A neural signal encoding method, device, equipment and readable storage medium
By performing wavelet transform and joint mutual information calculation on neural signals, and selecting and removing time-frequency features, the problems of ignoring higher-order dependence and redundancy of wavelet coefficients in existing technologies are solved, thus achieving more efficient neural signal encoding.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN READLINE BIOTECH CO LTD
- Filing Date
- 2023-08-07
- Publication Date
- 2026-04-21
AI Technical Summary
Existing techniques neglect higher-order dependencies between wavelet coefficients and potential redundancy of features in neural signal coding, resulting in insufficient effectiveness and accuracy of the coding results.
By quantizing neural signals into target matrices and performing wavelet transforms, the mutual information of time-frequency features is calculated in conjunction with the current target set. The largest time-frequency feature is selected as the target feature, and feature pruning is performed when the conditions are met to optimize the encoding process.
It improves the effectiveness and accuracy of neural signal encoding results, enabling more precise characterization of information in neural signals, reducing redundant features, and enhancing encoding performance.
Smart Images
Figure CN117009747B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of signal processing technology, and in particular to a neural signal encoding method, apparatus, device, and readable storage medium. Background Technology
[0002] Currently, wavelet analysis of neural impulse signals can capture detailed temporal dynamics of signal time sequence information. However, when selecting wavelet coefficients (i.e. signal features) from the signal time sequence information, the calculation of each wavelet coefficient is independent, ignoring the higher-order dependencies between wavelet coefficients and not considering the potential redundancy of features. This approach can only extract and encode signals in time sequence, making it difficult to guarantee the effectiveness and accuracy of the signal encoding results.
[0003] Therefore, how to improve the effectiveness and accuracy of neural signal encoding results is a problem that needs to be solved by those skilled in the art. Summary of the Invention
[0004] In view of this, the purpose of this application is to provide a neural signal encoding method, apparatus, device, and readable storage medium to improve the effectiveness and accuracy of neural signal encoding results. The specific solution is as follows:
[0005] In a first aspect, this application provides a neural signal encoding method, including:
[0006] Quantize neural signals into a target matrix;
[0007] Perform wavelet transform on the target matrix to obtain the time-frequency feature matrix;
[0008] Initialize the target set to an empty set and the information threshold to an initial value;
[0009] Calculate the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories by combining the current target set;
[0010] Among the time-frequency features with the largest first information content greater than the target value, the time-frequency feature with the largest first information content is selected as the target feature.
[0011] The target features are populated into the current target set, and the target features are deleted from the current time-frequency feature matrix;
[0012] Update the information threshold. If the updated information threshold meets the termination condition, perform feature reduction on the current target set and determine the reduced target set as the encoding result of the neural signal; otherwise, perform the calculation of the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories and subsequent steps in conjunction with the current target set.
[0013] Optionally, quantizing the neural signal into a target matrix includes:
[0014] Quantizing the neural signal of a single neuron into N trial ×N stim Okay, N t The target matrix is obtained by dividing the matrix into columns; N trial N represents the number of repetitions of a single stimulus category. stim N represents the number of stimulus categories. t This represents the response signal of a single neuron to a single type of stimulus;
[0015] or
[0016] Quantize the neural signals of the population of neurons into N trial ×N stim row, N t ×N cell The target matrix is obtained by dividing the matrix into columns; N trial N represents the number of repetitions of a single stimulus category. stim N represents the number of stimulus categories. t N represents the response signal of a single neuron to a single type of stimulus. cell This indicates the number of neurons.
[0017] Optionally, updating the information threshold includes:
[0018] Calculate the sum of the current value of the information threshold and the mutual information of the target feature;
[0019] The sum is used as the updated information content threshold.
[0020] Optionally, the initial value is zero; correspondingly, the termination condition is: the updated information threshold is equal to the sum of the information of all time-frequency features in the time-frequency feature matrix obtained by wavelet transform for different stimulus categories.
[0021] Optionally, the feature reduction of the current target set includes:
[0022] For each target feature in the current target set, other target features in the current target set besides the current target feature are made into an object set, and the object set is combined to calculate the second information content of the current target feature for different stimulus categories, so as to obtain the second information content corresponding to each target feature;
[0023] Determine the minimum second information quantity from the second information quantity corresponding to each target feature;
[0024] If the minimum second information content is less than the preset deletion threshold, the target feature corresponding to the minimum second information content is deleted from the current target set, and each target feature in the current target set is executed to make the other target features in the current target set except the current target feature form an object set. The second information content of the current target feature for different stimulus categories is calculated by combining the object set to obtain the second information content corresponding to each target feature and subsequent steps.
