Decoding of the neural representation of the visual RGB color model

Subjects and task-irrelevant

All subjects (N = 14, seven males, seven females, mean age 23.2 ± 2.34 years) had normal vision and agreed with the experimental consent. Each subject accepted financial support with $30/h. They all passed colorblindness screening; none of them are visually impaired. None of them had previously been experimented with on the P300. During the experiment, subjects will accept task-irrelevant random fast active visual color stimuli. Task-irrelevant stimuli do not require subjects to perform any additional tasks and do not require any behaviours and decisions that would have delayed effects. Subjects were only required to keep relaxed during the whole experiment when receiving stimulation (Fig. 1). However, to better control the subjects’ self-awareness when recording course, the stimulation paradigm also adopts a rapid, random and active method to assist. Fast stimuli can well evoke great discharge power of visual rod nerve cells within a very short time when the stimulation occurs. Typically, this process appears within a few milliseconds (Barbosaa, Pires & Nunes, 2016). While external visual stimuli can be reflected by P300 within a latency of 300 ms, the color visual interaction latency period remains unclear (Petrova, Chepin & Voznenko, 2021). In this work, epoch refers to each human-computer interaction of a single subject in one team, one trial of experiments. Each epoch contains the time taken for a complete visual color interaction and the corresponding EEG information. ERP indicates the occurrence of evoking stimulus events. Each epoch contains an active stimulus event (ERP) evoked by a random color. All electrophysiological signals converted by the optic nerve emanate through the cerebral cortex and form electroencephalogram signals (EEG). We have recorded the generation and changes of these EEG signals through experiments. Therefore, in experiment each epoch time windows of interaction were set to 1 s. At the same time, the use of random stimuli and fast stimuli can reduce distractions from other aspects during visual interaction. Because subjects can not respond for a short period, it can ensure that the EEG data contains only information on visual electrophysiological stimulation. In addition to random and fast stimulation, active stimuli use the theory based on ERP to record related random events which can be used in supervised learning to decode and classify RGB colors. In each interaction, there was a randomly generated only one event and one label. We defined this interaction for one epoch (Fig. 1). To decode the RGB color through binary classification, the color of red, green, blue are compared in pairs and cross-validated. We define groups of two colors as RG, RB, and GB. In each team, there were three trials. On each trial, a color have an average of 35 random crossover interactions. Two of the colors interacted a total of 70 times in each random trial. One interaction lasted 1 s. There were 70 epochs generated in one trial. A “yellow triangle” image on screen means the stimuli will start. A “white cross” image on screen means rest (Fig. 1). During the time of rest, the subjects can move freely, and the EEG data were not recorded in this period. Finally, all time of one team was 210 s. All three teams of one subject interacted for 630 s.

All experiments involving human subjects complied with the requirements stipulated by the Helsinki Treaty. All experiments were approved by the Ethical Committee of the Graduate School, Fudan University. Subjects supplied informed consent. The subjects informed consent we received during the experiment.

Visual stimuli

The stimulus was a pure color displayed by the screen after adjusting the uniform brightness (3/5) in an opaque darkroom (Fig. 2 middle four photos just a demo. It is impossible to shoot in the darkroom during the experiment process). The three colors were defined in the Ostwald color space system: this color space is defined according to the 360°, 24-color circle. The three colors were defined by the RGB color space. We fixed the saturability at 100% and fixed the luminance at 50% (Table S1). We designed the color such that (i) the variable in stimulus information vary only around RGB values, and (ii) the effects of interchromatic saturation and lightness were excluded between separate hues. We can approximate that the results of decoding are not to be influenced by other aspects.

Data acquisition and preprocessing

Subjects EEG data were collected by Start Neural Sen W system. We used Trigger Box designed the experimental paradigm. The raw EEG data sampling at the rate of 1,000 Hz.

