding layer, an activation layer, and an inhibition layer connected in sequence, and the inhibition layer is connected to the activation layer through lateral inhibition feedback; Step S2: Train the neural network emotion model by injecting the emotion pre-training data set, and use the memory buffer pool to collect the emotion decoding outputs of the activation layer corresponding to different emotions. Each emotion corresponds to a buffer list in the memory buffer pool. During emotion injection and training, any expanded data input will correspond to a decoding output and an emotion label, and then the decoding output is placed in the buffer list corresponding to the emotion label. For any neuron in the activation layer, the average pulse number of the neuron in different buffer lists is calculated, and the emotion label corresponding to the buffer list with the maximum average pulse number is selected as the emotion label of the neuron; Step S3: According to the real-time interaction results between the robot and the outside world, the connection weights of the designated connection layer of the neural network are continuously adjusted online, and the memory buffer pool corresponding to the interaction results is positively strengthened, and the memory buffer pool opposite to the interaction results is negatively attenuated.

2. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: The data dimension expansion in step S1 uses multiple input neurons to represent the same emotion injection. After the expansion, the middle layer neurons corresponding to the injected emotion label are more likely to trigger pulses, which, when combined with lateral inhibition, makes it more difficult for the middle layer neurons corresponding to the unexpanded emotion label to emit pulses.

3. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: In the step S1, the activation layer and the inhibition layer are connected one-to-one, the connection weight is greater than one, and the weight does not participate in the update; the inhibition layer and the activation layer are connected by lateral inhibition feedback, the diagonal of the connection weight matrix is ​​all 0, and the remaining weights are all less than negative one, and the weight does not participate in the update.

4. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: In the step S2, the expanded input data is divided into different groups according to different predetermined emotions. In the same group, the expanded input data belongs to different sub-groups of the same group. Sub-groups are randomly selected to construct emotion pre-injected discrete data with the selected sub-group as the main input. In the same group, data not in the same sub-group is used as a secondary input, and the remaining data is used as a noise input; emotion labels corresponding to different groups are pre-specified, and the neural network activation layer is trained in conjunction with the lateral inhibition mechanism.

5. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: In step S3, the neural network continuously updates the connection weights of the designated layer to memorize the abstract features of the input data for a long time; the lateral inhibition mechanism makes the neurons in the activation layer of the emotion model have competing pulse characteristics, and the earlier the neuron pulses in the activation layer, the more difficult it will be for other neurons to pulse.

6. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: In the step S3, the emotion label of the robot's interaction with the outside world is predefined. For the interaction of the current input data with positive emotion, if the number of negative emotion pulses output by the neuron decoded at the previous moment is greater than the number of positive emotion pulses, then the output of the emotion model at the previous moment is negative emotion, which is inconsistent with the emotion label of the interaction. Then, the decoded output at the previous moment is placed in the buffer list of positive emotion, and the decoded output is inverted and placed in the buffer list corresponding to the negative emotion at the previous moment. If the emotion model output at the previous moment is positive emotion, which is consistent with the emotion label of the interaction result, then it is only necessary to put the decoding output at the previous moment into the buffer list of positive emotion.

7. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: In step S3, the interaction intensity scalar signal is extracted as a reward signal, thereby influencing the pulse neural network to obtain different degrees of weight update for different input features.

8. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: In step S3, the emotion labels corresponding to the activation layer neurons are recalculated and assigned; since the contents of different emotion buffer lists are updated, the average pulse times of different neurons in different emotion lists also need to be updated, and thus the emotion labels corresponding to the activation layer neurons also need to be recalculated in step S2 and the buffer lists are updated.

9. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: The emotion model uses a pulse timing-dependent plasticity algorithm of a pulse neural network to update network weights.

10. The robot emotion expression method based on a neural network emotion model according to claim 1, characterized in that: Both the activation layer and the inhibition layer use LIF neurons.

Images