Volume 4 Issue 1, January 2022

For a third year in a row, we followed up with authors of several recent Comments and Perspectives in Nature Machine Intelligence about what happened after their article was published: how did the topic they wrote about develop, did they gain new insights, and what are their hopes and expectations for AI in 2022?

Although artificial reinforcement learning agents do well when rules are rigid, such as games, they fare poorly in real-world scenarios where small changes in the environment or the required actions can impair performance. The authors provide an overview of the cognitive foundations of hierarchical problem-solving, and propose steps to integrate biologically inspired hierarchical mechanisms to enable problem-solving skills in artificial agents.

Zeroth-order optimization is used on problems where no explicit gradient function is accessible, but single points can be queried. Hoffman et al. present here a molecular design method that uses zeroth-order optimization to deal with the discreteness of molecule sequences and to incorporate external guidance from property evaluations and design constraints.

To train machine learning models for medical imaging, large amounts of training data are needed. Zhou and colleagues instead propose a method of weak supervision which uses the information of radiology reports to learn visual features without the need for expert labelling.

Neural networks have become a useful approach for predicting biological function from large-scale DNA and protein sequence data; however, researchers are often unable to understand which features in an input sequence are important for a given model, making it difficult to explain predictions in terms of known biology. The authors introduce scrambler networks, a feature attribution method tailor-made for discrete sequence inputs.

Routine eye clinic imaging could help screen patients with cardiovascular risk as studies indicate strong associations between biomarkers in the retina and the heart. This potential is supported by a multimodal study, employing a deep learning model, that can infer cardiac functional indices based on retinal images and demographic data.

In artificial neural networks, a typical neuron generally performs a simple summation of inputs. Using computational and electrophysiological data, the authors show that a single neuron predicts its future activity. Neurons that predict their own future responses are a potential mechanism for learning in the brain and neural networks.

Reinforcement learning has shown remarkable success in areas such as game-playing and protein folding, but it has not been extensively explored in modelling cell behaviour. The authors develop an approach that uses deep reinforcement learning to uncover collective cell behaviours and the underlying mechanism of cell migration from 3D time-lapse images of tissues.

Piezoresistors can be used in strain sensors for soft machines, but the traditional design process relies on intuition and human ingenuity alone. Haitao Yang and colleagues present a method built on genetic algorithms and other machine learning methods to design and fabricate strain sensors with improved capabilities.