[0025] If the minimum second information content is not less than the preset deletion threshold, then the current target set is determined as the deleted target set.
[0026] Optionally, it also includes:
[0027] If there is no first information content greater than the target value, then calculate the third information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories, take the time-frequency feature with the largest third information content as the target feature, and perform the process of filling the target feature into the current target set and subsequent steps.
[0028] Optionally, it also includes:
[0029] The features in the encoding result are classified, and the encoding quality of the encoding result is evaluated based on the classification results;
[0030] and / or
[0031] The stimulus data is reconstructed based on the encoding results, and the encoding quality of the encoding results is evaluated based on the reconstruction results.
[0032] Secondly, this application provides a neural signal encoding device, comprising:
[0033] The signal quantization module is used to quantize neural signals into a target matrix;
[0034] The wavelet transform module is used to perform wavelet transform on the target matrix to obtain the time-frequency feature matrix;
[0035] The initialization module is used to initialize the target set to an empty set and the information threshold to an initial value;
[0036] The calculation module is used to jointly calculate the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories by combining the current target set;
[0037] The selection module is used to select the time-frequency feature with the largest first information content from the time-frequency features with a first information content greater than the target value as the target feature; fill the current target set with the target feature; and delete the target feature from the current time-frequency feature matrix.
[0038] The output module is used to update the information content threshold. If the updated information content threshold meets the termination condition, the current target set is pruned, and the pruned target set is determined as the encoding result of the neural signal. Otherwise, the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories and subsequent steps are calculated in conjunction with the current target set.
[0039] Optionally, the signal quantization module is specifically used for:
[0040] Quantizing the neural signal of a single neuron into N trial ×N stim Okay, N t The target matrix is obtained by dividing the matrix into columns; N trial N represents the number of repetitions of a single stimulus category. stim N represents the number of stimulus categories. t This represents the response signal of a single neuron to a single type of stimulus; or it quantifies the neural signals of a group of neurons into N. trial ×N stim row, N t ×N cell The target matrix is obtained by dividing the matrix into columns; N trial N represents the number of repetitions of a single stimulus category. stim N represents the number of stimulus categories. t N represents the response signal of a single neuron to a single type of stimulus. cell This indicates the number of neurons.
[0041] Optionally, the output module is specifically used for:
[0042] Calculate the sum of the current value of the information threshold and the mutual information of the target feature;
[0043] The sum is used as the updated information content threshold.
[0044] Optionally, the initial value is zero; correspondingly, the termination condition is: the updated information threshold is equal to the sum of the information of all time-frequency features in the time-frequency feature matrix obtained by wavelet transform for different stimulus categories.
[0045] Optionally, the output module is specifically used for:
[0046] For each target feature in the current target set, other target features in the current target set besides the current target feature are made into an object set, and the object set is combined to calculate the second information content of the current target feature for different stimulus categories, so as to obtain the second information content corresponding to each target feature;
[0047] Determine the minimum second information quantity from the second information quantity corresponding to each target feature;
[0048] If the minimum second information content is less than the preset deletion threshold, the target feature corresponding to the minimum second information content is deleted from the current target set, and each target feature in the current target set is executed to make the other target features in the current target set except the current target feature form an object set. The second information content of the current target feature for different stimulus categories is calculated by combining the object set to obtain the second information content corresponding to each target feature and subsequent steps.
[0049] If the minimum second information content is not less than the preset deletion threshold, then the current target set is determined as the deleted target set.
[0050] Optionally, the selection module is also used for:
[0051] If there is no first information content greater than the target value, then calculate the third information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories, take the time-frequency feature with the largest third information content as the target feature, and perform the process of filling the target feature into the current target set and subsequent steps.
[0052] Optionally, it also includes:
[0053] The verification module is used to classify each feature in the encoding result and evaluate the encoding quality of the encoding result based on the classification result; and / or reconstruct the stimulus data based on the encoding result and evaluate the encoding quality of the encoding result based on the reconstruction result.
[0054] Thirdly, this application provides an electronic device, comprising:
[0055] Memory, used to store computer programs;
[0056] A processor is configured to execute the computer program to implement the aforementioned disclosed neural signal encoding method.
[0057] Fourthly, this application provides a readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned disclosed neural signal encoding method.
