Self-organizing memristive networks have been demonstrated as physical substrates for in materia implementation of unconventional computing paradigms such as reservoir computing [1,2]. However, a detailed understanding of the relationship between electrical properties, emergent dynamics and computational properties in these multiterminal systems remain challenging. Here, we report on emergent functionalities in multiterminal nanowire networks through a combined experimental and modeling approach. Beyond extending the traditional two-terminal memristive concept to multiterminal architectures, we discuss experimental characterization techniques used to analyze these emergent behaviors, including conductance matrices, voltage maps, and conductivity maps obtained via electrical resistance tomography. By linking self-organizing memristive systems to dynamical systems theory, we introduce a new framework that enables the description of self-organizing nanowire networks as stochastic dynamical systems [3]. In this framework, emergent dynamics are modeled as an Ornstein-Uhlenbeck process, where stimulus-dependent deterministic trajectories coexist with stochastic effects such as noise and discrete jumps. Within this context, we demonstrate that the emergent dynamics of nanowire-based self-organizing systems can be harnessed for computation. We evaluate their computational capabilities through benchmark tasks, including nonlinear autoregressive moving average (NARMA) and nonlinear transformation (NLT) tasks, and discuss the implementation of pattern recognition, speech recognition, and time-series prediction.

Neuromorphic computing implemented with spiking neural networks (SNNs) using volatile threshold switching devices is a promising computing paradigm that may overcome future limitations of the von Neumann architecture. For the computation by dynamics and time new technologies for devices with high dynamics are required, for which volatile electrochemical memory (v-ECM), also known as diffusive memristor, is very promising [1,2]. v-ECM devices switch from the high-resistance to the low-resistance state when the voltage rises above the threshold voltage and automatically restores the high-resistance state after a relaxation time when the voltage falls below the holding voltage [3]. The advantage of v-ECMs compared to insulator-to-metal-based threshold switches is their extremely low leakage current and the low threshold voltages, while a drawback is the higher switching variability. We have investigated in detail the set kinetics and relaxation dynamics of v-ECM threshold switches based on analysis of the transient current in pulsed voltage measurements for Ag/HfO2/Pt crossbar devices made of 3 nm amorphous HfO2 by atomic layer deposition. The results show a clear correlation between the relaxation time and the operating parameters with reference to the set kinetics [4,5]. This is crucial for the application of v-ECM cells in neuromorphic circuits. There are several approaches to understand the physical origin of the relaxation behaviour, ranging from surface energy minimization [6] to the effect of electromotive force [7]. The presentation will summarize different approaches of device fabrication, physical understanding, and applications of v-ECM devices, also known as diffusive memristors. In addition, actual results based on integrate-and-fire circuits will be discussed.

Nowadays, the increasing popularity of extremely energy demanding AI models raises serious concerns in terms of environmental sustainability [1]. At the same time, the need to process vast amounts of often unstructured data highlights the urgency of exploring new and efficient computational paradigms. The mammalian brain performs complex tasks with remarkable efficiency compared to digital computers, exhibiting adaptability and learning capabilities. These features arise from brain peculiar architecture, comprising hundred billions neurons interconnected through an enormous number of synapses. Neurons integrate signals in a nonlinear fashion, while synapses modulate their transmission efficacy through plasticity, thus enabling learning. Neuromorphic Computing (NC) draws inspiration from the brain to develop viable alternatives to conventional digital computing.

Recently, gold cluster–assembled films (ns-Au) have attracted significant attention due to their topological complexity, memory effects, and non-linear electrical response, making them promising physical substrates for NC applications. Their intriguing neuromorphic properties arise from Resistive Switching (RS) phenomena displayed by the nano-junctions distributed thorough the network [2][3]. Notably, the implementation of a reconfigurable nonlinear threshold logic gate based on ns-Au has already been demonstrated [4].

We focused our study on the characterization of the switching behaviour exhibited by ns-Au with the aim to investigate the emergent electrical response of these systems. We employed a novel microthermography based imaging technique to track the regions responsible for the RS activity during electrical stimulation, referred as “switching sites”. This analysis allowed us to reliably study the correlation between different switching sites, gaining insight in the functional connectivity of the network [5]. Moreover, we highlighted that the geometry of the deposit has also a strong impact over the RS activity, particularly influencing the number of switching sites and the statistics of the RS, providing meaningful insight for future design and engineering of ns-Au based devices.