After collection, the raw EEG data were preprocessed to remove the effect of head movements, and then use notch filtering to continue to remove the effect of power-frequency noise. Power frequency noise is usually above 50 Hz, we processed the raw EEG data using notch filtering. The EEG signal effective frequency band was usually concentrated in the low-frequency area less than 30 Hz, so high-frequency information greater than 30 Hz was of no value to the experiment, we use a basic finite impulse response (FIR) to filter it out. Next, the data should be downsampled. We resampled the raw data frequency reduced from 1,000 Hz to 200 Hz. The resampling aim was to minimize the uncorrelated association between a few sampling points. After that, we continue processed the EEG data to remove artifacts caused by eye movement to ensure signal purity. Following this, each trial EEG data was extracted and saved by NSW 64 system (−200–1,000 ms around stimulus onset) at the same time removing the baseline mean (Table S1).

Control the environment to reduce the effect of the experiment

Before starting the prime experiments, we arranged the external environment of the experiment to reduce the impact on decoding. The whole experiment process was in a room without any light to prevent any impact of external light and color. During the whole experiment, the subjects were required to keep relax. At the same time, the subjects were asked to hold their jaws on a shelf (15 cm) and keep a stable sight distance (36 cm) to the screen during the whole experiment (Fig. 2). Then, they can put on the EEG hat (Fig. 2 left inset).

All channel positions on the EEG hat must be checked correctly. In this work, we used 59 channels international standard Fpz-O2 system (the channel ECG, Veol, Veou, Heor, Heol are the signals of electromyography (EMG) which were no use in our experiment) (Fig. 2 right inset). First, we must ensured the CZ channel at the sagittal line’s central. Next, adjusting the CZ channel equal spaces to left/right ears, forehead nasal tip, and inion. The position of the CZ channel must be accurate, and keep the error at a minimum for 0.2 cm.

When EEG data was recorded, a high-level impedance would generate more noise that affects the classification decoding accuracy. We kept the impedance of each channel lower than 3 kΩ (Table S2). Using a medical coupling agent can well reduce the impedance. Before EEG data was recorded, we applied it into each channel. Meanwhile, during the whole experiment, we overseed the impedance of the electrode channel. In addition, the synchronizer was used to reduce the time error of simulation and recording when interacting.

EEG signal data analysis and processing

The analysis and processing of EEG signal data was carried out using Matlab software. There were 210 empty epochs in the resting time when the screen showed the “white cross”; these empty epochs didn’t have any color information, and they were rejected (over three trials). The experiment was performed continuously in 1-s steps, and we added an empty epoch between two epochs containing EEG color signals. Continuous fast stimulation will make the eyes unable to adapt and adjust in time, and at the same time, disturbance signals will be generated due to eye movement and blinking. The empty epoch sandwiched between two consecutive stimulation epochs with EEG information allows the eyes to rest properly and prepare for the next stimulus interaction. The empty epoch also allows the eyes to adapt and perform physiological actions such as eye movement and blinking during this second. Therefore, the epochs containing EEG information are both independent and intermittently continuous with each other. The empty epoch can directly remove some unavoidable disturbance signals generated by blinking and eye movement from the source during the signal acquisition process, avoiding the trouble of separating from the EEG signal in the subsequent data processing. Because in many previous studies, the disturbance caused by eye movement and blinking will affect the screening of EEG signals, and it is easy to delete the required EEG signals by mistake in the later signal processing process. Other epochs which contained the information of RGB color in three trials were a totally of 210 epochs. To make sure that each RGB color was analysed for the same number in trials, these epochs contained the random colors of two in one team for 105 (team: RG, RB, and GB) (Table S2).

Two parts consisted of the whole decoding process. First, subjects saw the target, leading the active stimuli signal reach to the rod cell layer, after that transmission to the brain (Fig. 3). When the rod cell layer stimuli information was sensed by the brain, within a short time the nerve endings would discharge great power (Fig. 3). The Start NeuSen W system started recording EEG data. The system produced by Neuracle Technology (Changzhou) Co., Ltd. Next, the NSW 64 system saved EEG signal data and transferred it to the decoding model classifier (Fig. 3).

