publications

2025

1. 

   Decoding Movement from Neural Spike Trains : A Comparison of Linear and Nonelinear Models across Brain Regions and Temporal Delays

   Minsung Cho, Jaesung Yoo, Stefan M. Lemke, and 2 more authors

   In Conference on Cognitive Computational Neuroscience (CCN), 2025

   Abs PDF Website

   Neural spike trains, which represent the spiking activity of neural population over time, provide critical insights into how the brain encodes information and generates behavior. Despite significant advances, the extent to which these spike trains encode behavioral variables—particularly movement—remains not fully understood. In this study, we compare the performance of linear and nonlinear models in predicting behavior from neural spike trains, focusing on how prediction accuracy varies with different temporal lags between neural activity and movement onset. Furthermore, we examine how prediction performance depends on the specific brain regions from which the neural signals are recorded. Our findings provide new insights into the behavioral decoding with respect to both the temporal structure of neural spike activity and the specificity of brain regions.

2024

1. 

   Toward Structural Similarities between the Brain and Neural Networks

   Minsung Cho and Jay Hoon Jung

   In Conference on Cognitive Computational Neuroscience (CCN), 2024

   Abs PDF Website

   The emergence of artificial neural networks has brought significant advancements to the modeling of intelligence. However, there are still notable differences between their functioning and the structural operations of the brain. From a macroscopic perspective, the brain processes information received from various sensory organs and uses this input to make decisions associated with stored memories. In contrast, artificial neural networks perform cognition and decision-making simultaneously, a structure that reduces the model’s explainability and complicates the implementation of structural divisions. This research presents a framework that distinguishes between cognition and decision structures within artificial neural networks. Experimental analysis using a simplified model is conducted to showcase this distinction.

2023

1. 

   RaCUN : Research for Activation Function Based on Random Sampling to Increase the Robustness of Neural Network

   Minsung Cho, Jae Hyeon Kim, and Jay Hoon Jung

   Proceedings of the Korea Artificial Intelligence Conference, 2023

   Abs

   Since deep neural networks demonstrated their performance in the ImageNet competition, they have been applied across various fields. However, in domains that require high reliability, their adoption has been limited due to the low trustworthiness of neural network models. To address this issue, this study proposes an activation function called RaCUN, which enhances robustness to data noise by applying a sampling technique during training. RaCUN outputs sampled values following a normal probability distribution, thereby introducing nonlinearity while maintaining compatibility with various neural network architectures. Furthermore, experimental results show that models employing RaCUN achieve higher robustness compared to conventional models—improving by approximately 2 percentage points—while maintaining comparable accuracy on the test dataset.