Building artificial intelligence which mimics individual's preferences by combination of deep learning and functional MRI

Deep learning is a subset of machine learning, whereby human-level performance has been achieved in several cognitive tasks. Especially, in the field of computer vision, it has been demonstrated that artificial neural network trained by deep learning has achieved better performance than human. Interestingly, recent studies have revealed that the information structure represented by the trained neural network resembles that of humans (Yamins et al., 2014 PNAS, Güçlü et al., 2015, JNS). Our group works on building artificial intelligence that mimics individiual's preferences by combining deep learning with function MRI.

 

Looking for elementary unit of "mind" by big data analysis on neuroimaging

It is known that we can specify correspondence between brain regions and cognitive functions by appropriately-designed psychological experiments (forward inference). However, it is difficult to infer the mental states or individual's "mind" from brain activity (reverse inference). It's because a single cognitive function is implemented by multiple brain regions while a single brain region is associated with multiple cognitive functions, indicating that brain-function relationship is not one-to-one, rather, multiple-to-multiple (Barret and Satpute 2013; Yarkoni et al., 2011). We suppose that we can apply appropriate reverse inference by revealing the multiple-to-multiple correspondence. For the detailed discussion, please refer our 2017 presentation.

 

Neural correlates of taste in humans

Studies of rodents and monkeys demonstrated that the insular cortex is the center of gustatory processing. However, it still remains unclear where the human gustatory cortex resides. Our group revealed that basic tastes (i.e. sour, bitter, sweet and salty) are discriminated in the insular cortex (Chikazoe et al., Nature Communications, 2019). We will further reveal how human gustatory experience emerges, using machine learning technique.

​

​Development of biomarkers for psychiatric disorders

Diagnosis of mental disorders relies on the doctor's subjective assessment. From the aspect of evidence-based medicine, researchers/doctors have longed for biomarkers to make objective decisions on these.  Many labs have pursued to make classifiers of mental disorders based on resting functional MRI data, however, such classifiers' performance is not sufficient for clinical applications. Our group aims for developing biomarkers applicable to clinical practice with a novel analytical approach.