Machine learning was used binary classification for decoding. Supervised learning based support vector machine (SVM) and feed-forward neural network (FNN) techniques were used to train classifiers on all 14 subjects 59 channels EEG signal data. Linear function and layers.conv1d function was used in the data analysis (SVM) system contains many main functions for kernel function, Gaussian function and linear function. After many comparison tests, in this work, it was more acceptable to use the linear function on binary calssificaiton for processing EEG signal data. In FNN, the data was trained and classified by the neural network; next, using layers.conv1d function to extract the features. At last, for classification, putting the features into the FNN which fully connected (Fig. 4). The training datasets ratio and testing datasets ratio were adjusted for 8:2 (training datasets was 80% for whole datasets, testing datasets was 20% for whole datasets) (Fig. 4). For the EEG signal dataset we collected is a balanced data set and a small sample data set, it is suitable to use HoldOut cross-validation. In this technique, the whole dataset is cut into two sets with training sets and validation sets randomly. Based on experience, usually, 80% of the whole dataset is chosen for training and the resting 20% for validation. The advantage of this cross-validation is that it can be performed quickly: For we must cut the dataset once into two sets one for training and the other for validation, and there will be only one for the model built on the training set, it can be performed quickly. Since our dataset is a balanced dataset, and each group only has a two class (“red” and “green”, “red” and “blue”, “green” and “blue”) in two situations. Taking the two cases of “red” and “green” classes as an example, suppose 50% of the EEG signal data is class of “red” and the rest 50% of the data is calss of “green”. A train-test split was performed with a training set of 80% and a test set of 20% for the whole dataset. It cannot happen that all 50% of the data in the class of “red” is in the training set and all the data in the class of “green” is in the test set. Because the maximum limit of the test set is only 20%. However, in order to avoid some extreme situations, for example, 50% of “red” in the training set, or 50% of “green” in the training set, we need to iterate each result and find an average value. A single classification result is not enough to represent all the results. So our model generalizes well to our test data; since it has not seen the “red” or “green” class of data before; moreover, in the case of small datasets, all samples are used to test the model, so our model will not miss any important features since it was trained on all the data. In addition, iterations number was adjusted for 30 times. When the iterations number over 30 times, the system was overfitted. Therefore, the system fit iterations was 30 times. Time series cross-validation were used in this work. Data is collected at different time points. It were collected in adjacent time periods. We divide the EEG signal data into two sets of training and validation by time. The training were started with a small set of data. The later date points were predicted based on this set and then check for accuracy. Then the training EEG dataset of next contains the predicted samples as one part and forecasts on consecutive samples. In FNN, we use a single-layer network, and function layers.conv1d. Since our dataset is irreversible as the time series moves forward, FNN is very suitable. Because in FNN information also always moves in one direction; it never goes backwards. The kernel we use in SVM is mainly a linear function. It is a polynomial kernel function special case, simplicity is its great advantage, but disadvantage is that there is no solution for linearly inseparable data sets. The linear separable case is its main use. We find that from the feature space dimension to the input space dimension all in the same, and its parameters are scarce and speedy. It is ideal for classification effect to linear separable data, so a linear kernel usually first try to use. If the classification effect of the function is not ideal, tyr to use other kernel functions. To specify the object, the property-value pairs are used in the constructor. After the object is created, to access the values of the object property the dot notation can be used (Table S3). Since our dataset is not high-dimensional, it is not possible to use a kernel such as the Gaussian kernel function for testing. Because the Gaussian function can expand the input feature vector into an infinite-dimensional space. The value calculated by the Gaussian kernel function is always between 0 and 1. Samples can be mapped to higher dimensional space, which will cause the original characteristics of the data to be lost. In addition, due to the small amount of sample data, other kernel functions are not applicable. Finally, classification decoding used the 30 times iterations average value. EEG datasets which contain 210 epochs were cut into 168 epochs and 42 epochs for training and testing (Fig. 4). Whole datasets contained 5,040 and 1,260 epochs in training and testing after 30 times iterations (Fig. 4). By cross-validation got the classification decoding final result (Fig. 4). An epoch is defined as: the smallest and most basic interaction unit, including a one-time random single-color fast stimulus interaction time and corresponding EEG information. Acquired individual subject EEG data are usually stored in .bdf format. The data includes raw data and label data. The label data records the start time of the relevant event (color fast stimulus event) generated by ERP. In this work, we have incorporated related methods for performing data reading while building the model. Data can be read in MATLAB via importNeuracle, or in Python via MNE.

