Modeling the Effect of Prior Knowledge on Memory Efficiency for the Study of Transfer of Learning: A Spiking Neural Network Approach

Abstract

The transfer of learning (TL) is the process of applying knowledge and skills learned in one context to a new and different context. Efficient use of memory is essential in achieving successful TL and good learning outcomes. This study uses a cognitive computing approach to identify and explore brain activity patterns related to memory efficiency in the context of learning a new programming language. This study hypothesizes that prior programming knowledge reduces cognitive load, leading to improved memory efficiency. Spatio-temporal brain data (STBD) were collected from a sample of participants (n = 26) using an electroencephalogram (EEG) device and analyzed by applying a spiking neural network (SNN) approach and the SNN-based NeuCube architecture. The findings revealed the neural patterns demonstrating the effect of prior knowledge on memory efficiency. They showed that programming learning outcomes were aligned with specific theta and alpha waveband spike activities concerning prior knowledge and cognitive load, indicating that cognitive load was a feasible metric for measuring memory efficiency. Building on these findings, this study proposes that the methodology developed for examining the relationship between prior knowledge and TL in the context of learning a programming language can be extended to other educational domains.

1. Introduction

According to Gardner [1], true learning occurs when an individual comprehends a concept deeply enough to memorize the information related to the concept and retrieve it to apply appropriately in a novel situation. The transfer of learning (TL) refers to the ability to apply acquired knowledge and skills in new and diverse contexts. This ability is crucial for success in real-world scenarios, allowing individuals to adapt and perform effectively in various environments [2]. Achieving the TL is the ultimate goal of any educational system as the TL extends beyond simple memorization and recall; it involves making meaningful connections between prior and new knowledge, critically evaluating information, and applying learned concepts across different domains [3]. Despite its importance, achieving TL is not always straightforward. The extent to which knowledge and skills transfer to new contexts depends on various factors, such as the similarity between the learning tasks and the instructional conditions under which learning occurs.

The TL is commonly classified as ‘near transfer’ and ‘far transfer’. Near transfer occurs when the source and target tasks share significant similarities (e.g., learning Latin and Spanish). For example, Pellegrino and Hilton, eds. [4] suggest that learning ‘A’ affects learning ‘B’ only to the extent to which ‘A’ and ‘B’ have common elements. Thus, task similarity is one of the key conditions under which the TL can be demonstrated [5]. By contrast, far transfer involves applying knowledge across substantially different domains, involving the development of abstract principles and cognitive strategies [6]. However, far transfer is challenging to achieve as it requires extensive cognitive processing and abstract reasoning [7]. Woodworth and Thorndike’s [8] identical elements theory proposes that shared common features (i.e., the similarity between the source and the target domains) determine the extent to which learned skills transfer to new skills. Therefore, far transfer may rarely occur [7,9].

In addition to task similarity, the level of abstraction of the learned material, the complexity of the learning material, the degree of original learning, and the learning context also influence the TL [2]. Individual differences in a learner’s prior knowledge and cognitive abilities can also impact the transfer process. Educational strategies that facilitate the TL need to consider these factors. For example, analogical reasoning, where learners identify similarities between previously learned concepts and new ones, can aid in transferring knowledge across domains [10,11]. Cognitive flexibility (the ability to switch between different problem-solving approaches) has also been shown to promote the TL [12]. Additionally, providing multiple examples enhances the learner‘s ability to apply knowledge flexibly in diverse situations [13].

1.1. Challenges in Studying the TL

The concept of the TL has been studied by scholars, educators, and cognitive scientists for over a century [14]. Despite its longstanding presence in academic research, there remains a significant lack of research findings that can effectively aid learners in transferring their knowledge or skills from one context to another [4]. Additionally, understanding the factors hindering or facilitating the TL is still limited [15]. One of the main reasons is that measuring the TL using traditional methods, such as examinations, tests, surveys, or interviews, primarily tests recall and memorization but does not provide insights into how the TL occurs, and does not capture the subtle changes and differences in the cognitive processes underlying learner performance [16]. Second, self-reported data from surveys and interviews may introduce bias due to factors such as personal preferences and cultural background, while examination and test may not be an accurate measure of the TL and actual student progress as well [17,18,19].

Other approaches to studying and measuring the TL include transfer tests and transfer-appropriate processing (TAP). Transfer tests assess how learners apply previously learned knowledge to a new and unfamiliar situation. For instance, in the study by Gick and Holyoak [11], participants were asked to solve a medical problem that involved a tumor treatable by radiation. Some participants were exposed to a similar problem while serving in the military. They were given a hint about using an analogy-based problem-solving strategy and were more successful in transferring their existing knowledge to the new problem to solve it when compared to participants who were not exposed to a relevant prior experience.

TAP provides another perspective on measuring the TL. It suggests that the best memory performance occurs when the cognitive processes during information encoding (initial learning) and information retrieval (assessment for later use) are similar [20]. TAP posits that memory performance depends not just on associating meaning with information (depth of process) but also on the relationship between encoding and retrieval. If the cognitive processes used during encoding align with those during retrieval, memory will accurately recall the relevant information and the TL will occur [21]. However, the effectiveness of TAP as a measure of the TL requires careful instructional or experimental design to ensure that learners engage in similar cognitive processes both when learning and when testing their learning.

1.2. Cognitive Load

Cognitive load theory (CLT), proposed by John Sweller, posits a relationship between cognitive load and learning effectiveness [22]. Cognitive load is the total amount of resources in the brain’s working memory that are used while performing a cognitive task; learning is efficient when cognitive load is aligned with the individual’s cognitive capacity. Therefore, managing cognitive load is essential for successful learning [23].

CLT categorizes cognitive load into three distinct types, including intrinsic, extrinsic, and germane. Intrinsic cognitive load (ICL) is related to the task’s inherent complexity. It refers to the amount of mental effort in learning cognitively demanding content. Perceived CLT is less affected by the instructional design than by the number of interactions, such as between the type of the new information and the learner’s prior knowledge. For instance, learning complex mathematical concepts demands more cognitive resources, regardless of how the material is presented [24]. By contrast, extraneous cognitive load (ECL) depends on how information is presented to learners. If the instructional materials are poorly designed, communicated, and/or organized, ECL increases and may hinder learning. Finally, germane cognitive load (GCL) refers to the effort placed on memory to form new connectivity schemas and how information is processed in the human brain’s long-term memory [23].

Cognitive overload occurs when all three types are present; the learner is overwhelmed with the learning due to the increased mental effort. The learning task overloads the student’s working memory and impedes information retention, and the potential for the TL decreases [25]. Studies have found that a low cognitive load indicates that a learner has the knowledge or expertise related to the task, whereas a high cognitive load often reflects a lack of relevant knowledge or expertise.

Physiological measures using modern neuroimaging techniques have demonstrated the feasibility of using devices such as portable electroencephalogram (EEG) devices to measure objectively cognitive load. For example, EEG devices were used effectively to capture real-time neural activity related to mental workload and learning performance [17,26]. Furthermore, EEG data were used in machine learning (ML) and brain-inspired computational algorithms to evaluate working memory performance by assessing cognitive load [27]. Similarly, EEG data were used to classify learner expertise and compare the performance between novice and expert groups in mathematical learning tasks [28].

