Articles in 2021

Spiking neural networks promise fast and energy-efficient information processing. The ‘time-to-first-spike’ coding scheme, where the time elapsed before a neuron’s first spike is utilized as the main variable, is a particularly efficient approach and Göltz and Kriener et al. demonstrate that error backpropagation, an essential ingredient for learning in neural networks, can be implemented in this scheme.

Incorporating prior knowledge in deep learning models can overcome the difficulties of supervised learning, including the need for large amounts of annotated data. An approach in this area called deep reasoning networks is applied to the complex task of mapping crystal structures from X-ray diffraction data for multi-element oxide structures, and identified 13 phases from 307 X-ray diffraction patterns in the previously unsolved Bi-Cu-V oxide system.

The relationship between brain organization, connectivity and computation is not well understood. The authors construct neuromorphic artificial neural networks endowed with biological connection patterns derived from diffusion-weighted imaging. The neuromorphic networks are trained to perform a memory task, revealing an interaction between network structure and dynamics.

Deep learning has revolutionized image analysis, but applying it to voluminous and multimodal images can be challenging. The authors propose a neural network approach with a modality sampling strategy and an attention module for segmentation in fluorescence microscopy images.

Radiomics has been used to discover imaging signatures that predict therapy response and outcomes, but clinical translation has been slow. Using machine learning methods, the authors report tumour subtypes that are applicable across major imaging modalities and three cancer types. The tumour subtypes have distinct radiological and molecular features, as well as survival outcomes after conventional therapies.

Auction games present an interesting challenge for multi-agent learning. Finding the Bayes Nash equilibria for optimum bidding strategies is intractable for numerical approaches. In a new, deep learning approach, strategies are represented as neural networks, and policy iteration based on gradient dynamics in self-play enables learning of local equilibria.

Molecular simulations informed by experimental data can provide detailed knowledge of complex biomolecular structure. However, it is a challenging task to weight experimental information with respect to the underlying model. A self-adapting type of dynamic particle swarm optimization can tackle the parameter selection problem, which is demonstrated on small-angle X-ray scattering-guided protein simulations.

Particle image velocimetry is an imaging technique to determine the velocity components of flow fields, of use in a range of complex engineering problems including in environmental, aerospace and biomedical engineering. A recurrent neural network-based approach for learning displacement fields in an end-to-end manner is applied to this technique and achieves state-of-the-art accuracy and, moreover, allows generalization to new data, eliminating the need for traditional handcrafted models.

Deep learning-based methods to generate new molecules can require huge amounts of data to train. Skinnider et al. show that models developed for natural language processing work well for generating molecules from small amounts of training data, and identify robust metrics to evaluate the quality of generated molecules.

Methods are available to support clinical decisions regarding adjuvant therapies in breast cancer, but they have limitations in accuracy, generalizability and interpretability. Alaa et al. present an automated machine learning model of breast cancer that predicts patient survival and adjuvant treatment benefit to guide personalized therapeutic decisions.

With edge computing on custom hardware, real-time inference with deep neural networks can reach the nanosecond timescale. An important application in this regime is event processing at particle collision detectors like those at the Large Hadron Collider (LHC). To ensure high performance as well as reduced resource consumption, a method is developed, and made available as an extension of the Keras library, to automatically design optimal quantization of the different layers in a deep neural network.

Single-cell RNA sequencing efforts have made large amounts of data available for transcriptomics research. Simon and colleagues develop a neural network embedding approach that avoids batch effects, such that it can rapidly and efficiently integrate large datasets from different studies.

The response of the body to drugs follows complex dynamical processes that can be difficult to predict. Lu and colleagues combine a neural network approach with pharmacokinetic/pharmacodynamic modelling to learn these complex dynamics.

Neural networks are becoming increasingly popular for applications in various domains, but in practice, further methods are necessary to make sure the models are learning patterns that agree with prior knowledge about the domain. A new approach introduces an explanation method, called ‘expected gradients’, that enables training with theoretically motivated feature attribution priors, to improve model performance on real-world tasks.

Monoclonalization, the isolation and expansion of a single cell derived from a cultured population, is an essential step in large-scale human cell culture and experiments. A new deep learning-based workflow called Monoqlo automatically detects colony presence and identifies clonality from cellular imaging, enabling single-cell selection protocols to be scalable while minimizing technical variability.

The urgency of the developing COVID-19 epidemic has led to a large number of novel diagnostic approaches, many of which use machine learning. DeGrave and colleagues use explainable AI techniques to analyse a selection of these approaches and find that the methods frequently learn to identify features unrelated to the actual disease.

Gaining access to medical data to train AI applications can present problems due to patient privacy or proprietary interests. A way forward can be privacy-preserving federated learning schemes. Kaissis, Ziller and colleagues demonstrate here their open source framework for privacy-preserving medical image analysis in a remote inference scenario.

In the last few years, computational protein structure prediction has greatly advanced by combining deep learning including convolutional residual networks (ResNet) with co-evolution data. A new study finds that using deeper and wider ResNets improves predictions in the absence of co-evolution information, suggesting that the ResNets do not not simply de-noise co-evolution signals, but instead may learn important protein sequence–structure relationships.

Calcium imaging is a valuable tool for recording in vivo neural activity, but the task of extracting signals of individual neurons is computationally challenging. Bao and colleagues present a U-Net-based method that is both accurate and fast enough to potentially allow real-time processing and closed-loop experiments.

Experimental benchmarks such as ImageNet and Atari games play an important part in advancing artificial intelligence research. An analysis of results and papers linked to 25 popular benchmarks shows that research dynamics beyond conventional co-authorship has developed in this area.