Furthermore, our experiments allowed us to optimize electrical stimulation protocols for reliably tuning the electrical resistance map of multi-electrode device and the connectivity of the network by means of selectively triggering the formation and destruction of conductive paths. This way, we demonstrate that we have achieved a fine reversible control over the conduction state of the devices, which has never been previously reported for self-assembled materials. This characterization paved the way for the implementation of optimized stimulation protocols for implementing specific computing functionalities in reconfigurable ns-Au based multiterminal devices.

Nonlinear oscillators are increasingly recognized as core building blocks for physics-based computation, where information is processed through dynamical behaviors such as spiking, synchronization, and phase relationships. A wide variety of material platforms—including phase-change compounds, mixed ionic–electronic conductors, van der Waals materials, and nanoscale fluidic channels—have been explored for such dynamical devices. Yet across these diverse systems, a unifying challenge emerges: computational functionality is determined not only by materials properties, but by the underlying dynamical structure of the device. In this work, I present a cross-platform dynamical description of oscillator devices—spanning silicon thyristors, organic electrochemical transistors, and rectifying nanopores—demonstrating how nonlinear dynamics and bifurcation theory provide a general framework for designing, tuning, and optimizing oscillatory behavior for physical computation.

Despite their disparate physical mechanisms, these systems share a common architecture: a nonlinear element exhibiting negative differential resistance or negative transconductance, coupled to slow dynamical variables arising from capacitive charging, ionic motion, or configurational relaxation. Together these ingredients form slow–fast dynamical systems capable of self-sustained oscillations through a Hopf bifurcation, generating limit cycles whose amplitude, waveform, and frequency can be engineered through device parameters and external circuit elements.

I first discuss a compact two-terminal silicon thyristor oscillator with an ultrasmooth S-type NDR characteristic, enabling robust and hysteresis-free oscillations. Combined experimental and analytical work identifies a tunable Hopf bifurcation governed by input current and capacitance, producing a continuous transition from sinusoidal oscillations near onset to relaxation waveforms at higher capacitances. When operated close to the bifurcation threshold, the system exhibits stochastic resonance, illustrating how intrinsic nonlinearities can amplify weak signals and support temporal processing tasks.

I then extend the same dynamical framework to a single-transistor organic electrochemical oscillator, where the coupling between ion transport and electronic conduction creates a peaked transfer curve and an effective negative transconductance region. Despite the distinct physics of mixed ionic–electronic materials, the device maps onto the same two-variable oscillator class: a fast destabilizing electronic response combined with a slow ionic recovery. The Hopf criterion accurately predicts the emergence of autonomous oscillations without external amplifiers, showing how minimal circuit motifs can generate functional dynamical behavior in soft and biointegrated materials.

Finally, I show that fluidic nanopores—rectifying channels with hysteresis and deactivation dynamics—exhibit an analogous bifurcation structure. When described with fast activation and slow deactivation variables, the nanopore displays a negative resistance sector and undergoes a Hopf bifurcation, thereby functioning as a minimal liquid-phase oscillator. This demonstrates that the same mathematical formulation governing solid-state and organic oscillators extends naturally to fluidic systems, enabling dynamical computation in aqueous environments.

Across these platforms, a unified dynamical-systems perspective yields general design rules for oscillatory devices: (i) the shape and smoothness of the stationary I–V or transfer curve dictate the onset of instability; (ii) the ratio of fast to slow timescales controls the transition between harmonic and relaxation dynamics; (iii) external resistive–capacitive elements act as tunable parameters for frequency, amplitude, phase, and noise sensitivity; and (iv) operating near bifurcation points provides regimes of enhanced responsiveness, valuable for sensing, synchronization, and oscillator-based computational architectures.