1.3. Learning Programming as a Context for Studying the TL

Developing programming knowledge and skills is an essential part of today’s computer science education and professional training. As a knowledge domain, learning a programming language is characterized by a combination of problem-solving tasks that require a practical application of theoretical concepts. However, mastering programming languages is often difficult for learners. This can be attributed to several factors. First, learners need to understand the language’s logic, syntax, and semantics, which can be quite complex [29]. Second, programming requires learners to think logically and systematically, which can be challenging for some learners [30]. In addition, programming languages constantly evolve, which demands staying updated with the latest developments to remain relevant in the job market [31].

The challenges and cognitive demands associated with learning programming languages, and the characteristics of the programming knowledge domain make it a suitable context for studying the TL phenomenon. Research results indicate that the transfer of skills and knowledge within the same domain is more efficient compared to cross-domain transfer [32,33]. Therefore, it may be expected that an observable TL will occur when a learner studies a programming language after having already acquired some programming skills.

1.4. Research Question and Contributions

This study examines the cognitive and neural mechanisms underlying the TL in the context of learning a programming language. According to CLT, working memory load and task performance are inversely related [18]. Therefore, efficient memory use is paramount to good task performance. Memory efficiency determines how well prior knowledge is stored, retrieved, and applied to a new context; efficient memory retrieval is essential for transferring learning, as it allows learners to apply previously acquired knowledge to new problems and solve them effectively. Therefore, the main research question of this study was formulated as follows: ‘What is the effect of prior knowledge on memory efficiency when learning a new programming language?’.

Memory efficiency is a factor influencing learning performance as learning performance quality depends on the cognitive load experienced by the learner (cognitive load is the amount of mental effort required for comprehension) [25]. Higher cognitive load indicates increased mental effort, and tends to hinder retention and understanding; lower cognitive load indicates reduced mental effort and better performance [25]. In particular, this study proposes the following hypothesis:

H1. 

Having prior programming knowledge reduces cognitive load.

The primary objective of this study was to develop and validate a neuroscience-inspired computational approach for identifying cognitive load as a marker of memory efficiency and the TL. The experimental context was selected to explore this in a structured and controlled manner.

This study contributes to the understanding of the TL as a process facilitated by prior programming knowledge using a cognitive computing approach. This study validates the use of cognitive load as an indicator of memory efficiency in the context of studying a programming language. It identifies neural patterns in the alpha and theta waveband frequencies and demonstrates the effect of prior knowledge on cognitive load. Furthermore, this study develops an SNN-based methodology for investigating the TL processes that can be generalized and used in other educational domains.

2. Materials and Methods

2.1. Using an EEG to Collect Brain Activity Data

The human brain is always active, analyzing and responding to internal and external stimulants, learning and changing continuously through learning [34,35]. When learning occurs, the connections between the neurons in the brain (synapses) change. The ability of the brain to create new connections among neurons, causing transformative changes in the internal structure of the existing synapses, is called synaptic plasticity [36,37]. Synaptic plasticity is the biological mechanism for learning and memorizing.

Brain activity data gathered through neuroimaging techniques, such as EEG and functional magnetic resonance imaging (fMRI), can be used to study the changes in neuronal connections and how learning occurs in the brain [16,19].

EEG is a technique for measuring the electrical activity in the brain with high temporal resolution, capturing neural dynamics in milliseconds (ms) [38]. This is essential for studying fast-paced learning processes, enabling the real-time tracking of brain activity changes in response to stimuli in designed tasks [39]. Unlike fMRI, which provides high spatial but lower temporal resolution, EEG is particularly well-suited for capturing immediate changes in neural activity during cognitive tasks, making it an effective tool for analyzing the TL. Additionally, EEG is easier to use and more cost-effective than fMRI. EEG has been widely used in prior research in areas such as health sciences, medicine, and education [38,39,40]. For example, an EEG was used effectively in [41] to measure the changes in brain oscillations associated with memory encoding. In [42], an EEG was successfully used to explore the neural activities underlying motivation-related cognitive processes while solving mathematical problems, revealing specific brainwave patterns and activated brain regions. Due to its advantages and feasibility, EEG was chosen as the data collection method in this study.

Studies involving EEG as a data collection method have utilized brain wave frequency-based features, such as the alpha, beta, and theta power ratios, to measure cognitive workload and to provide insights into a learner’s mental state while performing learning tasks [43,44]. These ratios indicate how the power distribution across different frequency bands reflects cognitive processing demands. According to their findings, the ratios between the alpha and theta power and between the beta and alpha power and several related combinations can be used as the learners’ task engagement and cognitive load indicators [43]. Another study [45] examined the changes in the frequency bands (e.g., the delta wave band); it showed that the variations in the frequency bands correlated with different levels of memory workload. For example, an increase in the alpha waveband and a decrease in the theta waveband indicated good task performance [46].

2.2. Spiking Neural Networks and the NeuCube Architecture for EEG Data Analysis

The TL process relates to both space and time activities in the brain. Spiking neural networks (SNNs) can be used to capture the time–space relationship patterns emerging from neuronal activity data. SNNs were developed as the third generation of artificial neural networks (ANNs). SNNs were inspired by how biological neurons use a series of action potentials (spikes) to communicate and transform information through synaptic changes [47]. SNNs simulate realistically neural processing; they encode information using discrete spike trains, which capture the spatial and temporal features of neuronal activity data [48]. By using spike timing and rate-based computation, SNNs achieve low energy consumption and faster learning, and are highly effective when used to analyze spatio-temporal brain data (STBD), including EEG data, where the precision of the temporal data analysis is important [49,50].

This study uses the SNN architecture NeuCube [51] to analyze experimentally gathered EEG data. NeuCube captures both the spatial and temporal relationships in EEG signals to provide an advanced representation of the neural processes underlying incremental cognition and learning [52,53]. In particular, we used its 3D SNN framework that was specifically designed for modelling spatio-temporal brain data, such as EEG. It includes the following four key stages: (1) encoding, where EEG time-series data are converted into spike trains; (2) mapping, where input neurons (e.g., EEG channels) are placed into a 3D spatial structure based on brain region locations; (3) unsupervised training, where the reservoir (SNN cube, or SNNc) self-organizes via spike-timing-dependent plasticity (STDP), learning temporal, and spatial patterns in the data; and (4) supervised classification, where output neurons are trained using a dynamic evolving SNN (deSNN) to associate spatio-temporal firing patterns with output classes. Refer to the Appendix A for the algorithms, or one of the Figure for the steps.

STDP is a biologically inspired learning mechanism that strengthens the synaptic connections between highly interactive neurons; this way, SNNc learns autonomously spatio-temporal patterns in the data. The process mirrors synaptic adaptation in the brain, where connections are modified based on the timing of neuronal spikes [54,55]. The SNNc is designed to replicate the small-world connectivity structure of the brain and to provide biologically plausible neuronal activity representations [54,56]. It mirrors the networks of densely interconnected brain neurons (particularly in the regions that are functionally or structurally related in which neuronal connectivity is high), and preserves the spatial organization of the EEG channels.

NeuCube’s biological plausibility makes it particularly effective when studying the neural basis of memory efficiency, cognitive load, and the TL. A description of NeuCube’s meta-procedural algorithm and its functional characteristics, along with a Python v.3.10 code of a NeuCube implementation, are available for inspection and use at [57].

