Anagrafe della ricerca

In the field of unconventional computing, a wide variety of methods are used to process unknown and complex data, which can be roughly divided into statistical methods and machine learning methods. While the use of statistical methods already has a long tradition, machine learning (ML) has gained more and more importance in recent years and has been able to demonstrate its potential impressively. The reason for this impetuous development is that, in contrast to statistical methods, such as nonlinear vector- autoregression (NVAR), ML methods can cope with massive unstructured data, requiring little information about their context. This is of particular interest for pattern recognition, i.e. classification tasks, but can also be used very effectively for time series analysis. An interesting way to synergistically combine these two fields is offered by a methodology known as Reservoir Computing (RC). In contrast to other ML methods, like deep-learning networks (DNNs), where all layers need to undergo a learning process, RC offers the advantage that only the output layer of the network must be trained. For this advantage to be gained, a reservoir is needed. A reservoir is any dynamical system capable of generating a non-linear transformation of the input data, encoding it in some multi-dimensional space. If the reservoir is sufficiently large (has a sufficiently-high dimension), and the reservoir parameters are suitably selected, the resulting non-linear representation of the input data allows to solve challenging computational problems through the output (read-out) layer. Reservoirs can be efficiently realized in hardware. In particular, delay RC systems accomplish this purpose with a single physical node and realize the non-linear transformation via a temporal delay of the input data, i.e. by means of high-dimensional transient dynamics. The physical realization of this strategy typically leverages optical or electro-optical systems, while the special properties of memristor devices may enable a purely-electronic bio-inspired hardware realization with unprecedented efficiency. The possibility of a synergetic connection between RC and the field of statistical methods was recently investigated by Gauthier et al., which introduced a new paradigm named Next Generation Reservoir Computing (NGRC). The idea descends from a similarity between RC and static regression methods, which also use non-linear transformations of data. These methods, however, do not use a reservoir for implementing the nonlinear mapping. They directly operate on the input data, particularly on what is referred to as the feature vector. NGRC uses correlations, involving the feature vector, together with the Tikhonov regularization method to generate a suitable non-linear transformation of the input data in the reservoir. As a key advantage, due to the absence of a true reservoir, a significantly smaller number of parameters needs to be adjusted, and a shorter set of training data is required. This also makes the output layer smaller than what is required in the classical RC paradigm, which leads to shorter computation times, resulting, consequently, into a more effective hardware design. However, the NGRC method requires the optimization of the type and number of the non-linearities of the feature vector, which can be a more difficult task than the process of fine tuning the reservoir parameters, especially for the case where little is known about the input data set. In this context, the use of nonlinear circuit elements, and memristors in particular, can address this issue, since on one hand they can produce an appropriate nonlinear tunable representation of the input data, when integrated in a suitable circuit topology, and on the other hand they allow an efficient and compact realization of the signal processing concept in hardware. This is the motivation and goal of the RECOMMEND proposal. However, to realize the NGRC concept across neuromorphic hardware, a bridge must be built from the device level to the system level. This requires the development of a physical model of the memristor, which may predict accurately its complex behavior, may enable reliable investigations, and may be integrated into circuit design tools, the synthesis of circuit and system-theoretic methods to tailor the nonlinear dynamics of an analogue memristive circuit for achieving a suitable nonlinear representation of the input data, and the physical realization of a bio-inspired hardware platform which may demonstrate higher energy efficiency over the state-of-the-art. The successful execution of these multidisciplinary activities crucially requires the synergetic work of PIs, excelling in different yet complementary scientific fields, as those lining up in our consortium. Given that the open questions, the proposed project intends to find an answer to, envisage challenges of common interest to this DFG Priority Program (SPP), our research developments will stimulate discussions and collaborations with scientists engaged in other projects.