by Robert Haas (IBM Research Europe) and Michael Pfeiffer (Bosch Center for Artificial Intelligence)
The origins of artificial intelligence (AI) can be traced back to the desire to build thinking machines, or electronic brains. A defining moment occurred back in 1958, when Frank Rosenblatt created the first artificial neuron that could learn by iteratively strengthening the weights of the most relevant inputs and decreasing others to achieve a desired output. The IBM 704, a computer the size of a room, was fed a series of punch cards and after 50 trials it learnt to distinguish cards marked on the left from cards marked on the right. This was the demonstration of the single-layer perceptron, or, according to its creator, "the first machine capable of having an original idea". 
Computation in brains and the creation of intelligent systems have been studied in a symbiotic fashion for many decades. This special theme highlights this enduring relationship, with contributions from some of the region’s leading academic and industrial research laboratories. Ongoing collaboration between neuroscientists, cognitive scientists, AI researchers, and experts in disruptive computing hardware technologies and materials has made Europe a hotspot of brain-inspired computing research. The progress has been accelerated by EU-funded activities in the Future and Emerging Technologies (FET) programme, and more recently in the Human Brain Project. Over the last decade, similar large-scale interdisciplinary projects have started in the rest of the world, such as the BRAIN initiative in the US, and national initiatives in China, Canada and Australia. Brain-inspired computing projects from large industrial players such as IBM, Intel, Samsung and Huawei have focused on disruptive hardware technologies and systems. In technology roadmaps, brain-inspired computing is commonly seen as a future key enabler for AI on the edge.
Artificial neural networks (ANNs) have only vague similarities with biological neurons, which sparingly communicate using binary spikes at a low firing rate. However, compared with conventional ANNs, spiking neural networks (SNNs) require further innovation to be able to classify data accurately. Wozniak et al. present the spiking neural unit (SNU), a model that can be introduced into deep learning architectures and trained to a high accuracy over a broad range of applications, including music and weather prediction. Yin et al. show another application of backpropagation through time for tasks such ECG analysis and speech recognition, with a novel type of adaptive spiking recurrent neural network (SRNN) that achieves matching accuracy and an estimated 30 to 150-fold energy improvement over conventional RNNs. Masquelier also exploits backpropagation through time, but accounts for the timing of spikes too, achieving comparable classification accuracy of speech commands compared with conventional deep neural networks.
SpiNNaker, described by Furber, and BrainScaleS, outlined by Schemmel, are the two neuromorphic computing systems that are being developed within the Human Brain Project. With a million cores representing neurons interconnected by a novel fabric optimised to transmit spikes, SpiNNaker allowed the first real-time simulation of a model of the early sensory cortex, estimated to be two to three times faster than when using GPUs or HPC systems. BrainScaleS introduced novel chips of analogue hardware circuits to represent each neuron, which are then configured in spiking neural networks, performing over 1000 times faster than real time, and leverage general-purpose cores closest to the analogue cores to control the plasticity of each neuron's synapses. This enables experimentation with various learning methods: For instance, Baumbach et al. describe use cases that exploit these cores and simulate the environment of a bee looking for food, or perform reinforcement learning for the Pong game. In Göltz et al., the timing of spikes is exploited, with an encoding representing more prominent features with earlier spikes, and training of the synaptic plasticity implemented by error backpropagation of first-time spikes.
How can modern AI benefit from neuroscience and cognitive science research? Alexandre et al. present their interdisciplinary approach towards transferring neuroscientific findings to new models of AI. Two articles demonstrate how both hardware- and data-efficiency can be increased by following brain-inspired self-organisation principles. The SOMA project, presented by Girau et al., applies both structural and synaptic neural plasticity principles to 3D cellular FPGA platforms. The multi-modal Reentrant Self-Organizing Map (ReSOM) model presented by Khacef et al. highlights new opportunities to reduce the need for high volumes of labelled data by exploiting multi-modal associations. Ahmad et al. introduce more plausible approaches towards plasticity to replace the well-known, but biologically unrealistic, backpropagation algorithm used for training deep neural networks. Nagy et al. use results from human memory experiments to inform a new semantic compression technique for images, which captures the gist of visual memories. The highly efficient processing of the visual system is used as inspiration for a novel image and video compression technique by Doutsi et al., exploiting the temporal dynamics of spiking neurons to detect characteristics of visual scenes.
A promising approach to brain-like computation, which could be applied to machine learning and robotics, is computing with very high-dimensional, unreliable, highly efficient, and widely distributed neuronal representations. Rahimi et al. present an implementation of hyperdimensional computing on integrated memristive crossbars, showing how this computational paradigm is particularly well-suited for memristive hardware. Other examples of hardware–software co-design are described by Breiling et al., who present the results of a national competition in Germany, in which entrants developed energy-efficient AI hardware for ECG classification.
Brains naturally combine signals from different modalities, which can help develop better sensor fusion algorithms for artificial autonomous systems. A neuromorphic implementation of a visual-auditory integration model using bio-inspired sensors for accurate and stable localisation is presented by Oess and Neumann. Similarly, Janotte et al. describe a neuromorphic approach to both sensing and local processing of tactile signals. In their approach, event-driven sensors in the electronic skin encode tactile stimuli, and spiking neural networks learn to extract information that can be used for real time, low power control of robots or prostheses.
Computational neuroscience is an important source of information for brain-inspired computing. When trying to model the vast repertoire of learning rules at play at any time in different parts of the brain, manual modeling of plasticity rules beyond Hebbian learning and STDP becomes tedious. Mettler et al. demonstrate how new mechanistic principles and rules for plasticity can be discovered by evolutionary algorithms, using an evolving-to-learn approach. A new mechanistic model of the cerebellar Purkinje neuron, by Nilsson and Jörntell, is matched to in-vivo recordings of spike communications and provides a surprisingly simple explanation as to what such cells actually do.
Europe is currently a leader in advancing neuromorphic computing technologies, but there is still a long road ahead to translate the research and technology development results into novel products. The NEUROTECH project, presented by Payvand et al., has been very successful in building a community for European researchers in the field of neuromorphic computing. Among the highlights are monthly educational events on core themes such as event-driven sensing, and circuits for processing and learning. Projects such as NeuroAgents, described by Indiveri, aim to develop autonomous intelligent agents that show cognitive abilities on robotic platforms, entirely based on brain-inspired algorithms, sensors, and processors.
Although AI-based systems have reached performance levels beyond human capabilities in many specialised domains, such as image classification and game playing, it’s important to note that there is still a long way to go for us to reach the flexibility and efficiency of biological brains in real-world domains. Building increasingly brain-like AI also requires re-thinking fundamental principles of computation: the classical von-Neumann model with its separation of computation and memory is not an efficient way to implement large scale biologically inspired neural networks. Recent progress in neuromorphic computing platforms and in-memory computing provides new opportunities for fast, energy-efficient, massively parallel, and continuously learning large-scale brain-inspired computing systems based on spiking neural networks. Furthermore, envisioning more advanced capabilities, such as making analogies and reasoning on problems is another step that will hinge on an interplay of ANNs and symbolic AI, a recent and very active field of research.
 Melanie Lefkowitz, “Professor´s perceptron paved the way for AI – 60 years too soon”, Cornell Chronicle, https://news.cornell.edu/stories/2019/09/professors-perceptron-paved-way-ai-60-years-too-soon
IBM Research Europe, Zurich Research Laboratory, Switzerland
Bosch Center for Artificial Intelligence, Renningen, Germany