2.3. Experiment Design

2.3.1. Research Participants

The research participants comprised 21 males and 5 females, with the majority (22) aged between 18 and 21 years; one was over 40, and three were aged between 22 and 30 years. All participants were university undergraduates enrolled in a first-year computer programming (P1) course where they would learn the C programming language.

The participants completed a screening questionnaire to gather demographic data, and to gauge the type and level of their existing programming knowledge. The purpose of the screening was to exclude the participants with sufficient prior knowledge of the C language, and to ensure that the participants had comparable prior knowledge regarding the number and level of programming languages they had learned.

All 26 participants attended the first session of the experiment, which was held before starting the P1 course. Only 13 participants returned for the second session of the experiment, which was conducted after the P1 course was completed.

2.3.2. Learning Tasks Design and Presentation

The design of the learning tasks was partially inspired by the prior research [16,19]. The 17 programming comprehension tasks were formatted as multiple choice questions. Multiple choice questions were intentionally chosen to minimize body movement during the EEG recordings, ensuring the data quality essential for spatio-temporal neural analysis.

The participants were presented with C programming language code snippets designed to produce an output (see Supplementary Materials). After reviewing the code, the participants were required to select the correct output from four possible answers. This format minimized physical movement during the EEG recording as the participants were not required to type their answers. The learning tasks were organized in order of increasing difficulty, and were presented to the participants in the same predetermined order to ensure all participants worked through the tasks in the same sequence. During the experiment, the tasks were displayed on a laptop monitor. The participants were given unlimited time to solve the programming problems; this reduced time-limit pressure. The process is shown in Figure 1.

An instructional page was presented at the start of each individual experiment, with an OK button signaling the participant to proceed to the first multiple choice question. The participant was instructed to select an answer using a mouse click. Once the answer was selected, a screen displaying the word Relax appeared, followed by the automatic presentation of the next task after a two-second interval. While there was no fixed time limit imposed on the participants for answering the programming tasks, the two-second interval between each task was included to standardize transitions across the experiment. We followed Cohen’s [58] (p. 64) recommendation to provide an intertrial interval of at least 1000 ms between the end of a trial and the start of the next one to ensure adequate temporal separation.

This process continued throughout the 17 programming comprehension tasks. A page signaling the end of the experiment appeared once the participant had answered the last question. The calculated total score was presented and stored for further reference.

The programming tasks were presented using the open-source OpenSesame version 3.3.12 software platform for experiment support [59]. Two event markers were configured for each programming task to indicate the start and end of the task. The end of each task was marked when the participant clicked on their selected answer to the programming question; the start of each question was automatically marked after the two-second relaxation period.

2.3.3. Data Collection Method

For this study, the researcher collected primary data using an EEG device called Emotiv Epoch X [60]. It is a 14-channel EEG device that provides detailed brain activity data with high resolution. The device works with 14 electrodes: AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4. The electrodes are arranged according to the international 10–20 system to ensure standardized data collection from 14 locations on the scalp. The device is designed to detect six frequency bands, which include delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta low (12–16 Hz), beta high (16–25), and gamma (25–45 Hz).

The EEG data were recorded at two different time points, capturing the participants’ brain activity before and after completing the P1 course (Figure 1). The first session (at time point T1—pre-P1 course) was conducted before the start of the 12-week P1 course. The second session (at time point T2—post-P1 course) took place after the course completion. Thus, the study data comprised 39 datasets, 26 datasets that were collected at time point T1, and 13 datasets that were collected at time point T2.

2.4. Data Pre-Processing and Data Sampling

Standard EEG filtering techniques, such as band-pass filtering and artifact removal using independent component analysis (ICA) from MATLAB version R2012b in conjunction with EEGLAB version 8.7 toolbox, were applied to remove noise from the data. A high pass finite impulse response (FIR) filter at 0.5 Hz and a low-pass FIR filter at 45 Hz were applied offline. The channels were visually inspected to ensure that any remaining artefacts were minimal. To remove movement and muscle activity artefacts, ICA was performed on the data using the ‘runica’ algorithm implemented in MATLAB’s EEGLAB toolbox, using the default parameters.

The flowchart in Figure 2 presents the NeuCube processing pipeline used in this study. It includes the four main stages mentioned earlier (data preparation, encoding, mapping, and unsupervised training). Two different methods (average input interaction and neuron proportion) are used in the subsequent analyses. The last two stages (visualization and interpretation) present the outcomes of the unsupervised learning and interpret them in the context of the research question.

Each of the 39 sets of experimentally gathered data was used to generate spatio-temporal samples for input to NeuCube. A collection of 17 samples (each corresponding to one of the questions in the programming) was extracted from each of the 39 datasets. The sample contained the brain activity data gathered in a 5-s interval (3 s before and 2 s after the second marker). As the EEG drive collected the brain activity data every 4 ms, each sample contained data corresponding to 1250 time points. The specific interval described above was chosen as it comprises the important periods immediately before and after the participant makes a decision about the answer to the question in the programming task. Although there are no strict guidelines about the interval length [58], a 5-s interval was deemed appropriate given that the relaxation period after the second marker was 2 s.

The resulting spatio-temporal samples were formatted for input into NeuCube. All samples were saved as CSV files. Each CSV file comprised 1250 rows representing temporal features (timeline points), while the spatial features varied based on the focus of each analysis. In identifying the most active brain regions during the programming tasks (Section 3.1), all 14 spatial features were used, with each column corresponding to one of the 14 EEG channels.

Table 1. The experiment’s meta-procedural algorithm for determining the EEG channels representing the most active brain regions.

Step	Description
1	Convert continuous EEG data into discrete spike trains using the TBR encoding method.
2	Map each of the 14 EEG channels to the corresponding input neurons in the SNNc 3D small-world architecture using Talairach template.
3	Train the SNNc using the STDP learning rule (unsupervised training).
4	Perform spike communication analysis by calculating the average input interaction for 14 EEG channels.
5	Visualize spike activity patterns and identify the most active brain regions during the experimental task.
6	Interpret the results; thicker lines represent higher interaction and spike communication among channels.
7	Use the identified channels to inform subsequent analyses.

shows the meta-procedural algorithm followed.

To investigate the effect of prior programming knowledge on the cognitive load at time points T1 and T2, the spatial features (EEG channels F7 and T7) were represented by six spectral features, corresponding to the EEG frequency bands delta, theta, alpha, beta low, beta high, and gamma.

Table 2. The experiment’s meta-procedural algorithm used for the analyses of the EEG data representing the most active brain regions.

Step	Description
1	Extract power spectrum from channels identified in previous analysis.
2	Convert continuous EEG data into discrete spike trains using the TBR encoding method.
3.1–3.5	For each identified channel: Map each of the six frequency bands to the corresponding input neurons in the SNNc 3D small-world architecture using automatic graph matching. Train the SNNc using the STDP learning rule (unsupervised training). Perform clustering based on neuron proportion analysis which computes the percentage of neurons in the cube that belong to an input neuron cluster. Visualize neuronal activity patterns among all frequency bands. Interpret the patterns related to alpha and theta (wavebands of interest) as indicators of cognitive load and memory efficiency.

shows the meta-procedural algorithm followed.

