Neuromorphic engineering in wetware: the state of the art and its perspectives

摘要

The United Nations General Assembly (2015) has compiled an Agenda, containing 17 goals to be pursued worldwide to promote a sustainable future by 2030. Accomplishing these goals requires designing and implementing more effective strategies to manage Complex Systems, including human beings and their societies, the world economy, urban areas, natural ecosystems, and the climate (Gentili, 2021a). A promising strategy, which is literally blooming, relies on the development of Artificial Intelligence (AI) and Robotics. AI helps humans collect, store, and process the Big Data required to monitor the constant evolution of Complex Systems (Corea, 2019). AI also assists us in making up our minds for controlling the behavior of Complex Systems. Hard and soft robotics allow humans to access environments otherwise precluded. For instance, they help us (1) investigate colonize the geochemical characteristics of other planets and examine the abysses of our oceans to discover new mines of precious materials and energy resources, (2) access the interior organs of our bodies for less invasive surgery, (3) and work in dirty or dangerous places. Two are the principal and traditional approaches exploited to develop AI (Lehman et al., 2014) (Mitchell, 2019). The first approach entails writing "intelligent" software that runs on electronic computers based on von Neumann's architecture, whose principal drawback is having processing and memory units physically separated. Some software mimics rigorous logical thinking, while others imitate the structural and functional features of neural networks to learn how to perform tasks from data. The second approach for developing AI entails implementing artificial neural networks in hardware for neuro-prosthesis or designing brain-like computing machines, with processors and memory confined in the same space (the so-called mem-computing) (Sebastian et al., 2020). Artificial neural networks are rigid if they are made of silicon-based circuits or inorganic memristors; they are flexible if based on organic semiconductor films (Christensen et al., 2022) (Lee and Lee, 2019) (Wang et al., 2020) (Zhu et al., 2020). They can be designed with three distinct architectures: (A1) feedforward (having trainable unidirectional connections), (A2) recurrent (with trainable feedback actions), or (A3) reservoir (consisting of an untrained nonlinear dynamic system coupled to trainable input and output layers) network (Nakajima, 2020) (Tanaka et al., 2019) (Cucchi et al, 2022) (see Figure 1A).In the last decade or so, a novel promising strategy to develop AI has been put forward: it consists of mimicking human intelligence and the forms of intelligence exhibited by all the other living beings through molecular, supramolecular, and systems chemistry in wetware, i.e., liquid solutions (Gentili and Stano, 2023a) (Gentili andStano, 2023b) (Kazuya et al., 2023) (Murata et al., 2022), which is the peculiar phase supporting life. As we believe that this still not-well-explored field represents a huge opportunity to understand and exploit computation in the molecular realmthus closely mimicking the natural (biological) cognitive abilitieshere we would like to highlight the current methodologies. In particular, we focus on artificial neural networks in wetware and, hence, on the strategies to develop neuromorphic engineering in the fluid phase. The selection of topics presented in this short article is not meant to represent the whole diversity of this research areait rather mirrors our specific interests. The variegate methodologies proposed so far can be grouped into three distinct approaches (see Figure 1B) presented succinctly in the next three paragraphs. Some fFuture perspectives are shortly presented in the last paragraph. Any liquid solution containing two or more reactive solutes may display some of the brain's dynamic features, especially if considered as a useful model or even a simplified version of it. be conceived as a primit