[0058] As can be seen from the above scheme, this application provides a neural signal encoding method, including: quantizing the neural signal into a target matrix; performing wavelet transform on the target matrix to obtain a time-frequency feature matrix; initializing the target set to an empty set and initializing the information content threshold to an initial value; jointly calculating the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories with the current target set; selecting the time-frequency feature with the largest first information content as the target feature among the time-frequency features with the first information content greater than the target value; filling the target feature into the current target set and deleting the target feature from the current time-frequency feature matrix; updating the information content threshold; if the updated information content threshold meets the termination condition, then performing feature reduction on the current target set and determining the reduced target set as the encoding result of the neural signal; otherwise, performing the joint calculation of the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories with the current target set and subsequent steps.
[0059] As can be seen, after performing wavelet transform on the target matrix obtained from neural signal quantization, this application calculates the mutual information of the selected features (i.e., wavelet coefficients) in the time-frequency feature matrix at each step, in conjunction with the selected features in the target set, to obtain the corresponding joint mutual information (i.e., the first information content). Feature selection is then performed based on this joint mutual information. Joint mutual information effectively describes the degree of information correlation between features and stimulus categories. A larger joint mutual information indicates a greater correlation between the feature and the stimulus category, suggesting that these features respond to a wider range of stimuli. Furthermore, this application uses joint mutual information to select wavelet coefficients at each step, achieving a minimum loss of joint mutual information. Specifically, it selects the time-frequency feature with the largest first information content from those with a first information content greater than the target value, thereby selecting approximately optimal features containing more important information. After selecting features based on joint mutual information, this application further simplifies the selected features, that is, it reduces the number of features in the target set. The simplified features retain as much information as possible, thereby enabling the representation of more information contained in neural signals with as few features as possible, and thus more accurately encoding neural signals. Therefore, this application can improve the effectiveness and accuracy of neural signal encoding results.
[0060] Correspondingly, the neural signal encoding device, equipment, and readable storage medium provided in this application also have the above-mentioned technical effects. Attached Figure Description
[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0062] Figure 1 This is a flowchart of a neural signal encoding method disclosed in this application;
[0063] Figure 2 This is a flowchart of another neural signal encoding method disclosed in this application;
[0064] Figure 3 This is a schematic diagram of a neural signal encoding device disclosed in this application;
[0065] Figure 4 This is a schematic diagram of an electronic device disclosed in this application. Detailed Implementation
[0066] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0067] Currently, wavelet analysis of neural impulse signals can capture detailed temporal dynamics of the signal's time sequence. However, when selecting wavelet coefficients (i.e., signal features) from this temporal information, the calculation of each wavelet coefficient is independent, ignoring higher-order dependencies between them and neglecting potential feature redundancy. This approach only allows for signal extraction and encoding in the temporal sequence, making it difficult to guarantee the effectiveness and accuracy of the signal encoding results. Therefore, this application provides a neural signal encoding scheme that improves the effectiveness and accuracy of neural signal encoding results.
[0068] See Figure 1 As shown in the figure, this application discloses a neural signal encoding method, including:
[0069] S101. Quantize the neural signals into a target matrix.
[0070] In this embodiment, the neural signal is either a single neuron's neural signal or a group of neurons' neural signals. By stimulating a single neuron or a group of neurons with at least one category of stimulus data and acquiring the neural response signals of the single neuron or the group of neurons in response to the corresponding stimulus data, the neural signals of a single neuron or a group of neurons can be obtained. For example, under the stimulation of a natural image, the response signals of a single ganglion cell in the salamander's retina can be acquired using multiple electrodes to obtain the neural signal of a single neuron in the salamander's retina. As another example, under the stimulation of a natural image, the response signals of a group of ganglion cells in the salamander's retina can be acquired using multiple electrodes to obtain the neural signal of a group of neurons in the salamander's retina.
[0071] The data volume of neural signals from a group of neurons is greater than that of neural signals from a single neuron. This embodiment quantizes these signals as follows. In one implementation, quantizing the neural signals into a target matrix includes: quantizing the neural signals of a single neuron into N... trial ×N stim Okay, N t The matrix of columns is used to obtain the target matrix; N trial N represents the number of repetitions of a single stimulus category. stim N represents the number of stimulus categories. t This represents the response signal of a single neuron to a single type of stimulus; or it quantifies the neural signals of a group of neurons into N. trial ×N stim row, N t ×N cell The matrix of columns is used to obtain the target matrix; N trial N represents the number of repetitions of a single stimulus category. stim N represents the number of stimulus categories. t N represents the response signal of a single neuron to a single type of stimulus. cell This indicates the number of neurons.
[0072] S102. Perform wavelet transform on the target matrix to obtain the time-frequency feature matrix.