The EEG frequency bands used in this study—delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), low beta (12–16 Hz), high beta (16–25 Hz), and gamma (25–45 Hz)—were based on the standard frequency ranges supported by the Emotiv EPOC X device. Frequency-domain features were extracted using fast Fourier transform (FFT) applied to the EEG data from channels F7 and T7, centered around the second event marker. The resulting spectra were segmented into the device-defined wavebands, and the mean spectral power within each band was computed, producing a six-dimensional feature vector per trial. No fixed amplitude thresholds were applied; instead, the use of band-wise mean values allowed for a consistent, hardware-aligned representation of relative EEG activity across the trials and participants.

3. Analysis and Results

To test hypothesis H1, this study conducted three analyses of the data collected at time points T1 (pre-course) and T2 (post-course). The first analysis aimed to identify the most active brain regions and their corresponding EEG channels during programming learning tasks. The second analysis compared the brain activity patterns of the participants with or without prior programming knowledge before the start of the course, using the data from the EEG channels that corresponded to the most active brain regions (F7 and T7). The third analysis compared the neuronal change patterns underlying STBD across the two time points T1 and T2. Similar to the second analysis, the models were developed using the data from EEG channels F7 and T7 (representing the most active areas of the brain during the 5-s interval around the decision-making timepoint).

3.1. Identifying the Most Active Brain Regions

This analysis involved the samples derived from the 26 data collected at time point T1. The participants were divided into two groups based on the total score they achieved during the experiment. The participant scores varied between 2 and 11; the median score of seven was used as a threshold. A total of 14 participants scored seven or below; they were categorized as having insufficient prior knowledge (IPK). The 12 participants who scored above seven were classified as having sufficient prior knowledge (SPK). The aim of the analysis was to find out, for each individual IPK and SPK participant, which brain region(s) were most active when the participants were answering the programming task questions.

Each participant’s dataset (e.g., the collection of 17 samples) was processed as follows. First, the samples were loaded into the input encoding module of NeuCube. The continuous input data were converted into discrete spike trains using the threshold-based representation (TBR) encoding algorithm. While several spike encoding methods can be applied to EEG data, the TBR encoding method was selected due to its high temporal resolution, effectiveness in capturing the non-stationary dynamics of EEG signals, and its capacity to suppress noise and compress data. The TBR transforms continuous EEG signals into spike trains by applying adaptive thresholds based on signal fluctuations, thereby preserving meaningful temporal patterns, allowing the encoding to reflect only the meaningful changes in neural activity.

This approach enables precise detection of both excitatory and inhibitory activity by setting higher thresholds for positive changes and lower (or negative) thresholds for inhibitory responses. This study also used step-forward encoding, in which thresholds were automatically calculated based on the current changes in the signal, making it well-suited for dynamic EEG data.

Compared to alternative methods, such as rate coding, which aggregates spike frequency over time and can mask fast-changing neural events, or temporal coding, which may be overly sensitive to jitter and noise, the TBR provides more stable and biologically plausible encoding. The TBR is supported by the NeuCube architecture, and has demonstrated reliable performance in prior spatio-temporal brain data studies. In this study, we adopted commonly used parameter values, with the TBR spike encoding threshold set to 0.5 and the STDP learning rate set to 0.01. These parameters have been used in NeuCube studies involving EEG and spatio-temporal brain data, and have been validated in previous peer-reviewed studies [56,61,62].

For each dataset, the discrete spike trains were initialized by mapping the 14 EEG channels as input neurons into NeuCube’s three-dimensional recurrent SNNc. Next, each resulting SNN model was trained using unsupervised learning to capture the temporal and spatial relationships within the data. During this stage, the temporal relationships in each dataset were adjusted by neuronal connection weights based on the STDP rule. The resulting SNNc comprised 14 input neurons, each corresponding to one of the EEG channels. The input neuron sits at the center of a cluster, surrounded by the neighboring neurons it interacts with mostly. The interactions were visualized graphically using the calculated average input neuron interaction index (based on the average spike exchange of the input neuron with its neighbors during the EEG signal processing (Figure 3). The 14 input neurons in the diagrams correspond to the 14 spatial sampled features (i.e., EEG channels). In Figure 3, the EEG input neurons are arranged in a circular layout, visualizing the average interaction between them. The EEG channels are labeled as follows, starting from the top and moving clockwise: T7, FC5, F3, F7, AF3, AF4, F8, F4, FC6, T8, P8, O2, O1, and P7. The lines connecting the neurons represent the interactions between them. The thicker lines indicate a higher interaction level, meaning more information was exchanged among the two connected neuron clusters.

The models indicated significant input neuron interaction and spike activities in the frontal and temporal lobes of the brain. Across all participants, the highest level of information signal exchange was observed in the brain’s left hemisphere, between the input neurons corresponding to EEG channels F7 and T7.

The EEG channel T7 captures signals from the part of the brain’s left temporal lobe, which is involved in comprehending auditory and visual perceptions (e.g., reading and word recognition), visual linguistic perception, long-term memory, and comprehension [63]. Channel F7 is located in the left frontal lobe of the brain. This part of the brain creates and controls spoken and written language output and is responsible for attention-gaining working memory, and decision-making. These two most active clusters were selected for the subsequent analyses as the activities represented by the EGG data (e.g., comprehension, memorizing, and retrieving information) occur during the TL process.

The diagrams in Figure 3 represent the averaged input neuron interaction index computed by the NeuCube model for each participant. These interaction indexes reflect the strength of spike-based communication between the input neurons and their surrounding clusters, enabling the identification of the most active brain regions.

The visualizations serve to support the rationale for selecting channels F7 and T7 for further analysis. To ensure analytical focus and interpretability, subsequent analyses were restricted to the most active brain areas (F7 and T7) and the alpha and theta frequency bands. These channels demonstrated the strongest input neuron interactions during programming tasks, and the alpha–theta dynamics are widely associated with cognitive load and memory efficiency [46,64]. Analyzing all 14 channels or all six frequency bands would have avoided additional complexity without a clear theoretical or experimental basis for interpreting the results. Therefore, this study prioritized the most relevant spatial and spectral EEG features aligned with the objectives of investigating cognitive load and memory efficiency in the context of the TL.

3.2. Computing Neuronal Activity Patterns at Time Point T1

In this analysis, two comparative studies were carried out. The first comparative study examined the sample data extracted from EEG channel F7 at time point T1 (pre-course). It examined and compared the neural activities represented by the alpha and theta frequency bands of the IPK and SPK participants. The second comparative study explored the EEG channel T7 data at time point T1 (pre-course) in the alpha and theta frequency bands and compared the brain activity patterns of the IPK and SPK participants.

3.2.1. Change Patterns in the Alpha and Theta Frequency Bands in the Brain’s Left Frontal Lobe (EEG Channel F7) at Time Point T1

The EEG channel F7 data were extracted from the sample collections of the 26 IPK and SPK participants. The encoding, initialization, and unsupervised training were carried out following the process described in Section 3.1. However, instead of analyzing the data based on the averaged input neuron interaction, this study used a neuron proportion indicator as a measure of the neuronal relationship between the alpha and theta frequency bands. The value of the neuron proportion represents the percentage of neurons in the SNNc that belong to this input neuron cluster. For any given input neuron cluster, the neuron proportion indicator shows the strength of neuronal interactions and spike activities of the cluster in relation to the other clusters in the SNNc. For example, the NeuCube output diagrams in Figure 4 show the EEG channel F7 neuron proportions of six waveband frequencies (alpha, theta, delta, gamma, beta low, and beta high) extracted for two participant datasets.

