While neural network architectures are often said to be inspired by the brain, many design choices and optimizations are usually made by researchers. Evolutionary approaches to machine learning attempt to efficiently optimize these architectures through evolution. The cover image shows work by Bertens and Lee, who take a step further towards biological realism by proposing evolvable neural units that can mimic individual neuronal compartments and learning rules. Also in this issue, Tanneberg and colleagues use evolutionary training to learn robust and scalable algorithms.

See Bertens and Lee

Artificial intelligence can be manifested in corporeal and non-corporeal forms. In this issue, Miriyev and Kovač introduce the concept of physical artificial intelligence, which refers to the emerging trend in robotics to create physical systems by co-evolving the body, control, morphology, and actuation and sensing. To support their vision, the authors provide a blueprint for training researchers and establishing institutional environments. In our Editorial, we take a closer look at the history and promise of physical artificial intelligence.

See Miriyev and Kovač and the Editorial.

Neural networks will have limited utility in high-risk environments unless their outputs can be reliably explained. In the cover image, Hasani et al. show how a compact controller inspired by the neural architecture of a roundworm may provide more robust and explainable outputs in a lane-following task. Also in this issue, Jiménez-Luna et al. review how explainable artificial intelligence approaches could aid in drug discovery.

See Lechner et al.

Finding the ground state of a complex material or quantum system is an important task in condensed-matter physics. Mills et al. develop a method they call ‘controlled online optimization learning’ – COOL — based on reinforcement learning, which can find a temperature schedule to anneal and then slowly cool an atomic system with strong interactions, to find its ground state.

See Mills et al.

The cover image shows a miniature robotic manipulator that can be used in minimally invasive microsurgery. Its ‘remote centre of motion’ action, where a mechanism can rotate around a remote fixed point, is made possible by an origami-inspired design. The manipulator attains a positional precision of 26.4 μm and the authors demonstrate its potential utility as a precise tool for teleoperated microsurgery by performing a micro-cannulation procedure under a microscope on an imitation retinal vein. In addition, the device allows gravity compensation and back drivability for safety.

See Suzuki et al.

Identifying the molecular mechanisms that control gene expression is essential for progress in basic and disease biology. Taskaki et al. develop a systems biology model, using deep learning to predict differential gene expression and mine the biological basis of the underlying generative processes.

See Taskaki et al.

In three Comments this issue several groups of experts discuss what role AI can play in the fight against the COVID-19 pandemic. Though AI and machine learning researchers are ready and willing to play their part, it is not an easy task to identify where developments can be most useful. A close collaboration with health workers is required, as well as consideration of how new tools can make a global impact, with adaptability to local situations. One fast-emerging application of machine learning is in data-driven, digital solutions for tracing and tracking COVID-19 infections, but there are alarm bells ringing over the dangers of surveillance creep. In a series of short interviews we delve into the debate about contact track-and-trace apps and the whether it is possible to get the balance right between protecting public health and safeguarding civil rights with digital surveillance tools.

See Luengo-Oroz et al., Peiffer-Smadja et al., Hu et al. and Q&A

Lung cancer has a low survival rate and non-small cell lung cancer (NSCLC) accounts for 85% of lung cancer diagnoses. Mukherjee et al. developed and validated LungNet, an image-based deep learning approach trained on cohorts from various clinical centres, for the pre-diction of overall survival of patients with NSCLC. Their model outputs a risk score that can be used to stratify patients into high- and low-risk categories with regard to overall survival. Using transfer learning, the authors further showed that their survival model can be used for classifying benign versus malignant nodules.

See Mukherjee et al.

4D MRI scans can be used to track cardiovascular blood flow over time, and are important for diagnosing a range of cardiovascular diseases. The cover image in this issue shows blood flow reconstructed from these scans by a deep variational neural network developed by Vishnevskiy and colleagues. This approach may speed up diagnostic workflows, allowing clinicians to view blood flow in close to real-time.

See Vishnevskiy et al.

Neuromorphic chips are designed to use computational machinery inspired by the brain, but it has been challenging to use that machinery in real-world practical problems. In a paper in this issue, Imam and Cleland describe a neural algorithm for the learning and identification of odour samples based on the architecture of the mammalian olfactory system. They implement their neural algorithm in the Intel Loihi neuromorphic system.

See Nabil Imam and Thomas A. Cleland.

The cover image shows superficial upper extremity vein patterns in the human hand and forearm, extracted by a deep neural network under near-infrared light. Such patterns encode rich anatomical information that may be used to develop clinical decision software systems or to guide the supervised autonomous delivery of medical procedures such as vascular access and blood drawing.

See Alvin I. Chen et al.

Machine learning models based on trees are popular non-linear models. But are trees easy to interpret? In real-world situations, models must be accurate and interpretable, so that humans can understand how the model uses input features to make predictions. In a paper in this issue, Scott Lundberg and colleagues propose TreeExplainer, a general method to explain the individual decisions of a tree-based model in terms of input contributions. The utility of tree-based machine learning models for explainable artificial intelligence is further explored in a News & Views by Wojciech Samek.

See Lundberg et al.