Overall, this work argues that nonlinear dynamics offers a materials-agnostic design principle for next-generation dynamical devices. By treating oscillators as physical–computational entities governed by universal bifurcation structures, we open new pathways for engineering scalable, tunable, and energy-efficient computation across solid-state, organic, and fluidic technologies.^1,2

Artificial neural network (ANN)-based computing [e.g., deep learning with a multi-layer neural network (NN)] can provide excellent learning, classification, and inference characteristics that are close to, and in some cases beyond, those found in natural intelligence (i.e., the human brain), whereas the enormous amounts of power required by ANN (as in a typical multi-layer NN) are far higher than that required by human beings. To overcome the low energy efficiency of ANN computing, in-materio computing, which harnesses the inherent properties of materials to perform computation, has recently attracted attention. Among various types of computing, physical reservoir computing (PRC) is particularly attractive because it can significantly reduce the computational resources required to process time-series data by leveraging the nonlinear responses of a ‘reservoir’ (a material or device acting as a dynamical system) to input signals. Realizing nonlinear and diverse dynamics with nanomaterials and/or nanospace is thus a major challenge for the development of low-power, highly integrated ANN-based computing devices. Recently, we have developed high-performance PRC devices based on iontronic phenomena. One example is an ion-gating reservoir (IGR), which utilizes ion-electron coupled dynamics in the vicinity of a solid electric double layer [1,2]. Another example is a magnonic PRC device that utilizes the chaotic dynamics of interfered spin waves in a ferrimagnetic Y3Fe5O12 (YIG) single crystal [3]. In the presentation, the device performance and the operating mechanisms of the two PRC systems will be discussed.

Tiny processing entities, such as microcontrollers, are at the basis of the Internet of Things paradigm. The integration of Machine Learning models on small computing units require the development of architectures and hardware solutions that do not need the processing support from the cloud and the training processes typical of Artificial Intelligence, which are power consuming and involves data security and privacy risks [1]. To face this challenge, a novel threshold logic gate design, called Receptron, has been recently proposed as an alternative to multilayer architectures based on perceptrons typical of artificial neural networks [2]. The receptron is based on the use of nonlinear weights thus widening the spectrum of Boolean computable functions while simplifying training thanks to a random search protocol [3,4]. The hardware implementation of the Receptron model has been demonstrated to rurn on standard microcontrollers and CMOS components.

The reconfigurability and functional completeness of the receptron allows faster, more efficient and possibly edge, data processing compared to traditional architectures used for artificial neural networks. Here we report the fabrication of a receptron-based electronic board for edge data processing and classification working on analog inputs and capable to learn with an extremely reduced training.

Both in Neuroscience and in AI, the input-output transfer function of the units composing a large network is a pivotal element. Known as "activation function" (in Machine Learning) or as the frequency-current curve (in Neurobiology), it is a known mechanism of non-linearity as well as a biophysical primitive for neural computation.

We also know that its knowledge is instrumental (in Computational Neuroscience) to analyse and predict the collective behavior of real neuronal circuits by mean-field theories. Finally, in experimental Neuroscience, the estimate of the frequency-current curve of nerve cells has been used for over three decades to classify and make sense of neuronal diversitys. 

However, such a static description is inadequate to interpret how microcircuits and large networks of the brain process time-varying stimuli. Early theoretical studies and late experimental work from our group and others revealed how probing single-cell dynamical response properties is necessary to interpret of ultra-fast ensemble responses and other collective network phenomena. 

In this talk, I will review the results on dynamical response properties of neurons and neuronal ensembles and put them into context of the findings of unexpected differences between rodent and human cortical neurons.

Whethere these generalised biological features should be also considered for (Neuromorphic) implementation of AI systems, based on spiking neural models, is a question I will leave open to the audience.

References

Linaro D, Ocker GK, Doiron B, Giugliano M (2019) Correlation transfer by layer 5 cortical neurons under recreated synaptic inputs in vitro, J Neuroscience, 39 (39) 7648-7663, https://doi.org/10.1523/JNEUROSCI.3169-18.2019   

Linaro D, Biró I, Giugliano M (2018) Dynamical response properties of neocortical neurons to conductance-driven time-varying inputs, European Journal of Neuroscience 47(1):17–32, https://doi.org/10.1111/ejn.13761  

Testa-Silva, G., Verhoog, M.B., de Kock, C.P.J., Baayen, J.C., Meredith, R.M., Giugliano, M.*, Mansvelder, H.D.* (2014) High bandwidth synaptic communication and frequency tracking in human neocortex. PLoS Biology 12(11):e1002007. http://dx.doi.org/10.1371/journal.pbio.1002007  