In FNN, in order to improve the processing efficiency of the model, N samples are usually grouped into one team and calculated in batches. Supposing that the input of the first layer of the network is $A^{\left(l-1\right)}/in/mathbb{R}^{N/times{M}_l}$, where each row is a sample, the calculation formula of the first layer in the FNN is:

(1) $$Z^{\left(l\right)}=A^{\left(l-1\right)}{W}^{\left(l\right)}+b^{\left(l\right)}/in/mathbb{R}^{N/times{M}_l}$$

(2) $$A^{\left(l\right)}=f_l\left(Z^{\left(l\right)}\right)/in/mathbb{R}^{N/times{M}_l}$$

Among them, ${\mathrm{Z}}^{\left(\mathrm{l}\right)}$ is the neurons net activity value in layer l of N samples, ${\mathrm{A}}^{\left(\mathrm{l}\right)}$ is the neurons activity value in layer l of N samples, ${\mathrm{W}}^{\left(\mathrm{l}\right)}/\mathrm{i}\mathrm{n}/\mathrm{m}\mathrm{a}\mathrm{t}\mathrm{h}\mathrm{b}\mathrm{b}{\mathrm{R}}^{{\mathrm{M}}_{\mathrm{l}-1}/\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\mathrm{s}{\mathrm{M}}_{\mathrm{l}}}$ is the layer l weight matrix, and ${\mathrm{b}}^{\left(\mathrm{l}\right)}/\mathrm{i}\mathrm{n}/\mathrm{m}\mathrm{a}\mathrm{t}\mathrm{h}\mathrm{b}\mathrm{b}{\mathrm{R}}^{1/\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\mathrm{s}{\mathrm{M}}_{\mathrm{l}}}$ is the layer l bias. Here, a tensor of shape (number of samples × feature dimension) is used to represent a set of samples. The matrix X of samples is composed of N row vectors of x. In order to make the subsequent model building more convenient, we enhatsulate the computation of the neural layer into operators, and these operators all inherit the op base class.

All classification models can be run under MATLAB 2019A and Python environment. Codes of SVM and FNN can be found in Science Data Bank. All results in the manuscript can be reproduced.

