AI algorithms, and especially deep neural networks, are often considered as black boxes, and as a consequence are not easily understandable for humans. The drawbacks of such algorithms can include a) any bias within the training data is potentially transferred to the algorithm and remains undetected, b) users may not trust their predictions, and c) that they may lack robustness in operational environments. Explainable AI is an active field of research that aims to provide insights into the internal decision-making process of machine learning algorithms. Using these insights, algorithms can be developed whose predictions are not only correct but right for the right reasons.
Considering the importance of edge AI, there is a need for commitment to consider the impact of edge AI throughout its lifecycle. To that extent, the developed algorithms must be kept up to date and performant on new data, with the ability to integrate external sources through re-training. In addition to meet the requirements and defined metrics that indicate the training state of the AI system, the re-training must also consider any con- sequence it may have on other components or the system itself. The implication is that rather than having to spend time and resources on re-training from scratch, to incorporate slightly different insights, the re-training should focus on creating more generic models. The aim is to permit improvements in performance through a quick re-training of an edge AI model that has already been trained using previous data sets. Equally, the re-training of one model should not compromise the performance of other components within the system (or other systems within a system of systems). In simple terms, the re-training must enable improved performance through exploitation of new data and in parallel it must not negatively impact its surroundings. Additionally, changes in calibration (e.g. of sensors or actuators) should be permitted without the need to retrain the edge AI.
In distributed learning, a communication overhead is introduced in order for the edge platforms and the system aggregator to transfer data during training and inference. When compared to data processing in large central data centres, data produced on resource-constrained end devices in a decentralized and distributed setting is particularly vulnerable to security threats and the necessary level of protection against such risks should be considered carefully for specific applications. Further research is required to increase the security, privacy, and robustness of edge AI by reducing the overhead, or by adopting novel approaches such as clustered federated learning or federated distillations.
Training artificial neural networks at the Edge remains a challenge. Work has been done to optimize inference at the Edge by optimizing algorithms and accelerators for low precision, low memory footprint and feed-for- ward computations. However, an additional re-training phase of an artificial neural network can undo part of those optimizations as higher precision is needed to enable the iterative approach typically used and more storage is needed to keep track of the intermediate data required. Also, the frequent weight updates during training can pose additional challenges regarding energy efficiency as well as reliability. As such, neuromorphic-based architectures hold potential, as they allow on-line learning to be built in by modelling plasticity. Plenty of challenges remain to achieve this goal as it is difficult to make a single synapse and neuron device that allows the capture of a very wide range of time constants.
Integrating AI into the smallest devices: Recently a number of tools have been developed with the goal of implementing AI models which could fit the memory available in edge platforms. As an example, tinyML is about processing sensor data at extremely low power and, in many cases, at the outer- most edge of the network. Therefore, tinyML applications could be deployed on the microcontroller in a sensor node to reduce the amount of data that the node forwards to the rest of the system. These integrated “tiny” machine learning applications require “full-stack” solutions (hardware, system, software, and applications) plus the machine learning architectures, techniques, and tools performing on-device analytics. Furthermore, a variety of sensing modalities (vision, audio, motion, environmental, human health monitoring, etc.) are used with extreme energy efficiency (typically in the single milliwatt, or lower, power range) to enable machine intelligence at the boundary of the physical and digital worlds. With the increase in dedicated hardware for machine learning, an important direction for future work is the development of compilers, such as Glow, and other tools that optimize neural network graphs for heterogeneous hardware or train and handle specialized technologies and algorithms.
Data is the fundamental piece behind ML/AI. However, one of the major problems when developing AI solutions can be the lack of sufficient data to achieve the required performance in a specific application. In recent years several techniques have been considered to deal with this problem in the context of cloud-based solutions; for example, by using semi-supervised learning (to take advantage of the large amounts of unlabelled data generated by edge devices), by using data augmentation (via Generative Adversarial Networks (GANs) or transformations), or by transfer learning. These have become cutting-edge methods deployed to improve the overall performance in AI models. However, the adoption of these techniques in edge computing still needs to be thoroughly investigated. Moreover, edge systems need to interact with various types of IoT sensors, which produce a diversity of data such as image, text, sound, and motion. Edge analytics should be able to deal with those heterogeneous environments and adapt to be multimodal allowing learning from features collected over multiple modalities.