[0073] In this embodiment, when performing wavelet transform on the target matrix, methods such as Haar wavelet transform can be used, as long as the time-series information in the target matrix can be extracted. The time-frequency features in the time-frequency feature matrix are the wavelet coefficients. If Haar wavelet transform is used on the target matrix, the decomposition level can be set to 4. Each row of the target matrix is decomposed into 4 levels, and the obtained refined coefficients (4 levels D1, D2, D3, D4) and approximate coefficients (1 level A4) are recombine to form a row of the time-frequency feature matrix, thus obtaining the time-frequency feature matrix. The number of decomposition levels can be flexibly changed as needed.
[0074] S103. Initialize the target set to an empty set and initialize the information threshold to an initial value.
[0075] S104. Calculate the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories in conjunction with the current target set, and then execute S105 or S106.
[0076] S105. If there is a first information quantity greater than the target value, then select the time-frequency feature with the largest first information quantity as the target feature among the time-frequency features with the first information quantity greater than the target value, and execute 107.
[0077] S106. If there is no first information content greater than the target value, calculate the third information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories, take the time-frequency feature with the largest third information content as the target feature, and execute 107.
[0078] It should be noted that S106 can avoid feature omission and can supplement the features to be selected. Among them, the third information is the mutual information of the time-frequency features calculated without jointly calculating the target set.
[0079] S107. Fill the current target set with the target features and delete the target features from the current time-frequency feature matrix.
[0080] In the initial stage, the target set is empty, and the initial value of the information threshold is zero. Therefore, when selecting features for the first time, the current target set is empty. At this time, the joint mutual information calculated according to S104 is equal to the non-joint mutual information, that is: I(Y; w|S) = I(Y; w). This represents the first information content of any time-frequency feature w in the current time-frequency feature matrix for each stimulus category Y. Let w represent the third information content of the time-frequency feature w for each stimulus category Y, and S represent the target set. i ∈S, s i For each feature in the target set S that is not w, y∈Y, where y is one of the stimulus categories Y. i ) represents s i The probability of occurrence; p(y,w|s) i ) indicates that s is known i The probability of y and w occurring simultaneously under the given condition; p(y|s i ) indicates that s is known i The probability of y occurring under the given condition; p(w|s i ) indicates that s is known i The probability of w occurring under certain conditions.
[0081] Next, the first information content of each time-frequency feature w is compared with the target value (which is 0). If there is a first information content greater than the target value, the time-frequency feature with the largest first information content is selected as the target feature. If there is no first information content greater than the target value, the third information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories is calculated. That is, the mutual information of time-frequency features is calculated without jointly calculating the target set, and then the time-frequency feature with the largest third information content is selected as the target feature. The selected target feature is filled into the current target set, and the target feature is deleted from the current time-frequency feature matrix. In other words, the selected target feature is moved from the time-frequency feature matrix to the target set. After the first feature selection, the target set changes from an empty set to a non-empty set, and the number of features to be selected in the time-frequency feature matrix decreases by one. It can be seen that after each feature selection, a new target feature is added to the target set, and a time-frequency feature is removed from the time-frequency feature matrix.
[0082] It should be noted that when calculating the mutual information of features, joint mutual information can be obtained if the target set is joined, and non-joint mutual information can be obtained if the target set is not joined. In this embodiment, the first and second information values belong to joint mutual information, and the third information value belongs to non-joint mutual information.
[0083] S108. Update the information content threshold and determine whether the updated information content threshold meets the termination condition; if yes, proceed to S109; if no, proceed to S104.
[0084] In this embodiment, the termination condition is: the updated information threshold is equal to the sum of the information content of all time-frequency features in the time-frequency feature matrix obtained by wavelet transform for different stimulus categories. If W represents the time-frequency feature matrix obtained by wavelet transform, then the sum of the information content of all time-frequency features in W for different stimulus categories is: Where w∈W; p(y,w) represents the probability that y and w appear simultaneously; p(y) represents the probability that y appears; p(w) represents the probability that w appears.
[0085] In one implementation, updating the information threshold includes: calculating the sum of the mutual information between the current value of the information threshold and the target feature; and using this sum as the updated information threshold. The update of the information threshold is expressed by the formula: Updated information threshold σ' = σ + I max σ is the current value of the information content threshold, I max I represents the mutual information (joint mutual information or non-joint mutual information) of the currently selected target features. max =I(Y;w|S) or I(Y;w).
[0086] S109. Perform feature reduction on the current target set and determine the reduced target set as the encoding result of the neural signal.
