The technical achievements of the cooperation between the National University of Singapore, ByteDance and other institutions were recently published in the neurobiological journal "Nature Neuroscience". Reliable AI models are trained on medical data to improve the effect of precise medical treatment based on brain imaging.
Brain imaging technology can directly observe the neurochemical changes of the brain when processing information and responding to stimuli. In theory, AI models based on brain imaging can be applied to predict some representative characteristics of individuals, thereby promoting precision medicine for individuals. Despite the existence of large-scale human neuroscience datasets such as the UK Biobank, small-scale data samples of tens to hundreds of people are still the norm when studying clinical populations or addressing key neuroscience problems. Therefore, how to train a reliable AI model is becoming a focal issue in the fields of neuroscience and computer science when the amount of accurately labeled medical data is limited.
The researchers propose to use meta-learning in the field of machine learning to solve the above problems.
Meta-learning is one of the most popular learning methods in the past few years, and its goal is to allow models to quickly learn new tasks based on the acquired knowledge.
The researchers' analysis of previous small samples of data found an intrinsic correlation between an individual's representational characteristics, such as cognition, mental health, demographics and other health attributes, and brain imaging data. Based on this correlation between small sample data and large data sets, researchers propose a meta-matching method, which transfers the machine learning model trained on the large data set to the small data set, so as to train the more reliable model.
The new method, which has been evaluated on datasets from the UK Biobank and the Human Connectome Project, shows higher accuracy than traditional methods.
Experiments show that this new training framework is very flexible and can be combined with any machine learning algorithm, and can effectively train AI prediction models with good generalization performance on small-scale datasets.
(Original title, "AI Meta-Learning Enters Neuroscience for the First Time")
Brain imaging technology can directly observe the neurochemical changes of the brain when processing information and responding to stimuli. In theory, AI models based on brain imaging can be applied to predict some representative characteristics of individuals, thereby promoting precision medicine for individuals. Despite the existence of large-scale human neuroscience datasets such as the UK Biobank, small-scale data samples of tens to hundreds of people are still the norm when studying clinical populations or addressing key neuroscience problems. Therefore, how to train a reliable AI model is becoming a focal issue in the fields of neuroscience and computer science when the amount of accurately labeled medical data is limited.
The researchers propose to use meta-learning in the field of machine learning to solve the above problems.
Meta-learning is one of the most popular learning methods in the past few years, and its goal is to allow models to quickly learn new tasks based on the acquired knowledge.
The researchers' analysis of previous small samples of data found an intrinsic correlation between an individual's representational characteristics, such as cognition, mental health, demographics and other health attributes, and brain imaging data. Based on this correlation between small sample data and large data sets, researchers propose a meta-matching method, which transfers the machine learning model trained on the large data set to the small data set, so as to train the more reliable model.
The new method, which has been evaluated on datasets from the UK Biobank and the Human Connectome Project, shows higher accuracy than traditional methods.
Experiments show that this new training framework is very flexible and can be combined with any machine learning algorithm, and can effectively train AI prediction models with good generalization performance on small-scale datasets.
(Original title, "AI Meta-Learning Enters Neuroscience for the First Time")
Comments