Decoding RGB color classification feature

Figure 5 shows the scatter distribution of RG, RB, GB classification accuracy on 59 array channels distributed on the cerebral cortex using machine learning decoding. The consequence display that RGB color EEG can be decoded by SVM and FNN. When the time window of each epoch was cut to 500 ms after the stimulus, the results with higher classification accuracy were mostly distributed in the primary visual cortex (dark grey in Fig. 5). Figure 6 shows the specific performance of the classification results. We performed statistics on the EEG data of 14 subjects using boxplots. Each vertical line, it contains the classification accuracy distribution range of the same channel in the cerebral cortex of 14 subjects. The top of the vertical line represents the highest classification accuracy obtained by this channel (maximum) and the bottom represents the lowest classification accuracy obtained by this channel (minimum). The dot in the centre of the vertical line represents the average classification accuracy of the 14 subjects of this channel. The length of the vertical line represents the magnitude between the highest and the lowest classification accuracy extreme values of this channel. We divided the data into quartiles (perc: 25–70%), that is the box contains 50% data of one channel. The diamond-shaped symbols represent individual outliers with a large dispersion degree. The range of IQR is within 1.5. The horizontal line in the middle of the box represents the median Q2, Q1 is the box upper edge, Q3 is the box lower edge. We found that after removing the more discrete values, the rest of the 50% data in the valid box reflected that both the Q2 and the average value (dark grey area in the figure) were higher in the primary visual cortex than in the other cortex (Fig. 6). In the primary visual cortex of the RG team, box, Q2, and average positions were around a value of 70%; in the other cortex around 60% (Fig. 6A). In the primary visual cortex of the RB and GB team, box, Q2, and average position values were not as high as those in the RG team, but also around 60%; in the other cortex around 50% (Figs. 6B and 6C). It is important to note that the classification accuracy size does not offer an effect on absolute size weigh, but it is still a comparative effect dimension valid method in a certain research (Sato & Inoue, 2016). By comparison, it was found that the values of the primary visual cortex were slightly higher than other cortexes.

Let us use another method to intuitively compare the channels classification decoding results of the primary visual cortex (17 channels from Pz-O2) with the other cortex (42 channels from Fpz-TP8). Figure 7 shows each channel decoding classification results through a matrix heat map. The column vector of the matrix represents the 14 subjects in the experiment, the row vector represents the localization channels (59 channels from Fpz-O2) distributed on the cerebral cortex, and the color depth in each square represents the size of classification accuracy value (red > 70%, yellow 50–60%, green < 50%). We formed these three parameters set into three-dimensional data (3 × 3 data) and constructed a three-dimensional space matrix (3 × 3 matrix). In the column vector direction of the matrix, the classification accuracy is sorted according to the size of the gradient, and it is found that most of the regions are yellow (50–60%), which indicates that almost all EEG data are effectively classified by SVM and FNN. The red regions representing high classification accuracy are almost concentrated in the primary visual cortex (>70%). This result shows that RGB color information can be well classified and decoded in the primary visual cortex.

In addition, we use 3D color topographic map that stereoscopically rendered the 3D matrix (Fig. 8). 3D color topographic map more intuitively shows all chanels classification accuracy distribution. The 3D color topographic map on the left side of Fig. 8 shows the channel accuracy of classification gradient change from other cortex to visual cortex on the X-axis (channel axis). Starting from channel No. 43, the classification accuracy has a steep upward climbing trend to the last channel No. 59, and the classification accuracy shows a positive increase rate and reaches the extreme value (see Fig. 8D for the corresponding code of the channel). Channel No. 43 is a border between the primary visual cortex and the other cortex. Therefore, it can be intuitively found that the decoding effect of classification accuracy starting from the primary visual cortex has strong robustness. We further obtained contour maps by Fourier transforming the 3D color topographic map (Fig. 8 right inset). The results are consistent with the matrix heatmap shown in Fig. 7. Therefore, we can summarize the result: channels with higher classification accuracy are concentrated in the primary visual cortex.

Figure 9 shows the 14 subjects’ average classification results in comparison to SVM and FNN, the results are similar, indicating that the two classifiers have little error and the results are reliable. At the same time, we also analyzed the primal visual cortex with higher classification accuracy, it was found that the results obtained under the two classifiers were the same (in each team the two curves fit) (Fig. 9 right inset). This result suggests that the primary visual cortex is active when the brain processes the RGB color information sensing. Therefore, we can conclude that it is possible to interpret RGB color information through machine learning. These conclusions sustain a hypothesis: the brain can process RGB color information independently in the surrounding task-irrelevant decoding.

