Articles in 2020

Bertens and Lee propose an evolvable neural unit, a recurrent neural network-based module that can evolve individual somatic and synaptic compartment models of neurons. By constructing networks of these evolvable neural units, they can evolve agents that learn synaptic update rules and the spiking dynamics of neurons.

A set of predictive models can exist that predict equally well; however, the specific variables underlying these models may be important to some of them but not to others. Jiayun Dong and Cynthia Rudin demonstrate a method to visualize and quantify this effect of variable importance.

There is much interest in ‘explainable’ AI, but most efforts concern post hoc methods. Instead, a neural network can be made inherently interpretable, with an approach that involves making human-understandable concepts (aeroplane, bed, lamp and so on) align along the axes of its latent space.

The wealth of data generated by single-cell RNA sequencing can be used to identify gene sets across cells, as well as to identify specific cells. Lukassen and colleagues propose a method combining matrix factorization and variational auto encoders that can capture both cross-cell and cell-specific information.

Inspired by many examples in nature where organisms change shape to concur environments, there is much interest in designing robots that are capable of shape change. Shah et al. demonstrate a method for automatically discovering shape and gait changes for soft robots that can adapt to different terrains.

Autonomous drones can help find injured or missing people when a large or hard to traverse area has to be searched, but their view can be obscured in dense forests. David Schedl and colleagues have developed a method to reveal humans in thermal imaging recordings, even in the presence of dense foliage.

Neoantigens play a critical role in cancer immunotherapy. Tran et al. show how training a personalized deep learning model for each individual patient can improve the accuracy and identification rate of mutated neoantigens.

A hallmark of intelligent behaviour is the ability to learn abstract strategies that can be transferred across different tasks, but it has been challenging to incorporate this ability in artificial systems. The authors present a modular architecture for the learning of algorithmic solutions, and demonstrate generalization and scaling on 11 diverse algorithms.

Microrobots are usually too small to contain traditional computing substrates that could control their behaviour. Dekanovsky and colleagues have developed a microrobot swarm that removes hormonal pollutants when it senses a chemical signal in its environment.

Metal–organic frameworks (MOFs) are attractive materials for gas capture, separation, sensing and catalysis. Determining their water stability is important, but time-intensive. Batra et al. use machine learning to screen water-stable MOFs and identify chemical features supporting their stability.

The thickness of the retina is an important medical indicator for diabetic retinopathy. Holmberg and colleagues present a self-supervised deep-learning method that uses cross-modal data to predict retinal thickness maps from easily obtainable fundus images.

The wealth of data gathered from single-cell RNA sequencing can be processed with deep learning techniques, but often those methods are too opaque to reveal why a single cell is labelled to be a certain cell type. Lifei Wang and colleagues present an RNA-sequencing analysis method that uses capsule networks and is interpretable enough to allow for identification of cell-type-specific genes.

Across disciplines, there is a rising interest in interpreting machine learning models to derive scientific knowledge from data. Genkin and Engel show that models optimized for predicting data can disagree with the ground truth and propose a new model selection principle to prioritize accurate interpretation.

Neural network models can predict the socioeconomic wealth of an area from aerial views, but fall short of explaining how visual features trigger a given prediction. The authors develop a pipeline for projecting class activation maps onto the underlying urban topology, to help interpret such predictions.

To infer a previously unknown molecular formula from mass spectrometry data is a challenging, yet neglected problem. Ludwig and colleagues present a network-based approach to ranking possible formulas.

Inspired by the brain of the roundworm Caenorhabditis elegans, the authors design a highly compact neural network controller directly from raw input pixels. Compared with larger networks, this compact controller demonstrates improved generalization, robustness and interpretability on a lane-keeping task.

Magnetic endoscopes have the potential to improve access, reduce patient discomfort and enhance safety. While navigation of magnetic endoscopes can be challenging for the operator, a new approach by Martin, Scaglioni and colleagues explores how to reduce this burden by offering different levels of autonomy in robotic colonoscopy.

Recent advances have increased the dimensionality and complexity of immunological data. The authors developed a machine learning approach to incorporate prior immunological knowledge and applied it on clinical examples and a simulation study. The approach may be useful for high-dimensional datasets in clinical settings where the cohort size is limited.

Classifying cells from single-cell RNA sequences is challenging for cells where only limited data is available. Hu and colleagues show here that a clustering approach using transfer learning can use the variability of one dataset to cluster a smaller target dataset with high-quality results.

Advances in large-scale connectivity mapping of the brain require efficient computational tools to detect fine structures across large volumes of images, which poses challenges. The authors introduce a hybrid architecture that incorporates topological priors of neuronal structures with deep learning models to improve semantic segmentation of neuroanatomical image data.