[0087] In one implementation, feature reduction of the current target set includes: for each target feature in the current target set, forming an object set of other target features in the current target set besides the current target feature, and calculating the second information content of the current target feature for different stimulus categories in conjunction with the object set to obtain the second information content corresponding to each target feature; determining the minimum second information content among the second information contents corresponding to each target feature; if the minimum second information content is less than a preset reduction threshold, deleting the target feature corresponding to the minimum second information content from the current target set, and performing the following steps: for each target feature in the current target set, forming an object set of other target features in the current target set besides the current target feature, and calculating the second information content of the current target feature for different stimulus categories in conjunction with the object set to obtain the second information content corresponding to each target feature and subsequent steps; if the minimum second information content is not less than the preset reduction threshold, determining the current target set as the reduced target set. It is evident that the deletion process aims to delete the target feature w in the target set that minimizes I(Y; w|Sw) and has a second information content less than the preset deletion threshold γ. Sw represents the object set consisting of other target features in the target set besides the current target feature w. Then, the second information content of the target feature w for different stimulus categories Y calculated by combining Sw is I(Y; w|Sw).
[0088] In this embodiment, to verify the validity and accuracy of the encoding results, the features in the encoding results are classified, and the encoding quality of the encoding results is evaluated based on the classification results; and / or the stimulus data is reconstructed based on the encoding results, and the encoding quality of the encoding results is evaluated based on the reconstruction results. When classifying the features in the encoding results, a linear discriminant classification algorithm or other arbitrary classification algorithms can be used.
[0089] As can be seen, in this embodiment, after performing wavelet transform on the target matrix obtained from neural signal quantization, during the selection of time-frequency features (i.e., wavelet coefficients) in the time-frequency feature matrix, the mutual information of the selected features in the target set is calculated for each subsequent selection, yielding the corresponding joint mutual information (i.e., the first information content). Feature selection is then performed based on this joint mutual information. Joint mutual information effectively describes the degree of information correlation between features and stimulus categories. A larger joint mutual information indicates a greater correlation between the feature and the stimulus category, suggesting that these features respond to a wider range of stimuli. Furthermore, this embodiment, when using joint mutual information to select wavelet coefficients, can perform screening with minimal loss of joint mutual information. Specifically, it selects the time-frequency feature with the largest first information content from those time-frequency features where the first information content is greater than the target value, thereby selecting approximately optimal features containing more important information. After selecting features based on joint mutual information, this embodiment further simplifies the selected features, that is, it reduces the number of features in the target set. The simplified features retain as much information as possible, thereby enabling the representation of more information contained in the neural signal with as few features as possible, and thus more accurately encoding the neural signal. Therefore, this embodiment can improve the effectiveness and accuracy of the neural signal encoding results.
[0090] As described in the foregoing embodiments, this application provides a neural signal encoding method that combines wavelet analysis with joint mutual information (which may be called the wavelet-joint mutual information method), which can encode and represent the impulse response signals of a single neuron or a group of neurons, such as: two-photon calcium imaging slow signals, human intracranial electroencephalogram signals, etc.
[0091] Please see Figure 2 In one example, the neural signal encoding process based on wavelet-joint mutual information includes:
[0092] S1. Acquire data, including: stimulus data for each stimulus category and neuronal response data.
[0093] S2, Data Preprocessing.
[0094] Preprocessing for single-cell response data (i.e., response data of a single neuron to each stimulus category): The neural response signal of a single neuron to each stimulus category is organized into N... trial *N t The matrix, N trial N represents the number of times the stimulus was repeated during physiological experiment recording. t N represents the number of data points recorded for this neuron under this stimulus. stim N represents the number of stimulus categories. cell This represents the number of cells. Listing the neuron's response signals to all types of stimuli yields N. trial*N stim Okay, N t The data matrix M consists of columns.
[0095] For the preprocessing of ensemble neural responses (i.e., response data of multiple neurons to various stimulus categories): the neural response signals of a single neuron to each stimulus category are organized into N... trial *N t A matrix, listing the response signals of each neuron to all stimulus categories, yields an N... trial *N stim row, N t *N cell Column data matrix M pop .
[0096] Data matrix M and data matrix M pop Both can be used as target matrices.
[0097] S3. Perform wavelet analysis on the target matrix to obtain the time-frequency information of the neural signal.
[0098] The target matrix is decomposed into four levels using Haar wavelets to obtain four levels of fine coefficients D1, D2, D3, and D4 and one level of approximation coefficients A4. These are then recombined into a single row to obtain the time-frequency feature matrix W.
[0099] S4. Select wavelet coefficients using a strategy that minimizes joint mutual information loss.