Neuromorphic engineering is a ground-breaking approach to the design of computing technology that draws inspiration from powerful and efficient biological neural processing systems. Neuromorphic devices are able to carry out sensing, processing, and control strategies with ultra-low power performance. Today, the neuromorphic community in Europe is leading the State-of-the-Art in this domain. The community includes an increasing number of labs that work on the theory, modelling, and implementation of neuromorphic computing systems using conventional VLSI technologies, emerging memristive devices, photonics, spin-based, and other nano-technological solutions. Extensive work is needed in terms of neuromorphic algorithms, emerging technologies, hardware design and neuromorphic applications to enable the uptake of this technology, and to match the needs of real-world applications that solve real-world tasks in industry, health-care, assistive systems, and consumer devices. It is important to note that “neuromorphic” is most commonly defined as the group of brain-inspired hardware and algorithms.
Parallel to the advancement in neuromorphic computing, the underlying computation of such technology gets increasingly complex and requires more and more parameters. This triggers further development of efficient neuromorphic hardware designs, e.g. the development of neuromorphic hardware that can tackle the well- known memory wall issues and limited power budget in order to make such technology applicable on edge de- vices. The emerging memory technologies provide additional benefits for neuromorphic solutions, especially memory technology that can allow us to perform computation directly in the memory cells themselves instead of having to load and store the parameters, inputs, and outputs into computation cores.
Such technology, coupled with the properties of neuromorphic computing, delivers many benefits. Firstly, DL and spiking neural networks (SNN) parameters are often fixed and/or modified very seldom. This matches the capability of emerging non-volatile memories where write accesses are typically one or two orders slower than read accesses as the number of memory writes required is lower. Secondly, most computations are matrix addition and multiplication. This operation can be mapped efficiently in memory arrays. Thirdly, inference of such neuromorphic networks can be optimized for low-bit precision and coarse quantization without sacrificing the quality of the network outputs. Some tasks, such as classification, are proven to be good enough even when networks are optimized to binary and/or ternary representation. This provides an excellent opportunity as the underlying operation can be simply replaced by AND/XOR logic. Fourthly, neural networks are robust to error. Thus, process variations on the emerging memory technologies do not limit their capability to compute and/ or and load/store in the networks. These benefits can be achieved by in-memory compute technology using emerging memory technologies.
In most of today’s industrial applications of deep learning, models and related learning algorithms are tailor-made for very specific tasks. This procedure can lead to accurate solutions of complex and multidimensional problems but it also has visible weaknesses. Normally, these models require an enormous amount of data to be able to learn how to correctly solve problems. Labelled data can be costly as it may require the intervention of experts or not be available in real-time applications due to the lack of generation events. A question can therefore arise: in addition to having the correct formulation and the descriptive data for the problem, is it possible not only to try to solve it but also to learn how to solve it in the best way? Therefore: “is it possible to learn how to learn?” Precisely on this question, the branch of machine learning, called meta-learning (Meta-L), is based. In Meta-L the optimization is performed on multiple learning examples that consider different learning objectives in a series of training steps. In base learning, an inner learning algorithm, given a dataset and a target, solves a specific task such as image recognition. During meta learning, an outer algorithm updates the internal algorithm so that the model learned during base learning also optimizes an outer objective, which tries, for example, to increase the inner algorithm’s robustness or its generalization performance.
Intelligent extraction of information, by addressing the problem from a general point of view can also lead to the ability of the inner algorithm to handle new situations quickly and with little data available with a robust approach. Looking at the advantages of Meta-Learning and the possibility of using it together with Edge computing to increase its benefits, provides a good outline of how this branch of ML can soon find concrete uses in the most varied application scenarios.
