Kevin Portner

Last Name

Portner

First Name

Kevin

ORCID

0000-0002-1175-9122

Organisational unit

03925 - Luisier, Mathieu / Luisier, Mathieu

Show all metadata

Search Results

Publications 1 - 5 of 5

Towards Three-Factor Learning on Valence Change Memory Memristors
Portner, Kevin (2024)
In an era driven by data-intensive applications and artificial intelligence, conventional computing architectures based on the von Neumann architecture face significant challenges in terms of energy and latency due to the physical separation of their memory and processing unit. Inspired by the architecture of the brain, neuromorphic computing offers a promising alternative by co-locating the memory and computing operations and by allowing to implement artificial neural networks directly in hardware. Recently, more biologically plausible network architectures and computing schemes have been explored that enable so-called bio-inspired computing. Along this line, three-factor learning rules do not only respond to the input statistics but also condition learning on reward, punishment, or novelty, and are compatible with reinforcement learning algorithms. To realize them with neuromorphic hardware, three-terminal memristors are needed whose conductance can be independently modulated by the third terminal. The goal of this thesis is to develop electro-optical memristors that enable the realization of three-factor learning rules directly in hardware. Additionally, a bio-inspired framework is presented, allowing for the implementation of three-factor learning through a biologically plausible learning algorithm and network architecture. Together, the electro-optical memristive devices and the bio-inspired framework lay the foundation for future work on a complete three-factor learning hardware implementation. In the first part of this thesis, different variants of three-terminal electro-optical memristors are introduced. In the first generation, two-terminal valence change memory-based memristors in a horizontal configuration are integrated onto an optical silicon waveguide. Under light illumination, more gradual and symmetric switching properties are obtained, which increase the recognition accuracy when performing supervised learning on the MNIST dataset of handwritten digits. Moreover, light-induced switching, where the optical signal acts as the third factor, is demonstrated. A second generation of three-terminal memristors based on a bi-layer oxide is then presented, featuring a vertical electrode configuration. This scalable approach enables integration with crossbar arrays. Lastly, three-terminal electro-optical memristors with novel phase change materials are introduced, using both colloidal quantum dots and molecular telluride inks as active switching materials. The second part of the thesis is dedicated to a bio-inspired framework that enables the implementation of three-factor learning rules using purely electrical memristors. Specifically, temporal difference learning with an actor-critic network architecture is leveraged. The developed framework allows for more bio-inspired computing as it 1) employs a biologically plausible network architecture that mimics reward-based learning in the brain, and 2) performs in-memory training, thereby minimizing off-device calculations. Memristors are not only used for the in-situ training of their synaptic weights, but also to implement the learning rule, the so-called temporal difference error, directly in hardware. The demonstrated framework works with both physical and in-software-emulated memristors and is applied to two common neuroscience navigation tasks, the T-maze and the Morris water-maze. This thesis does not only advance the state-of-the-art in three-terminal memristor technology, but also establishes a foundation to integrate these devices into scalable and biologically inspired computing architectures. It paves the way towards realizing a full hardware implementation of three-factor learning rules using three-terminal electro-optical memristors.
Actor-critic networks with analogue memristors mimicking reward-based learning
Portner, Kevin; Zellweger, Till; Martinelli, Flavio; et al. (2025)
Nature Machine Intelligence
Advancements in memristive devices have given rise to a new generation of specialized hardware for bio-inspired computing. However, most of these implementations draw only partial inspiration from the architecture and functionalities of the mammalian brain. Moreover, the use of memristive hardware is typically restricted to specific elements within the learning algorithm, leaving computationally expensive operations to be executed in software. Here we demonstrate reinforcement learning through an actor-critic temporal difference algorithm implemented on analogue memristors, mirroring the principles of reward-based learning in a neural network architecture similar to the one found in biology. Memristors are used as multipurpose elements within the learning algorithm: they act as synaptic weights that are trained online, they calculate the weight updates associated with the temporal difference error directly in hardware and they determine the actions to navigate the environment. Owing to these features, weight training can take place entirely in memory, eliminating data movement. We test our framework on two navigation tasks-the T-maze and the Morris water maze-using analogue memristors based on the valence change memory effect. Our approach represents the first step towards fully in-memory and online neuromorphic computing engines based on bio-inspired learning schemes.
Conductance-Dependent Photoresponse in a Dynamic SrTiO₃ Memristor for Biorealistic Computing
Weilenmann, Christoph; He, Hanglin; Mladenović, Marko; et al. (2026)
Advanced Functional Materials
Modern computers perform pre-defined operations using static memory components, whereas biological systems learn through inherently dynamic, time-dependent processes in synapses and neurons. The biological learning process also relies on global signals-neuromodulators-that influence many synapses at once, depending on their dynamic, internal state. In this study, using optical radiation as a global neuromodulatory signal, nanoscale SrTiO3 (STO) memristors that can act as solid-state synapses are investigated. It is observed that the memristor's photo-conductance exhibits a long-term decay process after photoexcitation (10s of seconds) with an activation energy of 0.33 eV. Based on density functional theory calculations, this long-term photoresponse is attributed to the generation and migration of oxygen vacancies at the Pt-SrTiO3 interface. Additionally, the photo-conductance decay can be accurately controlled through an electrical bias signal and the magnitude of the memristor's photoresponse depends on its electrical conductance state, following a well-defined square root relation. These properties, in combination with the device's low power operation (< 1pJ per optical pulse) and small measurement variability, may pave the way for space- and energy-efficient implementations of complex biological learning processes in electro-optical hardware.
Electroforming Kinetics in HfOₓ/Ti RRAM: Mechanisms behind Compositional and Thermal Engineering
Kaniselvan, Manasa; Portner, Kevin; Falcone, Donato Francesco; et al. (2025)
ACS Nano
A critical issue affecting filamentary resistive random access memory (RRAM) cells is the requirement of high voltages during electroforming. Reducing the magnitude of these voltages is of significant interest, as it ensures compatibility with complementary metal-oxide-semiconductor (CMOS) technologies. Previous studies have identified that changing the initial stoichiometry of the switching layer and/or implementing thermal engineering approaches has an influence on the electroforming voltage magnitude, but the exact mechanisms remain unclear. Here, we develop an understanding of how these mechanisms work within a standard a-HfOₓ/Ti RRAM stack through combining atomistic driven kinetic Monte Carlo (d-KMC) simulations with experimental data. By performing device-scale simulations at atomistic resolution, we can precisely model the movements of point defects under applied biases in structurally inhomogeneous materials, which allows us to not only capture finite-size effects but also understand how conductive filaments grow under different electroforming conditions. Doing atomistic simulations at the device level also enables us to link simulations of the mechanisms behind conductive filament formation with trends in experimental data for the same material stack. We identify a transition from primarily vertical to lateral ion movement dominating the filamentary growth process in substoichiometric oxides and differentiate the influence of global and local heating on the morphology of the formed filaments. These different filamentary structures have implications for the dynamic range exhibited by formed devices in subsequent SET/RESET operations. Overall, our results unify the complex ion dynamics in technologically relevant HfOₓ/Ti-based stacks and provide guidelines that can be leveraged when fabricating devices.
Insights behind multi-level conductance transitions in HfO𝑥 memristors
Kaniselvan, Manasa; Mladenović, Marko; Clarysse, Jente; et al. (2024)
2024 Device Research Conference (DRC)

Publications 1 - 5 of 5

Kevin Portner

Last Name

First Name

ORCID

Organisational unit

Filters

Search Results