[0100] S41: Initial selection of a subset S of wavelet coefficients. First, S is set to an empty set, and the total information value I of the time-frequency feature matrix W is calculated. sum =I(Y;W), set the information threshold σ = 0, and Y represents the stimulus category. Then, in each selection, first calculate I(Y;w|S) for each feature in the current W. If there is an I(Y;w|S) greater than 0, then select the w with the largest I(Y;w|S) to enter the set S, remove the w from W, and update σ = σ + I(Y;w|S). If there is no I(Y;w|S) greater than 0, then calculate I(Y;w) for each feature in the current W, select the w with the largest I(Y;w) to enter the set S, remove the w from W, and update σ = σ + I(Y;w). Then the next selection process can begin, and so on, until σ = I. sum End of selection process.
[0101] S42: Simplify the S set. First, set the hyperparameter - the discrimination threshold γ. Iterate through the coefficients in S to find the w that minimizes I(Y; w|Sw). If the smallest I(Y; w|Sw) < γ, then remove w from S. At this point, S has been updated, so continue to find the w that minimizes I(Y; w|Sw) and compare the smallest I(Y; w|Sw) with γ. If the smallest I(Y; w|Sw) ≥ γ, then output the current S set as the final feature set.
[0102] S5. Use a linear classifier to classify the features in the feature set according to the stimulus category, and judge whether each feature in the feature set encodes the original neural signal well based on the classification results. It can be seen that using a classifier for decoding can verify whether the selected coefficients can well represent the amount of information contained in the neural signal.
[0103] This embodiment selects wavelet coefficients that contribute more information to the stimulus category based on joint mutual information, while taking into account the redundancy among the selected wavelet coefficients.
[0104] According to this embodiment, encoding of single and group neural responses in the salamander retina, single neuron responses of calcium signals in the mouse visual cortex, and human intracranial neural signals all yielded good encoding results.
[0105] Experiment 1: Based on the response data of a single ganglion cell in the retina of a salamander recorded under natural image stimulation, this response data was encoded according to this embodiment. The obtained features can be well decoded to determine the corresponding stimulus category. It can be seen that this embodiment can effectively extract important information from the impulse response of a single retinal ganglion cell.
[0106] Experiment 2: The response data of a single neuron in the calcium signal of the mouse visual cortex were recorded by two-photon under the stimulation of moving gratings in 8 directions. The decoding performance of our method is more than 20 percentage points higher than that of current related technologies, indicating that our method can effectively encode important information in relatively slow neural signals such as two-photon calcium signals.
[0107] Experiment 3: Based on the recorded response data of salamander retinal ganglion cells under natural image stimulation, selecting 18 features according to this scheme achieves a 50% coding performance, indicating that this method can effectively extract important information from the impulse response data of retinal ganglion cell populations. Compared to current related technologies such as WMI (Wavelet Mutual Information), the wavelet coefficients selected by this method originate from more dispersed neurons, demonstrating that this method can better eliminate redundant information.
[0108] Experiment 4: When subjects read 57 different syllables, a dataset of intracranial electroencephalogram (EEG) signals recorded from 13 active channels of the human ventral sensorimotor cortex showed significant cortical activity. The dataset was encoded according to this embodiment, and the resulting features showed good decoding performance, further demonstrating the applicability of this method to coarsely recorded intracranial EEG signals.
[0109] As can be seen, this method not only has good encoding performance for fast neural signals (pulse signals), but also for slow neural signals such as calcium signals. Furthermore, it can also be applied to extracting important information from coarsely recorded neural signals such as intracranial EEG. In addition, this application is applicable to both single neuron signals and group neural signals, and can extract high-dimensional spatiotemporal features from neural group signals. The encoded features can also be integrated, such as integrating multiple neural signals, to utilize multimodal data for various applications.
[0110] The following describes a neural signal encoding device provided in an embodiment of this application. The neural signal encoding device described below can be referred to in conjunction with other embodiments described herein.
[0111] See Figure 3 As shown in the figure, this application discloses a neural signal encoding device, including:
[0112] Signal quantization module 301 is used to quantize neural signals into a target matrix;
[0113] Wavelet transform module 302 is used to perform wavelet transform on the target matrix to obtain the time-frequency feature matrix;
[0114] Initialization module 303 is used to initialize the target set to an empty set and the information threshold to an initial value;
[0115] Calculation module 304 is used to jointly calculate the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories with the current target set;
[0116] The selection module 305 is used to select the time-frequency feature with the largest first information content from the time-frequency features with a first information content greater than the target value as the target feature; fill the target feature into the current target set; and delete the target feature from the current time-frequency feature matrix.
