Considering the potential for unintended harmful applications of AI tools can lead to deeply concerning findings. An urgent question is how to achieve the right balance between keeping science open and preventing misuse or malicious repurposing.

It is an outstanding challenge to develop intelligent machines that can learn continually from interactions with their environment, throughout their lifetime. Kudithipudi et al. review neuronal and non-neuronal processes in organisms that address this challenge and discuss pathways to developing biologically inspired approaches for lifelong learning machines.

Combinatorial optimization, the search for the minimum of an objective function within a finite but very large set of candidate solutions, finds many important and challenging applications in science and industry. A new graph neural network deep learning approach that incorporates concepts from statistical physics is used to develop a robust solver that can tackle a large class of NP-hard combinatorial optimization problems.

Swarms of microrobots could eventually be used to deliver drugs to specific targets in the body, but the coordination of these swarms in complex environments is challenging. Yang and colleagues present a real-time autonomous distribution planning method based on deep learning that controls both the shape and position of the swarm, as well as the imaging system used for swarm navigation to cover longer distances.

With the availability of a vast and growing number of digital publications, machine reading and other knowledge mining tools, computational methods can be applied at scale to extract insights from the scientific literature. Belikov et al. develop a Bayesian method to mine the biomedical literature that identifies robust scientific findings which could improve the planning of further experiments and scientific investigation.

The invariant causal prediction (ICP) framework tries to determine the causal variables given an outcome variable, but considerable effort is needed to adapt existing ICP methods to the clinical domain. The authors propose an automated causal inference method that could potentially address the challenges of applying the ICP framework to complex clinical datasets.

The use of deep neural networks for the automated analysis of behavioural videos has emerged as a tool in neuroscience, medicine and psychology. Marks and colleagues present a pipeline capable of tracking and identifying animals, as well as classifying individual and interacting animal behaviour in video recordings and even in complex environments.

Combinatorial optimization, the search for the minimum of an objective function within a finite but very large set of candidate solutions, finds many important and challenging applications in science and industry. A new graph neural network deep learning approach that incorporates concepts from statistical physics is used to develop a robust solver that can tackle a large class of NP-hard combinatorial optimization problems.

An important problem in microscopy analysis is to reduce the need for manual annotation of datasets. Bilodeau et al. develop a neural network method that can be trained on a simple main classification task using image-level annotations to solve more complex tasks such as semantic segmentation.

Considering the potential for unintended harmful applications of AI tools can lead to deeply concerning findings. An urgent question is how to achieve the right balance between keeping science open and preventing misuse or malicious repurposing.

Common-sense reasoning has recently emerged as an important test for artificial general intelligence, especially given the much-publicized successes of language representation models such as T5, BERT and GPT-3. Currently, typical benchmarks involve question answering tasks, but to test the full complexity of common-sense reasoning, more comprehensive evaluation methods that are grounded in theory should be developed.

The number of graph neural network papers in this journal has grown as the field matures. We take a closer look at some of the scientific applications.

An international security conference explored how artificial intelligence (AI) technologies for drug discovery could be misused for de novo design of biochemical weapons. A thought experiment evolved into a computational proof.