Deep Learning Model Reveals Mechanisms of Semantic Satiation

Semantic satiation, the phenomenon where a word or phrase loses its meaning after being repeated many times, is a well-known psychological phenomenon. However, the micro-neural computational principles underlying this mechanism remain unknown. This paper uses a continuous coupled neural network (CCNN) to build a deep learning model to study the mechanisms of semantic satiation and precisely describe this process through neuron components. The research results indicate that, from a mesoscopic perspective, semantic satiation might be a bottom-up process, differing from the top-down process suggested by existing macroscopic psychological studies. The simulation in this study uses an experimental paradigm similar to classic psychological experiments, observing similar results. The saturation of semantic targets is akin to the learning process of object recognition in our network model, depending on continuous learning and switching of objects, with neural coupling enhancement or attenuation affecting the saturation. In summary, both neural and network mechanisms play a role in controlling semantic satiation.

Research Background

In daily life, have you ever pondered a linguistic entity for so long that its semantic essence started to become vague and retreat? For instance, when repeatedly contemplating the word “cat,” this word, traditionally representing a cute domestic feline, might generate a strange feeling of detachment. This phenomenon is not limited to the language domain; in fact, it has been observed in various experiments with different protocols and techniques. In recent years, with the development of more advanced methods, this phenomenon has been continuously studied, uncovering new biomarkers.

Paper Source

This research was jointly completed by multiple research institutions, including Northwestern Polytechnical University and Lanzhou Jiaotong University. The relevant research results were published in the 2024 edition of the journal “Communications Biology.”

Full Research Details

Research Process

Procedure 1: Experimental Design and Verification of Semantic Satiation

The first step of the research was to design an experimental paradigm to simulate classic psychological experiments on semantic satiation by continuously inputting the same visual stimuli. This part of the research used the MNIST and Fashion-MNIST datasets, which contain images of handwritten digits and fashion items, respectively, as test objects for input.

Procedure 2: Specific Implementation of the Model

The study used a continuous coupled neural network (CCNN) to construct an artificial neural network model to simulate the behavior of neurons in the primary visual cortex. The model consists of multiple layers, gradually processing and classifying visual signals by setting different parameters.

Procedure 3: Model Simulation and Data Processing

To simulate complex neuronal behaviors, the model outputs were treated as electrophysiological signals. These signals were compared with brain waveforms recorded in classic experiments to verify their validity. Specific experiments included classification tasks with continuous input of the same stimulus and input of similar but different stimuli to examine the model’s performance during the semantic satiation process.

Research Results

Experimental Result 1: Semantic Satiation Induced by the Same Stimulus

The study found that with increased model time, the classification accuracy initially rose and then declined, a pattern consistent with the phenomenon of semantic satiation observed in psychological experiments. Similarly, changes in the size of the receptive field did not fundamentally affect the overall trend, validating the model’s effectiveness in simulating the semantic satiation process.

Experimental Result 2: Semantic Satiation Induced by Similar Stimuli

For classification tasks, input stimuli were presented with high, low, and no correlation. This part of the experiment also used the MNIST and Fashion-MNIST datasets. The study found significant differences in classification accuracy for inputs with varying degrees of correlation with the initial stimulus 1, which aligns with psychological experiments where participants respond differently to stimuli of different categories.

Visualization of Intermediate States of Visual Information Processing

The experiment further visualized the changing output features of image signals in the model over time to observe the phenomenon of semantic satiation. As time progressed, the input signals became increasingly blurred, especially the semantically related parts, which became larger and more obscure. This phenomenon also affected subsequent image processing and classification accuracy.

Research Conclusions and Value

The study reveals the neural mechanisms behind the phenomenon of semantic satiation, proposing that its essence is a bottom-up process, challenging traditional psychological views of this phenomenon. Through simulation results, the research provides a new paradigm and numerical reference for psychological research and proposes a foundation for verification with real neuron experiments.

Research Highlights and Value

• Novelty: The study is the first to reveal the neural mechanisms of semantic satiation from a mesoscopic perspective.
• Method Innovation: The use of continuous coupled neural networks to simulate neuron behavior in the primary visual cortex demonstrates the effectiveness of this method in simulating the semantic satiation process.
• Research Paradigm: The study provides a novel numerical paradigm for psychological research, potentially opening new research directions in brain science and computational neuroscience.

References

This research cites a large number of related field references, which can be found in the online version of the journal “Communications Biology.”

This study provides a new explanation for the long-standing perplexing phenomenon of semantic satiation in psychology and neuroscience and offers new directions for future research and applications.