Köndgen, H., Geisler, C., Fusi, S., X.-J. Wang, Lüscher, H.-R., Giugliano, M. (2008) The dynamical response properties of neocortical neurons to temporally modulated noisy inputs in vitro, Cerebral Cortex 18(9), 665-670. http://dx.doi.org/10.1093/cercor/bhm235http://dx.doi.org/10.1093/cercor/bhm235 

Performances of computers based on von Neumann architecture [1], with separated  data processing and memory units, are hindered by power dissipation, execution time and sustainability issues [2]. This makes it challenging to process in real time a substantial amount of data for complex tasks, that are carried out with great efficiency by the human brain. This is a network composed by basic computing units, neurons, which are densely connected in a self-assembled and redundant way through synapses. Their conductive strength depends on the signals previously received, which are spike trains whose frequency encodes information; computation and memory thus take place together.

            Computational models and engineering solutions emulating the biological nervous system are called neuromorphic; thin films made of metallic clusters, fabricated by Supersonic Cluster Beam Deposition [4], can be an experimental strategy used to implement in hardware such paradigm. These systems show non-linear electrical properties and resistive switching [5,6]: their resistance has a spiking activity and explores distinct levels when constant voltage is applied.

            Here, we report how gold nanostructured networks change their conductive state depending on the features of voltage trains as input signals, in a controllable and reproducible way.

            We proved that the device resistance is affected by varying the temporal features and voltage levels of pulsed signals, switching reversibly between levels, several orders of magnitude apart; this switching mechanism is controlled by varying time and polarity features of the signal. Furthermore, devices preserve their conductive state in time and can cycle dozens of times between different resistance levels.

This behaviour, observed at first in two terminal devices, can be found also in multielectrode ones. Not only reversibility and input dependence are preserved, but the more complex terminal geometry allows to study the non-local electrical response of such devices: by applying the proper signal between two terminals it modifies not only the resistance of the path connecting them, but also those of the paths that do not cross it, in a non-trivial and reproducible way.

This type of device emulates the heterosynaptic plasticy featured in the brain, that makes it possible to change its state instantaneously and globally depending on input signal features. We propose the use of multi-electrodes devices based on cluster-assembled thin films both as a neuromorphic multiplexer and reconfigurable threshold logic gate.

To further engineer these devices, we plan to fabricate cluster assembled films with tailored metallic materials, featuring different thermic and electrical properties, as also to study their real time response to analog inputs.

Self-assembled conducting-polymer architectures offer a powerful route to create adaptive bioelectronic systems without relying on rigid templates or conventional patterning. Among these materials, PEDOT dendritic fibers formed through AC electropolymerization stand out for their ability to grow into hierarchical, branched networks whose morphology and conductivity emerge directly from local electric fields and ionic dynamics. This template-less growth not only simplifies fabrication but also yields architectures with intrinsic electronic responsiveness and mixed ionic/electronic transport.

In this talk, we present recent advances demonstrating how these dendritic networks can serve as adaptive interfaces for biological systems. Their complex microstructure provides both a physical scaffold and an active electrochemical environment that supports cell adhesion and organization. Because the fibers modulate—and are modulated by—the surrounding electrolyte, they enable a form of reciprocal bioelectronic feedback, where cellular activity influences the polymer’s state and vice versa. This creates opportunities for spatially localized stimulation, sensing, and computation within a single soft-matter platform.

To guide the rational design of such systems, we introduce a computational framework that models field-driven dendritic morphogenesis and predicts the resulting electrotactic landscape for cells. The simulations capture key experimental trends and identify regions that promote directed cell motion or alignment, providing a foundation for programming template-less interfaces capable of influencing cellular behavior.

Overall, these results outline a pathway toward electrochemically-assembled, biologically integrated, and computationally capable materials—advancing the vision of soft, adaptive bioelectronics rooted in emergent physical principles.