Data-based and knowledge-based modelling can be combined into hybrid modelling approaches. Some solutions can take advantage of a-priori knowledge in the form of physical equations describing known causal relationships in the behaviour of the systems or by using well known simulation techniques. Whereas dependencies not known a priori can be represented by many kinds of machine learning methods using big data based on observing the behaviour of the systems. The former type of situation can be seen as white box modelling as the internal states possess a physical meaning, while the latter is referred to as black box modelling, using just the input-output-behaviour, but not maintaining information on the internal physical states of the system. However, in many cases, a model is not purely physics-based nor purely data-driven, giving rise to grey box modelling methods that can be formulated. The assignment of models to the scale varies within the literature: For instance, a transfer function can be derived from physical considerations (white), identified from measurement data with a well-educated guess of the model order (grey) or without (black).
Approaches for combining machine learning and simulation, by simulation-assisted machine learning or by ma- chine-learning-assisted simulation and combinations are described by von Rueden et al. in “Combining Machine Learning and Simulation to a Hybrid Modelling approach: Current and Future Directions” and in “Informed machine learning – towards a taxonomy of explicit integration of knowledge into machine learning.” advantage of hybrid modelling is avoiding the necessity of learning a-priori the behaviour of systems from huge amounts of data, if they can be described by simulation techniques. Also, in the case of missing data, hybrid modelling is a possible approach.
A practical example of combining physical white-box modelling and machine learning to improve a model for the highly non-linear dynamic behaviour of a ship, described by a set of analytical equations has been recently investigated by Mei et al.. Another example is hybrid modelling in process industries.
Reducing energy consumption is a general goal, not only, but especially for smart systems providers to address the challenges of global warming and enable a higher degree of miniaturization of intelligent devices. For a long time power reduction has been a challenge in micro and nano electronics and also a target for all AI applications, regardless of whether data is processed in the cloud or at the edge. But at the edge, this target is especially important as applications usually have only limited power resources available. They often have to be battery powered or even use energy harvesting.
Special energy-efficient neural network architectures have been investigated. Not only is the hardware crucial for low-power AI applications, but also the implemented methods and models have great influence on the energy consumption. This has been examined for the example of computer vision.
Moving away from traditional von Neumann processing solutions and using dedicated hardware allows for additional power reduction. Even more can be achieved with neuromorphic architectures.
The “ultimate benchmark” in power consumption for artificial intelligence would be the “natural intelligence” in form of the human brain, which has 86 bn. neurons and approximately 1014–1015 synapses with an energy consumption of less than 20W, based on glucose available to the brain, or only 0.2W, when counting the ATP usage instead of glucose. Current GPU based solutions with that complexity are far from this energy efficiency. There is obviously plenty of headroom for further development.
 Samek, Wojciech, et al., eds. Explainable AI: interpreting, explaining and visualizing deep learning. Vol. 11700. Springer Nature, 2019.
 D. Silver, A. Huang, C. Maddison, A. Guez, L. Sifre, G. Van Der Driessche, J. Schrittwieser, I. Antonoglou and V. Panneershelvam, “Mastering The Game Of Go With Deep Neural Networks And Tree Search,” Nature, 2016.
 A. Krizhevsky, I. Sutskever, G. Hinton, Ilya and E. Geoffrey, “Imagenet classification with deep
convolutional neural networks,” Communications of the ACM, vol. 60, no. 6, pp. 84–90, 2017.
 G. Marcus, “Deep learning: A critical appraisal,” arXiv preprint arXiv:1801.00631, 2018.
 F. Došilović, M. Brčić and N. Hlupić., “Explainable artificial intelligence: A survey,” in 2018 41st International convention on information and communication technology, electronics and microelectronics (MIPRO), IEEE, 2018, pp. 0210–0215.
