Explainable Deep Learning Methods for Neuroscience Data to Analyze the Extracted Features in The Hidden Layers رسالة ماجستير

Abstract

In recent years, deep learning models have revolutionized neuroscience by uncovering the complex layers of hidden information. However, unraveling the meaning of features within these hidden layers remains a challenge. This study, titled "Examining the Features in Hidden Layers of Deep Learning Models Applied to Neuroscience Data" employs explicable deep learning techniques to elucidate the decision-making processes of convolutional neural networks (CNNs) when applied to neuroscience data. Understanding hidden layers in deep learning models, especially in applied applications such as connectivity analysis in neuroimaging, is challenging due to their opaque representations, non-linear transformations, and increased dimensionality. In such cases, this opacity hinders the interpretation of neural representations learned by the model, limiting insights into brain function. Balancing model complexity is crucial to accurately capturing meaningful patterns in neuroscientific data. Overcoming these challenges requires specialized interpretability tools and techniques to unravel hidden layer representations and gain deeper insights into neural systems and their computational mechanisms. When analyzing Magnetoencephalography (MEG) data with six frequency bands as important features, the complexity of MEG connectome images hinders the visual inspection of Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) results. Undeterred, the study examines high-accuracy CNN models for dogs, cats, and the MNIST dataset, showcasing the efficacy of LIME and SHAP. This is VI achieved through visual inspection and mathematical equations for SHAP values. Crucially, the research aims to demystify the traditionally elusive hidden layers in CNN models, offering insight into their decision-making processes and enhancing overall model transparency . Across the MEG dataset, the study identifies Gamma1 as having the highest SHAP values, indicating a significant influence on CNN predictions. This nuanced understanding contributes to interpreting the decision-making process of the model and provides insight into the hierarchical influences between frequency bands within the MEG dataset. In conclusion, this study addresses the challenge of interpretability of neuroscience data by unlocking the black box of hidden layers, thereby promoting informed decision making in neuroscience applications.