[0117] Output module 306 is used to update the information content threshold. If the updated information content threshold meets the termination condition, feature reduction is performed on the current target set, and the reduced target set is determined as the encoding result of the neural signal; otherwise, the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories and subsequent steps are performed in conjunction with the current target set.
[0118] In one implementation, the signal quantization module is specifically used for:
[0119] Quantizing the neural signal of a single neuron into N trial ×N stim Okay, N t The matrix of columns is used to obtain the target matrix; N trial N represents the number of repetitions of a single stimulus category. stim N represents the number of stimulus categories. t This represents the response signal of a single neuron to a single type of stimulus; or it quantifies the neural signals of a group of neurons into N. trial ×N stim row, N t ×N cell The matrix of columns is used to obtain the target matrix; N trial N represents the number of repetitions of a single stimulus category. stim N represents the number of stimulus categories. t N represents the response signal of a single neuron to a single type of stimulus. cell This indicates the number of neurons.
[0120] In one implementation, the output module is specifically used for:
[0121] Calculate the sum of the current value of the information threshold and the mutual information of the target feature;
[0122] We will use and as the updated information content threshold.
[0123] In one implementation, the initial value is zero; correspondingly, the termination condition is: the updated information threshold is equal to the sum of the information of all time-frequency features in the time-frequency feature matrix obtained by wavelet transform for different stimulus categories.
[0124] In one implementation, the output module is specifically used for:
[0125] For each target feature in the current target set, other target features in the current target set besides the current target feature are made into an object set, and the object set is used to calculate the second information content of the current target feature for different stimulus categories, so as to obtain the second information content corresponding to each target feature;
[0126] Determine the minimum second information quantity from the second information quantity corresponding to each target feature;
[0127] If the minimum second information content is less than the preset deletion threshold, the target feature corresponding to the minimum second information content is deleted from the current target set, and the execution is performed for each target feature in the current target set, so that the other target features in the current target set except for the current target feature constitute an object set, and the object set is combined to calculate the second information content of the current target feature for different stimulus categories, so as to obtain the second information content corresponding to each target feature and subsequent steps;
[0128] If the minimum second information content is not less than the preset deletion threshold, then the current target set is determined as the deleted target set.
[0129] In one implementation, the selection module is further configured to:
[0130] If there is no first information content greater than the target value, then calculate the third information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories, take the time-frequency feature with the largest third information content as the target feature, and perform the process of filling the target feature into the current target set and subsequent steps.
[0131] In one implementation, it further includes:
[0132] The verification module is used to classify the features in the encoding results and evaluate the encoding quality of the encoding results based on the classification results; and / or to reconstruct the stimulus data based on the encoding results and evaluate the encoding quality of the encoding results based on the reconstruction results.
[0133] For more detailed information on the working process of each module and unit in this embodiment, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0134] As can be seen, this embodiment provides a neural signal encoding device that can improve the effectiveness and accuracy of neural signal encoding results.
[0135] The following describes an electronic device provided by an embodiment of this application. The electronic device described below can be referred to in conjunction with other embodiments described herein.
[0136] See Figure 4 As shown in the figure, an embodiment of this application discloses an electronic device, including:
[0137] Memory 401 is used to store computer programs;
[0138] Processor 402 is configured to execute the computer program to implement the method disclosed in any of the above embodiments.
[0139] The following describes a readable storage medium provided in an embodiment of this application. The readable storage medium described below can be referred to in conjunction with other embodiments described herein.
[0140] A readable storage medium is provided for storing a computer program, wherein the computer program, when executed by a processor, implements the neural signal encoding method disclosed in the foregoing embodiments. Specific steps of the method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.
[0141] The terms “first,” “second,” “third,” “fourth,” etc., used in this application (if applicable) are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, or apparatus.
[0142] It should be noted that the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed in this application.
[0143] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0144] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of readable storage medium known in the art.
