An attention-based fuzzy CNN-LTSM network for visual object recognition from fMRI images.

Decoding visual stimuli is an increasing challenge in the field of neuroscience. It is hoped to classify the patterns of brain activity that occur when people look at visual stimuli. This paper introduces a new method for detecting visual objects from functional magnetic resonance imaging (fMRI) data using an attention-guided convolutional neural network (CNN) with a long short-term memory (LSTM) model. The new model is end-to-end trained for visual object recognition. In CNNs, the method proposed receives frequency and spatial information, whereas the LSTM network learns temporal features. But the fully connected network and a fuzzy neural block (FNB) are included to examine the classification results. The method is verified through performance using the DS105, DS107, and DS116 databases, and two approaches of hyperparameter tuning, Bayesian optimization and the coordinate descent algorithm, are compared. The introduced architecture, an FNB, utilizing Bayesian optimization to tune shows more than 97% recognition accuracy rates on all databases, demonstrating superior classification accuracy to the state of the literature. The superior performance of the proposed model along with fuzzy units’ generalizability to other classifiers suggests promising enhancements for visual object recognition from fMRI data.