The neuron proportions of the theta and alpha waveband frequencies of each of the IPK and SPK groups are shown in

Table 3. Neuron Proportion Values of the IPK and SPK Groups (Timepoint T1, EEG Channel F7).

IPK Group Neuron Proportion (in %)			SPK Group Neuron Proportion (in %)
Participant Dataset ID	Theta	Alpha	Participant Datasets ID	Theta	Alpha
2	23	5	18	18	8
3	19	10	13	12	24
19	13	28	4	1	23
9	14	13	25	12	17
11	10	16	7	9	17
17	19	11	20	18	10
6	10	16	14	13	24
8	12	22	16	5	16
23	17	10	26	3	17
15	12	23	12	11	24
1	10	6	10	7	17
24	12	18	22	12	14
21	23	11
5	17	5
Average	15.07	13.85	Average	10.08	17.58

. The comparison between the groups indicated that, in the IPK group, the average neuron proportion of the theta waveband was 15.07%, while the average neuron proportion of the alpha waveband frequency was lower, at 13.85%. The SPK group showed a different pattern; the average neuron proportion of the theta waveband frequency (10.08%) was much lower than the neuron proportion of the alpha waveband frequency, which was 17.58%. In other words, in the IPK group, the neuronal activity in the alpha waveband was at a lower level compared to the theta waveband; in the SPK group, the neuronal activity in the alpha waveband was at a higher level compared to the theta waveband.

To compare the brain activity patterns of the SPK and IPK participants, we used a within-group paired t-test using the neuron proportion data of each of the two groups of study participants (Test 1, Appendix A,

Table A1. Summary of statistical tests.

Test	Condition	Group	Waveband	Mean (SD)	t-Value	p-Value	Cohen’s d
Test 1 (F7)	Within group (T1)	IPK (n = 14)	Alpha vs. Theta	A: 13.86 (7.03) T: 15.07 (4.60)	0.44	0.66	0.12
		SPK (n = 12)	Alpha vs. Theta	A: 16.25 (7.92) T: 12.92 (6.55)	−2.85	0.016	0.82
Test 2 (F7)	Between group (T1)	IPK vs. SPK	Alpha	IPK: 13.07 (6.58) SPK: 16.25 (7.92)	−1.53	0.14	−0.42
			Theta	IPK: 15.14 (4.67) SPK: 12.92 (6.55)	2.53	0.019	0.38
Test 5 (F7)	Within group (T2)	IPK (n = 8)	Alpha vs. Theta	A: 16.62 (7.01) T: 9.50 (3.85)	4.26	0.004	1.06
		SPK (n = 5)	Alpha vs. Theta	A: 16.60 (3.36) T: 10.50 (0.84)	7.44	0.002	2.83
Test 7 (F7)	Longitudinal (IPK)	T1 vs. T2	Alpha	T1: 13.00 (7.41) T2: 16.63 (6.81)	−1.60	0.153	0.57
			Theta	T1: 15.63 (4.66) T2: 9.50 (3.87)	2.32	0.053	0.82
Test 8 (F7)	Longitudinal (SPK)	T1 vs. T2	Alpha	T1: 15.40 (4.93) T2: 19.60 (3.36)	−1.57	0.159	−1.00
			Theta	T1: 14.80 (5.63) T2: 10.40 (0.89)	1.73	0.156	1.09
Test 9 (T7)	Within group (T1)	IPK(n = 14)	Alpha vs. Theta	A: 16.50 (6.33) T: 13.79 (4.76)	0.99	0.34	0.26
		SPK(n = 12)	Alpha vs. Theta	A: 15.50 (8.05) T: 13.00 (6.90)	0.65	0.53	0.19
Test 10 (T7)	Between group	IPK vs SPK	Alpha	IPK:16.50 (6.33) SPK:15.50 (8.05)	0.35	0.73	0.14
			Theta	IPK: 13.79 (4.76) SPK: 13.00 (6.90)	0.33	0.74	0.13
Test 11 (T7)	Within group (T2)	IPK(n = 8)	Alpha vs. Theta	A: 19.12 (4.32) T: 8.75 (6.45)	3.27	0.014	1.16
		SPK(n = 5)	Alpha vs. theta	A:15.60 (4.98) T: 19.60 (5.32)	−1.14	0.32	−0.51
Test 11A (T7)	Between groups (T2)	IPK vs. SPK	Alpha	IPK: 19.12 (4.32) SPK: 15.60 (4.98)	1.30	0.23	0.77
			Theta	IPK: 8.75(6.54) SPK: 19.50(5.32)	−3.27	0.008	−1.77
Test 12 (T7)	Longitudinal (IPK)	T1 vs. T2	Alpha	T1: 14.88 (9.40) T2: 19.12 (4.32)	1.16	0.27	0.58
			Theta	T1:14.12 T2: 8.75	−1.77	0.10	−0.88
Test 13 (T7)	Longitudinal (SPK)	T1 vs. T2	Alpha	T1: 14.20 (11.82) T2: 15.60 (4.98)	0.24	0.82	0.15
			Theta	T1: 14.80 (8.64) T2: 19.60 (5.32)	1.06	0.33	0.67

). The results indicated that there was no statistical difference between the alpha and theta neuron proportion in the SPK group (p-value = 0.66—not significant; Cohen’s effect size effect d = 0.122—very small). By contrast, there was a statistically significant difference between the alpha and neuron proportions in the IPK group, with one of the wave bands significantly more active than the other (p-value = 0.016—significant; Cohen’s effect size d = 0.82—large). The results indicated that, at time point T1, the two study groups were different in terms of the brain activity patterns, which provided the necessary confirmation of the way the groups were formed.

To find the indices that differentiated the brain activity patterns of the SPK and IPK groups at time point T1, we conducted a between-group unpaired t-test (Test 2, Appendix A, Table A1) comparing the neuron activities for each of the band’s alpha and theta across the two study groups. There was no statistical difference in the alpha neuron proportion between the IPK and SPK participants (p-value = 0.14—not significant; Cohen’s effect size d = −0.42—small). By contrast., there was a statistically significant difference in the theta neuron proportion between the IPK and SPK participants (p-value = 0.019—not significant; Cohen’s effect size d = 0.38—small to moderate). The results indicated that, at time point T1, the brain activity patterns in the theta wave band were the ones that differentiated the two participant groups (IPK and SPK).

3.2.2. Change Patterns in the Alpha and Theta Frequency Bands in the Brain’s Left Temporal Lobe (EEG Channel T7) at Timepoint T1

This study replicated the study presented in Section 3.2.1, using the EEG channel T7 data. The neuron proportions of the alpha and theta waveband frequencies for the individual IPK and SPK participants are shown in

Table 4. Neuron Proportion Values of the IPK and SPK Groups (Timepoint T1, EEG Channel T7).