[0145] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A neural signal encoding method, characterized in that, include: The neural signal is quantized into a target matrix; wherein, a single neuron or a group of neurons is stimulated using stimulus data of at least one category, and the neural response signal of the single neuron or the group of neurons in response to the corresponding stimulus data is collected to obtain the neural signal. Perform wavelet transform on the target matrix to obtain the time-frequency feature matrix; Initialize the target set to an empty set and the information threshold to an initial value; Calculate the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories by combining the current target set; Among the time-frequency features with the largest first information content greater than the target value, the time-frequency feature with the largest first information content is selected as the target feature. The target features are populated into the current target set, and the target features are deleted from the current time-frequency feature matrix; Update the information threshold. If the updated information threshold meets the termination condition, perform feature reduction on the current target set and determine the reduced target set as the encoding result of the neural signal; otherwise, perform the calculation of the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories and subsequent steps in conjunction with the current target set. The feature reduction of the current target set includes: For each target feature in the current target set, other target features in the current target set besides the current target feature are made into an object set, and the object set is combined to calculate the second information content of the current target feature for different stimulus categories, so as to obtain the second information content corresponding to each target feature; Determine the minimum second information quantity from the second information quantity corresponding to each target feature; If the minimum second information content is less than the preset deletion threshold, the target feature corresponding to the minimum second information content is deleted from the current target set, and each target feature in the current target set is executed to make the other target features in the current target set except the current target feature form an object set. The second information content of the current target feature for different stimulus categories is calculated by combining the object set to obtain the second information content corresponding to each target feature and subsequent steps. If the minimum second information content is not less than the preset deletion threshold, then the current target set is determined as the deleted target set.
2. The method according to claim 1, characterized in that, The process of quantizing neural signals into a target matrix includes: Quantizing the neural signal of a single neuron into OK, The target matrix is obtained by dividing the columns of the matrix. Indicates the number of repetitions of a single stimulus category. Indicates the number of stimulus categories. This represents the response signal of a single neuron to a single type of stimulus; or Quantify the neural signals of the population of neurons into OK, The target matrix is obtained by dividing the columns of the matrix. Indicates the number of repetitions of a single stimulus category. Indicates the number of stimulus categories. This represents the response signal of a single neuron to a single type of stimulus. This indicates the number of neurons.
3. The method according to claim 1, characterized in that, Updating the information content threshold includes: Calculate the sum of the current value of the information threshold and the mutual information of the target feature; The sum is used as the updated information content threshold.
4. The method according to claim 3, characterized in that, The initial value is zero; correspondingly, the termination condition is: the updated information threshold is equal to the sum of the information content of all time-frequency features in the time-frequency feature matrix obtained by wavelet transform for different stimulus categories.
5. The method according to any one of claims 1-4, characterized in that, Also includes: If there is no first information content greater than the target value, then calculate the third information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories, take the time-frequency feature with the largest third information content as the target feature, and perform the process of filling the target feature into the current target set and subsequent steps.
6. The method according to any one of claims 1-4, characterized in that, Also includes: The features in the encoding result are classified, and the encoding quality of the encoding result is evaluated based on the classification results; and / or The stimulus data is reconstructed based on the encoding results, and the encoding quality of the encoding results is evaluated based on the reconstruction results.
7. A neural signal encoding device, characterized in that, include: The signal quantization module is used to quantize neural signals into a target matrix; wherein, a single neuron or a group of neurons is stimulated using stimulus data of at least one category, and neural response signals of the single neuron or the group of neurons in response to the corresponding stimulus data are collected to obtain the neural signals. The wavelet transform module is used to perform wavelet transform on the target matrix to obtain the time-frequency feature matrix; The initialization module is used to initialize the target set to an empty set and the information threshold to an initial value; The calculation module is used to jointly calculate the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories by combining the current target set; The selection module is used to select the time-frequency feature with the largest first information content from the time-frequency features with a first information content greater than the target value as the target feature; fill the current target set with the target feature; and delete the target feature from the current time-frequency feature matrix. The output module is used to update the information content threshold. If the updated information content threshold meets the termination condition, the current target set is pruned, and the pruned target set is determined as the encoding result of the neural signal. Otherwise, the first information content of each time-frequency feature in the current time-frequency feature matrix for different stimulus categories and subsequent steps are calculated in conjunction with the current target set. Specifically, the output module is used for: For each target feature in the current target set, other target features in the current target set besides the current target feature are made into an object set, and the object set is combined to calculate the second information content of the current target feature for different stimulus categories, so as to obtain the second information content corresponding to each target feature; Determine the minimum second information quantity from the second information quantity corresponding to each target feature; If the minimum second information content is less than the preset deletion threshold, the target feature corresponding to the minimum second information content is deleted from the current target set, and each target feature in the current target set is executed to make the other target features in the current target set except the current target feature form an object set. The second information content of the current target feature for different stimulus categories is calculated by combining the object set to obtain the second information content corresponding to each target feature and subsequent steps. If the minimum second information content is not less than the preset deletion threshold, then the current target set is determined as the deleted target set.
8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as claimed in any one of claims 1 to 6.
9. A readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1 to 6.