Neurophysiology of creativity and machine learning applications for creative process’ stages differentiation through assessment of EEG/VP signals
- Authors: Shemyakina N.V.1, Nagornova Z.V.1
-
Affiliations:
- Sechenov Institute of Evolutionary Physiology and Biochemistry of the Russian Academy of Sciences
- Issue: Vol 18, No 4 (2023)
- Pages: 882-885
- Section: Conference proceedings
- URL: https://genescells.ru/2313-1829/article/view/623517
- DOI: https://doi.org/10.17816/gc623517
- ID: 623517
Cite item
Open Access
Access granted
Subscription or Fee Access
Abstract
Of particular interest for researching the cognitive specializations of neurons is their correlation with environmental variables and animal behavior. Mutual information (MI) is a preferable method for measuring such correlations, as it allows for the assessment of non-linear relationships between variables, detects synchronization, and provides both significance and strength quantification. However, calculating MI for real data is significantly challenging. In this study, we used updated MI calculation techniques to analyze the connection between calcium fluorescence signaling and behavioral variables. Our approach encompasses novel strategies which we compiled into a software program known as INTENS (Information-Theoretic Evaluation of Neuronal Specializations), and it enabled to identify specialized neurons in mice hippocampal calcium activity data while they explored the arena with varying levels of novelty.
Numerous methods exist for analyzing the relationship between neuron spikes and behavioral variables, including information-theoretical approaches [1]. Extracting information about the relationship between calcium fluorescent signals and behavior is of particular interest due to the signal’s ability to provide crucial information about subthreshold activations of the neuron. In this study, we use the GCMI Gaussian copula entropy method to calculate mutual information [2]. This method relies on the fact that mutual information between two random variables is independent of their marginal distributions and only depends on the type of copula used (a multidimensional distribution where each marginal distribution is uniform).
The actual MI was compared to its corresponding values computed on the time-shifted signals for assessing the statistical significance of the computed information association between the calcium signal and the behavioral variable. Additionally, we devised a technique for gauging the strength of the coupling effect. This involved normalizing the mutual information between the fluorescence signal and the behavior with the entropy value of both variables, previously calculated as random variables. Importantly, the approach outlined earlier is effective for analyzing continuous variables such as calcium signal and animal speed, as well as pairs of continuous and discrete variables such as calcium signal and the presence or absence of grooming.
The analysis of calcium signals recorded from the CA1 region of the hippocampus revealed neuronal specializations related to the animal’s external environment, such as place cells, and specializations related to its behavioral activities, including neurons activated during running, rearing, and freezing. Some neurons selectively activated in response to discrete parameters included the animal’s location within the arena (center, walls, and corners) and its speed (rest, slow, and fast). A total of 781 specializations were detected across 472 neurons throughout all four sessions of the experiment. Notably, a single neuron could have several specializations. However, more than half (55%) of the neurons were found to have only one specialization.
Full Text
The study presents the findings of investigations into the neurophysiological characteristics of verbal and nonverbal (artistic) creativity, alongside a comparative analysis of various approaches to classifying time series and time-frequency maps (using continuous wavelet transformation) for EEG signals.
The studies included various groups of subjects who completed different creative tasks while undergoing EEG or ERP registrations. These tasks included creating unique endings to familiar proverbs using the ERP paradigm [1–3], developing a plot based on a provided image called Story [4] (referred to as continuous activity during changing states), completing sketches using the Torrence test called Sketches, and creating freeform artistic drawings on canvas and receiving feedback from a professional artist, referred to as sketching, canvas, review, etc. [5].
It was demonstrated that searching for the original ending, in contrast to the control task of recalling the ending of the proverb, led to increased power values of 8–9 Hz in the right frontal and left parietal regions of the cortex. Additionally, inventing a story based on a picture, as opposed to simply describing the picture, was associated with a higher percentage of individual alpha frequencies in the EEG structure. Under the conditions of unconstrained sketching, a greater proportion of theta (5–6 Hz) and alpha frequencies (12–13 Hz) in the frontal and parietal regions of the artist’s cortex were discovered in contrast to the control task of drawing lines.
The study aimed to explore the potential of accurately classifying states of creative activity by analyzing EEG/ERP characteristics during creative tasks. This will further aid the development of a machine learning algorithm that distinguishes stages/states of the creative process through EEG/ERP characteristics.