IPK Group Neuron Proportion (in %)			SPK Group Neuron Proportion (in %)
Participant Dataset ID	Theta	Alpha	Participant Dataset ID	Theta	Alpha
2	19	7	18	17	12
3	10	17	13	1	24
19	12	25	4	15	10
9	7	19	25	13	29
11	12	22	7	16	7
17	9	17	20	22	1
6	12	22	14	13	24
8	11	23	16	1	20
23	14	11	26	8	18
15	13	23	12	23	14
1	23	10	10	13	17
24	12	18	22	14	10
21	22	10
5	17	7
Average	13.78	16.50	Average	13.00	15.50

. The comparison of the averaged neuron proportion values showed that, in the SPK group, the level of neuronal activity in the theta waveband frequency (13%) was lower than that in the alpha waveband frequency (15.5%). The pattern in the IPK group was similar, with average neuron proportions in the alpha and theta wave band frequencies at 13.7% and 16.5%, respectively.

The neuronal activity pattern in the IPK group did not follow the expected pattern of increased activity in the theta waveband frequency and reduced neuronal activity in the alpha waveband frequency. However, the neuron proportion of the theta waveband in the IPK group was still higher than in the SPK group, indicating that the IPK participants experience a higher cognitive load than the SPK participants.

To compare statistically the brain activity patterns of the SPK and IPK participants, we used a within-group paired t-test using the neuron proportion data of each of the two groups of study participants (Test 9, Appendix A, Table A1). The results indicated that, similarly to EEG channel F7, there was no statistical difference between the alpha and theta neuron proportion in the SPK group (p-value = 0.34—not significant; Cohen’s effect size d = 0.26—very small). However, no statistically significant difference between the alpha and neuron proportions in the IPK group was found either, with one of the wavebands significantly more active than the other (p-value = 0.053—not significant; Cohen’s effect size d = 0.19—large). The results indicated that, at EEG channel T7, the brain activity patterns of the participants in the two study groups were similar. Still, in both channels, the average of the neuron proportion in the alpha waveband was higher than the average in the theta waveband. For this reason, the data in T7 were retained for further analyses.

3.3. Comparing Neuronal Activity Patterns at Time Points T1 and T2

The neuron proportion data gathered at time point T2 are presented in the last two columns of

Table 5. Neuron Proportion Comparison at Time Points T1 and T2 for 13 Participants (EEG Channel F7).

Dataset ID	Participant Scores		Neuron Proportion at Time Point T1		Neuron Proportion at Time Point T2
	T1	T2	Theta	Alpha	Theta	Alpha
2	6	11	23	5	4	8
3	4	11	19	10	12	24
19	5	8	13	28	12	24
9	3	8	14	13	12	24
11	6	14	10	16	6	13
17	7	8	19	11	9	17
6	7	10	10	16	6	16
13	9	10	12	22	11	24
25	9	8	12	17	11	17
7	11	10	9	17	10	16
20	8	12	18	10	11	22
21	8	10	23	11	9	19
5	7	11	17	5	15	7

. To compare the brain activity patterns of the SPK and IPK participants after they had completed course P1, we used a within-group paired t-test of the alpha and theta neuron proportions at channel F7 (Test 5, Appendix A, Table A1) and T7 (Test 11, Appendix A, Table A1).

The results for channel F7 indicated that there was a significant statistical difference between the alpha and theta neuron proportion in the SPK group (p-value = 0.004—significant; Cohen’s effect size effect d = 1.06—large) as well as in the IPK group (p-value = 0.002—significant; Cohen’s size effect d = 2.83—large). The results indicated that, at time point T2, the brain activity patterns of the participants in both study groups were similar, with a much larger effect size in the SPK group compared to the IP group. The result was expected, as prior to time point T2 all participants had undergone an intervention that had allowed the acquisition of programming knowledge. The larger effect size in the SPK group indicates a larger difference between the neural activities in the alpha and theta wavebands (compared to the IPK group), indicating a more pronounced effect of prior knowledge on the learning outcome in the SPK group.

The results for channel T7 showed that there was a significant statistical difference between the alpha and theta neuron proportion in the IPK group (p-value = 0.014—significant; Cohen’s effect size effect d = 1.16—large). However, there was no significant statistical difference between the alpha and theta neuron proportion in the SPK group (p-value = 0.32—significant; Cohen’s effect size effect d = −0.51—moderate). The subsequent between-group test (Test 11A, Appendix A, Table A1) indicated that the SPK and IPK groups were differentiated by the level of neuronal activity in the theta wave band, with p-value = 0.008 (significant) and Cohen’s effect size d = 1.77 (very large). This result is incongruent with the result for EEG channel F7, and may be due to factors not considered in this study.

While previous analyses focused on time point T1 (pre-course), this analysis tracked individual participants’ neural activities over time. It identified the change patterns in the alpha and theta wavebands between time point T2 (post-course) compared to time point T1 (pre-course) for each participant. The aim was to assess how these changes aligned with the participants’ programming task scores, and whether the inverse relationship between the alpha and theta waveband frequency found in the time point T1 data (refer to Section 3.2) was manifested at time point T2 as well.

The analysis included two comparative studies, as described below. The first comparative study examined and compared, for each participant, the neural activities represented by the alpha and theta frequency bands extracted from EEG channel F7 at time points T1 and T2. The second comparative study similarly analyzed the data extracted from EEG channel T7.

3.3.1. Change Patterns in the Alpha and Theta Wavebands in the Brain’s Left Frontal Lobe (EEG Channel F7)

The datasets of the 13 individuals who participated in the data collection at both time point T1 and time point T2 were used for this analysis. The data pre-processing and the encoding, initialization, and unsupervised training processes followed the procedures described in Section 3.1. As in the analyses described in Section 3.2, the neuron proportion value (as calculated by NeuCube) was used as a neuronal activity indicator.

The neural activity of each participant at each time point was examined to explore the relationship between the alpha and theta waveband frequencies. In particular, the neuron proportion values related to the theta and alpha wavebands were extracted and compared to identify patterns of change in neural activity from time point T1 to time point T2.

The results indicated a reduction in the theta neuron proportion at time point T2 compared to time point T1 for all participants, except for participant 7 (who demonstrated a slight increase in the theta waveband activity). In the alpha waveband frequency, eight participants exhibited an increase in the alpha neuron proportion at T2, while three participants demonstrated a decrease; three participants displayed no change between the time points. Table 5 shows the neuron proportion values (in percentage) for the alpha and theta wavebands in time points T1 and T2 for each participant.

The participants’ programming task scores were used to investigate the relationship between neural activity change patterns and the knowledge presumably acquired after completing the P1 course. As shown in Table 5, the participants’ programming task scores at time point T2 improved for all the participants except for participants 7 and 25. At time point T2, participant 25’s theta neuron proportion decreased slightly (when compared to time point T1), while the alpha neuron proportion remained the same. For Participant 7, the neuron proportion of the theta waveband frequency increased slightly, while the neuron proportion of the alpha waveband showed a slight decrease.

3.3.2. Change Patterns in the Alpha and Theta Wavebands in the Brain’s Left Temporal Lobe (Channel T7)

This study replicated the study described in the preceding section using the EEG channel T7 data. As shown in

Table 6. Neuron Proportion Comparison at Time Points T1 and T2 for 13 Participants (EEG Channel T7).

