AI Model Predicts How We Remember Emotional Experiences Using Brain Signals

When people look back on an emotional experience, they rarely judge it by its full duration. Instead, decades of research in cognitive psychology have shown that our overall memory of an event is dominated by two moments: the emotional peak and the way it ends. This phenomenon, known as the peak-end rule, was famously articulated by Daniel Kahneman and colleagues through pain perception experiments, and it has since been confirmed across consumer experiences, everyday well-being, and clinical symptom reporting. A recent study published in the journal Cognitive Computation has now taken this behavioral insight and translated it directly into the architecture of a deep learning system, building an artificial intelligence model that can predict how a person will retrospectively evaluate an emotional experience from their brain activity alone.

The research, conducted by Zhongtang Guo of Duke Kunshan University, addresses a long-standing gap in EEG-based affective computing. Most existing studies of brain-driven emotion recognition have focused on classifying a person’s immediate emotional state at a given moment. Far less attention has been paid to predicting the summary judgment a person forms after an entire emotional sequence has unfolded. Yet it is precisely this retrospective evaluation that shapes clinical symptom recall, customer satisfaction, and everyday judgments of happiness or distress. The new work proposes that the neural signatures of the peak-end effect can be captured in event-related potentials, or ERPs, the millisecond-scale electrical ripples the brain produces in response to each stimulus.

To gather the necessary data, thirty right-handed participants aged 18 to 28 completed a sequential emotion induction paradigm built from the International Affective Picture System. Each trial presented six images in a row, arranged along a preset arousal curve so that each sequence contained a clear emotional peak and a defined endpoint. The researchers systematically manipulated where the peak occurred within the sequence and how intense the final image was, using a Latin square design to cross these variables fully. After each sequence, participants rated their overall experience on a nine-point scale, and these ratings became the training targets for the predictive model. Brain activity was recorded from a 64-channel EEG cap at a sampling rate of 1000 Hz, with rigorous preprocessing including independent component analysis to remove ocular, muscular, and cardiac artifacts.

Three ERP components formed the neural backbone of the system. The Early Posterior Negativity, appearing roughly 150 to 350 milliseconds after stimulus onset, indexes early attentional capture by emotionally salient material. The P300, measured between 300 and 500 milliseconds over central-parietal electrodes, reflects cognitive evaluation and stimulus categorization. The Late Positive Potential, or LPP, spanning roughly 400 to 800 milliseconds over midline parietal sites, is widely regarded as the core electrophysiological marker of emotional arousal and motivational salience. The study confirmed that all three components were significantly modulated by arousal intensity, with high-arousal negative images producing the largest LPP amplitudes, and the peak of each sequence was operationally defined as the image evoking the maximum absolute LPP amplitude.

On top of these neural features, the researchers built a hybrid deep learning architecture called TCN-Attention-PeakEnd Gate, abbreviated TAPE. The model begins with a temporal convolutional network, or TCN, whose dilated causal convolutions expand the receptive field exponentially across layers, allowing the network to capture long-range temporal dependencies without the vanishing-gradient problems of recurrent networks. The TCN output then passes through an eight-head multi-head self-attention module, which models global relationships among all time steps in the sequence. The distinctive element, however, is the peak-end gating module. This module takes the ERP feature vectors of the peak and endpoint moments and computes a learnable sigmoid gating signal that adaptively reweights the contribution of each time step. In effect, the peak-end rule is embedded into the network as a differentiable operation, constraining the model to mirror the cognitive bias that human memory actually exhibits.

The performance results were striking. Under leave-one-subject-out cross-validation, the most demanding evaluation scheme in EEG research, in which each participant serves as the test subject exactly once, TAPE achieved a mean absolute error of 1.038 rating points, a Pearson correlation of 0.654 between predicted and actual retrospective ratings, and a three-level classification accuracy of 70.4 percent. These figures significantly outperformed eight baselines spanning shallow regression, classical EEG feature engineering, and deep architectures including LSTM, CNN-LSTM, and Transformer models, with all differences confirmed by Holm-Bonferroni-corrected paired-sample t-tests. Across fifty independent fivefold cross-validation evaluations, TAPE simultaneously attained the highest median correlation of 0.682 and the smallest interquartile range of 0.054, indicating that its advantage was stable rather than an artifact of favorable data splits.

External validation on the publicly available SEED dataset replicated the model’s superiority, where TAPE reached 84.7 percent three-class accuracy compared with 81.9 percent for the strongest deep-learning baseline, a Transformer. An ablation study then dissected the contribution of each module. Removing the peak-end gating module increased error by roughly 11 percent, removing self-attention reduced the correlation by nearly 10 percent, and replacing the TCN encoder with an LSTM cost 6.3 percentage points of accuracy, confirming that each component makes an independent and synergistic contribution to performance.

Perhaps the most scientifically compelling findings came from the interpretability analysis. When the researchers visualized the attention weights learned by the trained model, they discovered that under high-arousal conditions the model automatically assigned its highest attention to the sequence positions where peak stimuli most frequently occurred, precisely as peak-end theory predicts. Under neutral conditions, where no salient peak existed, attention became nearly uniform across positions, consistent with the theoretical corollary that retrospective evaluation reverts to an averaging strategy when arousal variability is low. The model also revealed a previously underexplored valence asymmetry: endpoint attention weights were higher for positive sequences than negative ones, hinting at a neural basis for the everyday wisdom of ending experiences on a high note. Correlation analysis further showed that the gating signals tracked LPP amplitude more strongly than the other ERP components, with a correlation of 0.483, providing computational-level evidence that sustained emotional elaboration carried by the LPP is the primary neural signal underlying peak-end integration.

The implications reach well beyond the laboratory. In clinical psychology, retrospective symptom assessments such as the PHQ-9 depression questionnaire are known to be vulnerable to peak-end recall bias, and a theory-constrained decoding system of this kind could eventually support wearable-EEG monitoring that quantifies and corrects such distortions in depression, anxiety, and post-traumatic stress. In user experience research, combining EEG with peak-end gating offers a more objective way to identify the key moments that dominate a customer’s lasting impression than traditional questionnaires ever could. The authors also outline a roadmap for future work, including multi-center validation on larger and more diverse cohorts, real-time implementation on edge devices through model compression, multimodal integration with heart rate, skin conductance, and eye-tracking, and longitudinal clinical translation. The study’s broader methodological message may prove its most durable contribution: embedding mature cognitive theories directly into neural network architectures as differentiable constraints can simultaneously improve predictive accuracy, stability, and interpretability, offering a template for a new generation of theory-guided affective computing systems.

Subject of Research: A deep learning framework that predicts retrospective evaluations of emotional experiences from ERP neural markers of the peak-end effect

Article Title: A Deep Learning-Based Retrospective Evaluation Prediction System for Emotional Experiences: Temporal Dynamic Feature Extraction and ERP Neural Mechanisms of the Peak-End Effect

Article References: Guo, Z. (2026). A Deep Learning-Based Retrospective Evaluation Prediction System for Emotional Experiences: Temporal Dynamic Feature Extraction and ERP Neural Mechanisms of the Peak-End Effect. Cognitive Computation, 18(1), Article 111. https://doi.org/10.1007/s12559-026-10657-9

Image Credits: AI Generated

DOI: 10.1007/s12559-026-10657-9

Keywords: peak-end effect, deep learning, EEG, event-related potentials, emotion recognition, temporal convolutional network, self-attention, LPP, retrospective evaluation, affective computing, Cognitive Computation, brain-computer interface