 M. Masud, C. Woolam, J. Gao, L. Khan, J. Han, K. Hamlen and N. Oza, “Facing the reality of data stream classification: coping with scarcity of labeled data,” Knowledge and information systems, vol. 33, no. 1, pp. 213–244, 2012.
 J. Vanschoren, “Meta-learning: A survey,” arXiv preprint arXiv:1810.03548, 2018.
 T. Hospedales, A. Antoniou, P. Micaelli and A. Storkey, “Meta-learning in neural networks: A survey,” arXiv preprint arXiv:2004.05439, 2020.
 A. Madela, A. Picon, C. Saratxaga, O. Belar, V. Cabezón, R. Cicchi, R. Bilbao and B. Glover, “Few shot learning in histopathological images: reducing the need of labeled data on biological datasets,” in 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), IEEE, 2019, pp. 1860–1864.
 Y. Wang, Q. Yao, J. T. Kwok and L. M. Ni, “Generalizing from a few examples: A survey on few-shot learning,” ACM Computing Surveys (CSUR), vol. 53, no. 3, pp. 1–34, 2020.
 D. Chen, Y. Liu, B. Kim, J. Xie, C. S. Hong and Z. Han, “Edge Computing Resources Reservation in Vehicular Networks: A Meta-Learning Approach,” IEEE Transactions on Vehicular Technology, vol. 69, no. 5, pp. 5634–5646, 2020.
 Schoukens, J., & Ljung, L. (2019). Nonlinear System Identification: A User-Oriented Roadmap.
ArXiv:1902.00683 [Cs]. http://arxiv.org/abs/1902.00683
 Laura von Rueden, Sebastian Mayer, Rafet Sifa, Christian Bauckhage and Jochen Garcke, “Combining Machine Learning and Simulation to a Hybrid Modelling approach: Current and Future Directions” IDA 2020, LNCS 12080, pp. 548–560, 2020.
 Von Rueden, Laura, et al. “Informed machine learning–towards a taxonomy of explicit integration of knowledge into machine learning.” Learning 18 (2019): 19-20.
 Urko Leturiondo, Oscar Salgado, Lorenzo Ciani, Diego Galar, Marcantonio Catelani, “Architecture for hybrid modelling and its application to diagnosis and prognosis with missing data,” http://dx.doi. org/10.1016/j.measurement.2017.02.003
 Mei, B., Sun, L., & Shi, G. (2019). White-Black-Box Hybrid Model Identification Based on RM-RF for Ship
Manoeuvring. IEEE Access, 7, 57691–57705. https://doi.org/10.1109/ACCESS.2019.2914120
 Glassey, Jarka, and Moritz Von Stosch, eds. Hybrid modelling in process industries. CRC Press, 2018.
 Hsi-Shou Wu, Energy-Efficient Neural Network Architectures, dissertation, University of Michigan,
 Abhinav Goel, Caleb Tung, Yung-Hsiang Lu, and George K. Thiruvathukal, A Survey of Methods for Low-Power Deep Learning and Computer Vision, Purdue University and Loyola University, Chicago
 Yu Wang, Lixue Xia, Tianqi Tang, Boxun Li, Song Yao, Ming Cheng, Huazhong Yang, Low Power Convolutional Neural Networks on a Chip, Tsinghua University, Beijing, China, IEEE 2016
 Arianna Rubino, Melika Payvand, and Giacomo Indiveri, Ultra-Low Power Silicon Neuron Circuit for Extreme-Edge Neuromorphic Intelligence, Institute of Neuroinformatics, University of Zurich and ETH Zurich, IEEE 2019
 William B Levy, Victoria G. Calvert, Computation in the human cerebral cortex uses less than 0.2 watts yet this great expense is optimal when considering communication costs, doi: https://doi. org/10.1101/2020.04.23.057927
Parts of this roadmap is reproduced from the European Platform on Smart Systems Integration (EPoSS) 2021 Whitepaper