Figure 10 shows the parameter analysis on different machine learning models. The ROC curves (receiver operating characteristic curve) show the data training and data testing decoding results with SVM and FNN from EEG collected on 10 channels of the primary visual cortex of 14 subjects. In each team, there were included all 826 data. We divide 826 data to test the two classifiers’ performance and carry out statistical significance testing. Each team input data range is 826. Smaller and larger test measurement values respectively indicate a positive tests and a biased test. Channel POz-O2 is the actual positive state (all: 826; positive: 140; negative: 686). Standard error is 95%, and confidence levels is also 95%. The direction of test is positive v.s. high. Diagonal reference lines are included in results. In the input data, there are no missing values, and the bad data and invalid values are not used in calculations The RG, RB, GB (SVM and FNN) ROC curves analysis results were listed in Table S4. Analysis results indicate that AUC of SVM and FNN in RG: 0.89 > 0.5, 0.82 > 0.5. In RB: 0.78 > 0.5, 0.64 > 0.5. In GB: 0.73 > 0.5, 0.69 > 0.5. The results all over 0.5 have a valid accuracy (Table S4). The two curves sensitivity and 1-specificity are biased towards the diagonal reference line upper left, which means our two models have a bigger examination accuracy. The two curves are fitting means the model results are similar and stable.

Cross-temporal decoding for visual color ERP

The classification decoding results we have discussed so far have been based on evaluating the performance of the classifier using test data obtained at the same period after stimulation as training. The classifier showed significant decoding performance within 300 ms after stimulus onset. The possibility of this phenomenon arises is that the stimulation of RGB color-related events evokes transient activity patterns of visual potentials (Cherepanova et al., 2018). To decode the time course features of transient activity patterns, we performed visual and statistical analyses of the EEG data. Figure 11 shows ERP plots of 14 subjects’ with red, green, blue stimuli in 500 ms on primal visual cortex 17 channels (Pz-O2). Each curve represents a channel. The size of the peak shows the average result of the optic nerve cell firing voltage for all epochs of the same color. We usually think of P300 as an endogenous component of ERP (King & Dehaene, 2014). Therefore, the position of the peak mapped on the time axis (horizontal axis) in Fig. 11 is also an endogenous component of ERP, which is consistent with the P300 theory. The cross-temporal distribution features of all VEPs can be found from the peak-to-time correspondence in Fig. 11: The peak of power appeared within 300 ms latency after stimulation. We performed sample statistics on ERP time points drawn from 14 subjects. All ERP’s 300 ms scatter distribution was arranged in three teams of red, green, and blue (Fig. 12A). Figure 12A shows these three colors in each time segment are concentrated at similar time points. Three teams are distributed continuously. However, some data have a larger degree of dispersion (The reason for the larger error is mainly because the individual differences of some subjects are too large or generated during the experiment). To exclude the influence of these small outliers, we use the method of median statistics. Since the median is obtained by sorting, changes in some data do not affect it. When the individual data in a group of data changes greatly, it still can describe the central tendency of this group of data. The data was cut into quartiles and used boxplots and violin plots to present them. Box perc: 25–75%; IQR is within 1.5. After sorting obtained red: Q3 = 185 ms, Q1 = 195 ms; green: Q3 = 200 ms, Q1 = 233.75 ms; blue: Q3 = 222.75 ms, Q1 = 255 ms. The time segments of RGB color latency are red: 185–195 ms; green: 200–233.75 ms; blue: 222.75–255 ms. Time segments are distributed continuously, and the boxes are also distributed continuously (Fig. 12B). Figure 12B violin plot further shows that the data has a larger kernel density near the median (Q2), indicating that the distribution of data near the median (Q2) is probability larger (the distribution of kernel density is normal). Finally, the cross-temporal decoding results are obtained by performing median (Q2) statistics on the three-time segments of RGB colors: red latency (Q2) = 190 ms; green latency (Q2) = 215 ms; and blue latency (Q2) = 238.5 ms (Fig. 12B the point in the middle of the box).