Linear classification was used for the time series (raw signal) in both reference values and conversion to current source density (CSD). Continuous EEGs were divided into 2-second fragments for the classification of time series, and 1500 ms intervals were used for the classification of ERPs after stimulus presentation. In addition, the EEG/ERP data was presented as time-frequency maps, calculated using the continuous wavelet transform (Morlet) for frequencies ranging from 3 to 30 Hz. Each image was generated for a 4-second interval with increments of 100 milliseconds (for continuous EEG recordings) or for a 1500-millisecond interval after presentation of the stimulus (for evoked potentials). These images consisted of time-frequency maps of zones of interest, specifically the frontal (Fz) and parietal (Pz) leads. The time-frequency map images were classified using a modified convolutional neural network with 47 layers based on the SqueezeNet architecture. EEG classification procedures were carried out using the Classifier learning and Deep Network Designer software package options in the Matlab environment. The training and test samples never overlapped and were proportioned 80:20%.
When classifying states for the Story creative task model, four categories were used: describing an image, creating a plot, advancing a plot, and a background with eyes open. SqueezeNet yielded the highest accuracy rate of 53.4%.
When classifying the stages of Free Artistic Creativity, three states of a professional artist were used for classification: background with open eyes, painting on canvas, and viewing and evaluating work on canvas. Similarly, the Sketches model for non-verbal creativity classified three states: background with open eyes, creative drawing in the Torrance test, and drawing of given objects.
In the first instance of spontaneous drawing and self-evaluation, the Kernel Naive Bayes classifier achieved an 86.94% accuracy in time series classification with CSD conversion. In the second scenario, a support vector machine (Gaussian radial basis function) classifier achieved a 66.9% accuracy in distinguishing between background states and creative/non-creative drawing. The use of time-frequency maps coupled with a convolutional neural network resulted in a classification accuracy of 98.2% in the first case and 96.5% in the second case.
In studies of ERP paradigms using the Proverbs model, a single-trial approach was used to classify three states: generating a novel ending for a well-known proverb, identifying a semantic synonym for the ending, and recalling the original ending of the proverb. Classification was conducted individually for each of the 15 participants. All outcomes, determined using linear discriminant analysis, surpassed the threshold for random recognition, with results ranging from a minimum of 37.8% to a maximum of 58.5% and an average of 46±6% for the group.
Since the Proverbs model only included tasks with a maximum of 100 samples, it was not possible to acquire an adequate number of time-frequency maps for training a convolutional neural network. To address this limitation, the training sample was formed by using image samples from all participants, resulting in 922 samples for finding the original answer, 1102 for finding a synonym, and 1180 for recalling the end. In this instance, the categorization of the sample population as a whole did not surpass the arbitrary threshold of 36%, which may be attributed to the variability between subjects within the dataset.
Convolutional neural networks have demonstrated superior performance in classifying continuous and long-term states of creative activity. On the other hand, estimating fast transients proves to be more efficient in classifying time series.
ADDITIONAL INFORMATION
Authors’ contribution. All authors made a substantial contribution to the conception of the work, acquisition, analysis, interpretation of data for the work, drafting and revising the work, final approval of the version to be published and agree to be accountable for all aspects of the work.
Funding sources. The study was supported by RSF grant No. 22-28-02073.
Competing interests. The authors declare that they have no competing interests.
About the authors
N. V. Shemyakina
Sechenov Institute of Evolutionary Physiology and Biochemistry of the Russian Academy of Sciences
Author for correspondence.
Email: shemyakina_n@mail.ru
Russian Federation, Saint Petersburg
Zh. V. Nagornova
Sechenov Institute of Evolutionary Physiology and Biochemistry of the Russian Academy of Sciences
Email: shemyakina_n@mail.ru
Russian Federation, Saint Petersburg
References
- Bechtereva NP, Danko SG, Medvedev SV. Current methodology and methods in psychophysiological studies of creative thinking. Methods. 2007;42(1):100–108. doi: 10.1016/j.ymeth.2007.01.009
- Shemiakina NV, Dan’ko SG, Nagornova ZhV, et al. Changes in the power and coherence spectra of the eeg rhythmic components during solution of a verbal creative task of overcoming a stereotype. Fiziol Cheloveka. 2007;33(5):14–21.
- Shemyakina NV, Nagornova ZV. Does the instruction “Be original and create” actually affect the EEG correlates of performing creative tasks? Hum Physiol. 2020;46:587–596. doi: 10.1134/S0362119720060092
- Shemyakina NV, Nagornova ZV. EEG “signs” of verbal creative task fulfillment with and without overcoming self-induced stereotypes. Behav Sci (Basel). 2019;10(1):17. doi: 10.3390/bs10010017
- Shemyakina NV, Potapov YG, Nagornova ZhV. Dynamics of the EEG frequency structure during sketching in ecological conditions and non-verbal tasks fulfillment by a professional artist: case study. Human Physiology. 2022;48:506–515. doi: 10.1134/S0362119722700050
Supplementary files
