Machine learning, practically speaking

ML projects might involve training a system to find and classify patterns indicative or predictive of disease in images or gene expression data, to predict protein structures from genetic sequence or to design chemical scaffolds in drug discovery. MIT computer scientist Regina Barzilay likes seeing how popular and modular deep learning frameworks for building ML systems, such as PyTorch or Google's TensorFlow, have become. "Now you have the big Lego blocks and you can put it together," she says. Collaborating with computer scientists is still advisable to better understand what the system does, "but you can start using some of these methods even though you are not expert in them," says Christos Davatzikos of the University of Pennsylvania Perelman School of Medicine. But Barzilay sees some biomedical researchers try AI, make big claims that don't materialize and then turn their backs on these methods.