This result suggests that it is feasible for independent decoding RGB color across time courses based on the theoretical mechanisms of ERP and P300. In addition, the research result indicates that at the time the brain senses the stimuli from RGB color, red was the shortest evoked, then green and blue. It also reveals why human visionual is so sensitive to red. This can also be used to explain why red is often used to inform and remind others of danger (Acqualagna, Treder & Blankertz, 2013).

Comparing the spatial position of classification decoding and power distribution

To directly compare the space position of classification decoding and power distribution, the 59 channels international standard system were used (Fpz-O2) (Engel, Zhang & Wandell, 1997) (EMG are not used in this work) (Sitaram et al., 2017). Our goal is to use the results of classification accuracy and power distribution obtained from EEG data channels to reverse the calculus of their position in the cerebral cortex. We note that RGB color information exists in the primary visual cortex through previous classification experiments. Meanwhile, the cross-temporal course independent decoding of VEPs shows the features of power distribution. Therefore, we also need to use a simple random sampling analysis of 14 subjects to account for the correlations between classification decoding and power distribution. By comparing the correlation between these results to determine whether these features all come from the same channel position.

Figure 13 ABC shows the comparison results of 4 subjects’ simple random sampling analysis (N = 14, n = 4, 4/14): the 3D topographic heat map (use MNI coordinate file for BEM dipfit model) shows the distribution of power, which is evident in the primary visual cortex; darker areas are found in the primary visual cortex (Fig. 13 right inset, with the channel POz-O2, are prominent), confirming the previous classification and decoding results. We also generated time-frequency plots of 16 consecutive epochs of POz channels with high classification accuracy and power distribution (Fig. 14, random sample of nine subjects).

EEG signals localized to channel POz achieved high results in both machine learning classification and VEP power discharge. The frequency, time and power parameters 3D data spectrograms composed of short-time Fourier transforms (STFTs) exhibit strong correlations. These epochs are arranged at certain intervals according to time series and different color labels. The power discharge is concentrated in the low-frequency region (<10 Hz), and some high power appears within 300 ms of each epoch (0.6–0.8 μV, Fig. 14). These results provide potential correlations between classification accuracy decoding, power distribution, and time course.

Comparing the correlation results of space position, we can get that, first, in the primary visual cortex, RGB color can be independently decoded, and it is easier to decode in the POz channel; Second, the decoding results are interoperable and consistent, which can be cross-validated by the channel space position. Third, this is also the position where the neural representation of RGB color appears.

The distribution features of space position reveal the brain processing RGB color information neural mechanism. We explain the results of position distribution as neural representations evidence in the brain for visually received RGB color stimuli, and we will return to this in the discussion below. The latency to the peak value of decoded RGB color EEG power is often affected by the designed task. Task-relevant sequential stimuli will impact the information’s feed-forward flow (Zhang & Luck, 2009), and after 200–300 ms stimulus the task effects typically emerge onset (Hebart et al., 2018; Grootswagers et al., 2021). Constant with the task effect late reach, EEG feedback from RGB color stimuli was disrupted at times <300 ms. Compared with task-relevant random stimuli, using task-irrelevant random stimuli can obtain purer EEG information. In visual sensing, the extent of the latency incarnates the differences in stimuli perception and classification, and more perceptually different stimuli can be decoded at a later time, and are compared to those performed by regions further down the visual processing hierarchy calculation associated (Proklova, Kaiser & Peelen, 2019; Carlson et al., 2013; Cichy, Pantazis & Oliva, 2014). And the latency of ERP power to reach the peak value caused by the perception of the stimuli was less than 300 ms. Within the same time window (the interaction time window for an epoch is 1 s), the results of stimulus perception and classification are the same, implying that the neural mechanism of RGB color is therefore consistent with the hypothesis that the brain can independently receive and process RGB color information. There are functional areas in the cerebral cortex that can independently perceive and decode RGB color information.