Dataset ID	Participants Scores		Neuron Proportion at Time Point T1		Neuron Proportion Time Point T2
	T1	T2	Theta	Alpha	Theta	Alpha
2	6	11	19	7	1	20
3	4	11	24	1	13	10
19	5	8	12	25	9	20
9	3	8	11	24	10	22
11	6	14	12	22	12	24
17	7	18	6	11	1	22
6	7	10	12	22	20	17
13	9	10	1	24	24	16
25	9	8	13	29	23	16
7	11	10	16	7	16	8
20	8	12	22	1	12	22
21	8	10	22	10	23	16
5	7	11	17	7	4	18

, there was a reduction in the neuron proportion of the theta waveband frequency at time point T2 (compared to time point T1) for seven participants. Conversely, there was an increase in the theta neuron proportion for four participants; two participants exhibited no change. Regarding the alpha waveband frequency, at time point T2 there was an increase in the neuron proportion for seven participants (compared to time point T1). For the remaining six participants, a decreased neuron proportion in the alpha waveband frequency at time point T2 was observed (compared to time point T1).

As already noted, the programming task scores at time point T2 were higher than those at time point T1 for all participants except for participants 7 and 25, whose scores were lower (refer to Table 6). At time point T2, participant 25’s theta neuron proportion increased (when compared to time point T1), while the alpha neuron proportion decreased. For Participant 7, the neuron proportion of the theta waveband frequency remained unchanged, while the neuron proportion of the alpha waveband showed a slight increase.

We compared statistically the alpha and theta neuronal activities of the participants in the IPK and SPK groups at time points T1 and T2 for EEG channels F7 and T7 (Tests 7, 8, 12, and 13, Appendix A, Table A1). In the IPK group (channel F7), there was no statistical difference between the neuron proportion in the alpha wave band between time points T1 and T2 (p-value = 0.15—not significant); Cohen’s effect size d = 0.57—moderate.) However, the moderate effect size suggests a meaningful increase in alpha activity at time point T2 in comparison to time point T1.

All eight comparisons provided similar results. For example, the statistical difference between the neuron proportion in the theta waveband (EEG channel F7) was marginally insignificant (p-value = 0.053). Coupled with the large Cohen’s effect size (d = 0.57) indicating a decrease of the activities in the theta wave band. Given the moderate to large effect size values, the results can be interpreted as providing evidence, although limited, for decreased cognitive load at time point T2 (for all participants).

4. Discussion and Conclusions

This study examined TL as a process contextualized through prior knowledge. The main research question of the study was formulated as follows: What is the effect of prior knowledge on memory efficiency when learning a new programming language? To investigate the research question, a working hypothesis H1 was formulated, linking prior knowledge to cognitive load.

A series of experiments were conducted to test hypothesis H1: Having prior programming knowledge reduces cognitive load. The next two sections analyze the experimental results and discuss their implications.

4.1. Prior Knowledge and Cognitive Load

4.1.1. The Effect of Prior Knowledge on Cognitive Load at Time Point T1

To test this hypothesis, we analyzed first the neuronal interactions and spike activity at time point T1 (before the participants started the P1 programming course) to determine which brain areas were most active while the participants were answering the questions in the programming task. We found that, during the time intervals used in this study, the brain’s left hemisphere, particularly the regions corresponding to EEG channels F7 and T7, was more engaged compared to the right hemisphere. This finding aligns with the previous research results about the involvement of the brain’s left frontal and temporal regions in supporting higher-order cognitive processes necessary for programming comprehension. For instance, one study [65] has established that the left inferior frontal gyrus and posterior superior temporal gyrus/sulcus are actively engaged during language comprehension tasks. Similarly, ref. [66] reported increased activity in the left prefrontal cortex during programming tasks.

The neural change patterns in the brain’s left frontal lobe, associated with EEG channel F7 measured at time point T1, indicated that the participants who had significant programming knowledge (the SPK group) and the participants who did not have significant prior programming knowledge (the IPK group) exhibited distinctly different cognitive load patterns, including an increase in cognitive load for the IPK participants (average neuron proportion in the theta and alpha waveband frequencies: 15.07% and 13.85%, respectively) and a decrease in cognitive load for the SPK participants (average neuron proportion in the theta and alpha waveband frequencies: 10.08% and 17.58%, respectively) (refer to Table 3). The established pattern of inverse relationship between the neuronal activities in the alpha and theta waveband frequencies supports hypothesis H1.

The analysis of neural change patterns in the brain’s left temporal lobe, associated with EEG channel T7 measured at time point T1, demonstrated a reduced cognitive load in the SPK group, supporting hypothesis H1. However, the results of the IPK participants provide only partial support for hypothesis H1, as the cognitive load patterns of the IPK group were similar to those of the SPK participants. As shown in Table 4, the average neuron proportions in the theta waveband frequencies in the IPK and SPK groups are 13.78% and 13.00%, respectively. In the alpha waveband frequency, the average neuron proportions of the IPK and SPK groups were 16.50% and 15.50%, respectively.

This inconsistency may be explained by the functional role of the T7 region in the temporal lobe. While the left frontal lobe (the channel F7 region) is highly engaged in executive functions, such as decision-making, attention, and working memory, the temporal lobe is more involved in processing auditory and spoken language signals [67]. Answering the programming task questions requires relatively little processing of such signals; therefore, channel T7 may not capture cognitive load patterns as well as channel F7. Although the neuronal activity pattern in the IPK group did not follow the expected pattern of increased theta and reduced alpha, the comparison of the average neuron proportion in the theta waveband frequency between the IPK and SPK participants showed higher neuronal activity in the IPK group. This finding may suggest that the IPK participants still experienced higher cognitive loads than the SPK participants.

4.1.2. The Effect of Prior Knowledge on Cognitive Load at Time Point T2

The findings suggest that the 12 weeks of learning programming language enhanced the prior programming knowledge of the majority of the 13 participants. As shown in Table 5, the comparison of the brain activities associated with EEG channel F7 for nine participants demonstrated the inverse relationship pattern between the theta and the alpha wavebands, including a decrease (eight participants) or increase (one participant) in the theta waveband frequency and a corresponding increase or decrease in the alpha waveband frequency. The changing pattern was consistent with the difference between the task scores obtained at the two time points. Ultimately, eight participants experienced reduced cognitive load at time point T2, indicating efficient memory use. These findings align with the previous research using cognitive load levels as expertise indicators [25,68].

The remaining four participants demonstrated a reduction in the theta and a reduction or no change in the alpha waveband activities. The programming task scores of three of the respective participants were higher at time point T2, which may suggest that they also experienced a lower cognitive load. However, the time point T2 task score of the fourth participant was lower.

The variations in alpha waveband activity among these participants may be explained by the impact of external factors on the experienced cognitive load. While an increase in alpha waveband activity is associated with relaxed alertness and efficient handling of known information [69], stability or slight activity reductions (particularly when paired with reduced theta activity) may indicate external influences, like emotional stress or attentional shift [70,71]. For example, the participants were very likely to experience attention shifts as the experiment site was not entirely isolated from the noisy campus environment. Overall, the findings of this comparison provide support for hypothesis H1.

The comparison of the neural activity patterns in the brain’s left temporal lobe, associated with channel T7 (refer to Table 6), showed that five participants (datasets) experienced a cognitive load decrease. However, variations were exhibited by the other eight participants, including different combinations of increases and decreases in the activities in the theta and alpha wavebands. As already mentioned, this variation may be attributed to the specific roles of the brain region around EEG channels F7 and T7. The region around channel F7 is involved in decision-making and working memory, which are closely related to cognitive load [72]. By contrast, channel T7 primarily involves auditory and speech processes [67], which may not be directly related to the neural mechanisms underlying cognitive load during programming tasks. Another reason for the inconsistency is related to the nature of the neural oscillations at specific frequency bands related to certain cognitive processes. For example, the theta waveband (4–8 Hz) has been related to the coding phase of short-term memory tasks [64], maintenance [73], information retrieval [74], and differences in memory load [75]. Changes in the alpha range (8–13 Hz), with variable amplitude, have been associated with attention processes and the inhibition of irrelevant information [75,76,77].