The future edge computing applications, such as the Internet of Things (IoT), demand substantial data processing performed locally. This makes the conventional von Neumann computing architecture unfit for such applications due to the limited energy supply in edge computing. In this circumstance, highly energy-efficient computing paradigms are needed, such as neuromorphic computing and in-memory computing (IMC) [1]. IMC refers to local information processing where data is generated and stored, making it particularly appealing for data-intensive applications. Emerging non-volatile memory technologies are the key enablers for IMC, and among them, memristors are regarded as the most promising memory devices due to their low-energy and fast switching characteristics. Memristor-based stateful logic [2] forms the fundamental IMC primitive, where the output is stored in the same cells that perform the computation, thereby eliminating write operations to a separate non-volatile memory block. In this talk, we present experimental demonstrations of five stateful logic operations—AND, OR, XOR, NOR, and NOT—implemented with inkjet-printed Ag/ZnO/Au memristors [3]. We discuss that the switching characteristics of the printed memristors, namely the asymmetric switching voltages of SET (1.43 V) and RESET (−0.41 V), and a compliance current during logic operation, prevent input switching of memristors in an anti-series circuit configuration. This allows for higher output accuracy rates and crossbar-compatible circuit topologies. The presented results show high output accuracy rates of 94% for OR and 98.5% for AND. With the reliable OR and AND gates, we demonstrate the realization of a compare-and-swap (CAS) sorting circuit for an in-memory unary sorting application. The presented results highlight the potential of using printed memristive stateful logic as a general-purpose in-memory computing solution for emerging applications requiring local data processing, such as wearable electronics for healthcare and low-power sensor nodes for agricultural monitoring.

This work presents a simulation study of the circuit-level translation of degree-one piecewise-linear B-splines into spiking-compatible electronic primitives, supported by a robust selection network.
The core contribution is a methodology to translate each linear segment of a B-spline—used to implement the univariate nonlinearities of Kolmogorov–Arnold Networks (KAN)—into a compact spiking circuit element: a single 2T1R-based supercell synthesizes each linear piece, while a comparator-based thresholding network generates a one-hot selection signal that gates only the active segment. The thresholding network, based on cascaded comparators and simple digital logic, produces a one-hot selection signal that isolates the active spline segment for a given input voltage, preventing contributions from non-relevant segments and enabling localized, spike-timed computation. Circuit prototypes (including comparator stages and representative 2T1R supercells used as linear-segment primitives) are validated with LTspice simulations that confirm correct segment isolation, controllable analog outputs compatible with spike-to-voltage encoding, and repeatable device-level behaviour for the chosen memristive model. System-level Python experiments use a limited assembly of supercells to verify functional composition, spike timing, and correct segment selection under representative input patterns.
Importantly, the present study deliberately scopes the contribution to the supercell-level spline-to-spiking translation and the selection mechanism, while mapping to large-scale crossbar arrays and in-situ learning mechanisms are identified as future work. This focused contribution clarifies a practical hardware pathway for implementing interpretable univariate nonlinearities in neuromorphic systems.

Randomly assembled materials composed of nanoparticles or nanowires exhibit remarkable functional properties that can be leveraged to mimic brain-like data processing. This capability stems from the inherent memristive character and the intricate wiring of the junctions connecting the building blocks within the nanostructured network.

Specifically, nanostructured cluster-assembled metallic films, such as those fabricated by assembling gas-phase produced platinum clusters, demonstrate nonlinear conduction properties. These arise from the extremely high density of grain boundaries and the resulting complex arrangement of nanojunctions which result in a current response to applied voltage ramps that reveals the presence of negative differential resistance.

Crucially, different stable values of electrical resistance can be accessed and reversibly set using unipolar voltage pulses. This process enables a long-term memory effect due to the stability of these distinct resistive levels over time. Such traits have allowed the integration of these nanostructured films with conventional electronic components in programmable analog circuits, including devices like gain amplifiers and relaxation oscillators. Exploiting the unique properties inherent in these platinum cluster-assembled films, a novel class of neuromorphic systems can be realized for unconventional electronic devices. The reprogrammability of the electrical resistance, the complex connectivity of the network that results in nonlocal modification of the resistance map, and long-term memory effects can be exployted to implement electrical components that function as reconfigurable nonlinear threshold logic gates. Furthermore, the peculiar nonlinear electrical response of platinum-based cluster-assembled devices is demonstrated to be well-suitable for hardware solutions that can significantly decrease energy and time consumption of data processing, particularly for classification tasks.