4.2. Prior Knowledge and Memory Efficiency

Overall, the analyses above provide evidence that prior programming knowledge reduces the cognitive load when learners engage in programming tasks related to a new programming language, and support hypothesis H1. The brain activity patterns of the participants with prior programming knowledge indicated a state of cognitive readiness and relaxed engagement, facilitating the efficient retrieval of known information (prior knowledge). Given the positive impact of reduced cognitive load on memory efficiency suggested in the prior research [25,65], it can be inferred that, in this context, prior knowledge had a positive effect on memory efficiency. Furthermore, the consistent patterns of reduced theta waveband activity in the participants with prior programming knowledge also support the inference; it was suggested in the prior research [64] that a reduction of neuronal activity in the theta waveband frequency reflects increased memory efficiency. Enhanced memory efficiency is likely to contribute to better learning outcomes and more effective TL, as the reduced intensive cognitive processing allows learners to solve new problems without overloading their working memory

The findings of this study about the inverse relationship between the neuronal activities in theta and alpha wavebands align with the results reported in the literature and, in particular, in the context of completing outcome-driven tasks [78]. In the prior research, theta waveband activity was found to be associated with mental load and the encoding of new information, while the alpha waveband activity was linked to the efficient processing and retrieval of stored information [79]. For example, it was found by [69] that increased theta activity was associated with increased task difficulty and demanding attention control requirements, while increased alpha activity reflected cognitive readiness and better attention control.

To summarize, the findings of this study provide plausible evidence supporting the study hypothesis. They demonstrate that prior knowledge positively influences memory efficiency by reducing cognitive load, as reflected by high alpha and low theta waveband activities. Low cognitive load improves memory efficiency, supporting the better retrieval of prior knowledge. This efficient memory usage facilitates an effective TL process and enhances learning performance. These findings have important implications for computer science education, particularly in addressing the challenges students face when learning programming languages. In particular, the results may motivate educators to design instructional materials that are aligned more closely with the students’ cognitive capacities. Specifically, customized interventions can be developed for students with varying levels of prior knowledge, thereby improving engagement and performance. By aligning course content and teaching strategies with cognitive load requirements, educators can facilitate better learning experiences and outcomes for learning programming languages.

4.3. Comparison with Prior Work and Study Contributions

There is limited research on the use of neuronal activity patterns to study the role of prior knowledge in the TL process. Authors, such as [25], have proposed a method (adapted a subjective survey instrument) to measure cognitive load in introductory computer science courses, focusing on self-reported data during lectures. However, their study did not incorporate objective neurophysiological measures, such as EEG, to assess cognitive load. By contrast, our study utilizes EEG-based measurements to objectively evaluate cognitive load, providing a more direct assessment of the learners’ mental effort during programming tasks.

A similar EPOC device was used in [68] to collect experimental EEG data from the participants who were working on problem-solving programming exercises. The authors computed direct measures of cognition using the alpha and theta wave frequencies and attempted to classify participants according to the level of their programming expertise. By contrast, this study identified the specific neuronal change patterns that signify cognitive load reduction or increase, and thus more or less affect memory use.

While some researchers [66,80] have studied neuronal signals in relation to cognitive load, their work does not explicitly focus on how prior knowledge affects brain dynamics during learning. Ref. [66] reviewed ML techniques for cognitive workload recognition using EEG, and [80] proposed multimodal approaches to enhance measurement accuracy. However, these studies do not explore how cognitive load might serve as an indicator of memory efficiency within a TL study framework. Our study extends this work by identifying waveband-specific neural patterns—particularly in the alpha and theta bands—that reflect the cognitive impact of prior knowledge in programming tasks, offering a novel method to quantify the TL.

Borgheai et al. [79] identified the relationships between activities in the alpha and theta wavebands similar to the ones identified in this study. However, their study was not conducted in an educational context, and did not investigate how cognitive load patterns matched task performance. In addition, the data used in the analyses were collected through a complex brain–computer interaction interface informing EEG and fMRI devices

Other related works [44,65] have explored cognitive absorption and cognitive load in e-learning environments, as well as predictive coding mechanisms across the left temporal and frontal brain regions during language comprehension. However, these studies do not incorporate EEG-based models related to the TL and do not differentiate between learners based on prior knowledge, or identify the specific brain areas involved in processing information during computer programming tasks. By focusing on EEG signals from key brain regions, namely the left frontal and temporal lobes, our study investigates how prior knowledge facilitates the TL, specifically through patterns of change in alpha and theta waveband activity that reveal how the brain applies previously acquired knowledge to new programming tasks.

In summary, while prior research has advanced our understanding of cognitive load and predictive brain functions, few studies have offered a direct approach to measuring the TL in the brain.

This study makes several important contributions:

• This study identifies neural patterns in the alpha and theta wavebands that demonstrate the effect of prior knowledge on cognitive load, and validates their use as a framework for the evaluation of cognitive load in the context of studying a programming language.
• This study proposes and validates the use of cognitive load as an indicator of memory efficiency in the context of studying a programming language.
• By integrating advanced computational models with neuroscience-based techniques, this study introduces a replicable framework for investigating the cognitive effects of prior knowledge. The novel SNN methodology for investigating the TL processes using spatio-temporal neural patterns emerging from the EEG data has the potential to be generalized and used in other educational domains.

4.4. Study Limitations and Directions for Further Research

The major limitation of this study is the relatively small study sample (39 datasets based on data collected from 26 participants. Recruiting research participants was a major challenge in this study due in part to the impact of the COVID-19 pandemic, with some courses still conducted online. Additionally, misconceptions about the EEG device and its capabilities discouraged participation. A larger participant sample would facilitate a more precise group of participants according to their prior knowledge potentially distinguishing better between the neuronal activity patterns of the two groups.

Apart from conducting the same type of experiments on a larger dataset, future research can explore the generalizability of the methodology by applying it to the study of the processes in educational domains. The scope of the research can be expanded to developing SNN models that consider additional cognitive factors, such as stress, motivation, and their impact on the neural activity patterns associated with attention, memory efficiency, and cognitive load. Enabled by the functionalities of the NeuCube platform [47], this approach will lead to developing a more holistic understanding of how internal states and emotional responses impact learning performance and the TL

While this study focused on establishing a reliable neural measure of cognitive load through simple programming tasks, future research can extend this work by incorporating more complex and practical programming scenarios. This would allow for the examination of how cognitive load evolves over time and under realistic problem-solving conditions, further contributing to our understanding of the TL in real-world learning environments. Finally, as the SNNc is designed according to a brain template, the resolution used for the template defines the size of the SNNc. We used the Talairach Daemon softwareas part of NeuCube version 1.3 to extract a template for 1471 neurons, which is practical and large enough for this type of data. As the NeuCube architecture is scalable, from hundreds to millions of spiking neurons, larger templates can be used for larger sets of data with no limitations.