Periodic Reporting for period 4 - BACKUP (Unveiling the relationship between brain connectivity and function by integrated photonics)
Berichtszeitraum: 2023-05-01 bis 2024-04-30
“BACKUP” addressed the role of neuron activity and plasticity in learning. We developed a simple in-vitro model of the brain's cellular network to study memory formation and neuronal activity. Concurrently, we created an artificial brain unit using photonic integrated circuits, forming a hybrid circuit where biological and artificial neurons collaboratively compute. Light-based optogenetic tools enabled data exchange between biological neurons and photonic circuits, facilitating brain-computer interfacing.
This novel platform aims to heal neurological disorders and add plasticity to hardware, allowing it to change based on experience. This development offers less invasive approaches to neurological disorders and enhances AI by reducing power consumption.
The long-term vision for hybrid neuromorphic photonic networks includes:
• Understanding brain function
• Computing beyond von Neumann architecture
• Controlling and supplementing neuronal functions in disorders.
We demonstrated memory formation through synaptic engrams, induced and suppressed hyper-excitability in neuronal cultures for epilepsy treatment, and showed photonic neural networks performing complex tasks. These advancements enable learning and recalling in applications like optical signal equalization and counterfeit banknote recognition. Finally, we interfaced neuronal cultures with a photonic neural network, showing the potential for hybrid artificial neuronal circuits
This novel platform aims to heal neurological disorders and add plasticity to hardware, allowing it to change based on experience. This development offers less invasive approaches to neurological disorders and enhances AI by reducing power consumption.
The long-term vision for hybrid neuromorphic photonic networks includes:
• Understanding brain function
• Computing beyond von Neumann architecture
• Controlling and supplementing neuronal functions in disorders.
We demonstrated memory formation through synaptic engrams, induced and suppressed hyper-excitability in neuronal cultures for epilepsy treatment, and showed photonic neural networks performing complex tasks. These advancements enable learning and recalling in applications like optical signal equalization and counterfeit banknote recognition. Finally, we interfaced neuronal cultures with a photonic neural network, showing the potential for hybrid artificial neuronal circuits
The first brain-inspired photonic integrated circuits have been designed, fabricated, tested, and validated to perform complex computational tasks. Memory is incorporated using the intrinsic nonlinearity of optical microring resonators and external optical feedback. Demonstrated neural networks range from simple optical perceptrons for telecom applications to arrays of nonlinear optical nodes for digit recognition and time series forecasting. A significant advancement is the creation of photonic integrated spiking neural networks with microring resonators, featuring short- and long-term memories, proposed for edge computing in sensing applications.
Using optogenetics and patterned light excitation in cultured neurons, we investigated synaptic strengthening and memory formation. We developed an optical platform for single-cell optogenetic experiments in vitro, utilizing multiple wavelengths to induce excitation or inhibition. This platform, coupled with labeling techniques or electrophysiological monitoring, allows selective excitation or inhibition of cells in a neuronal network, aiding molecular, physiological, and morphological characterization.
A photonic chip and neuronal culture have been designed, allowing optical signals to activate specific neurons, while neuronal activity influences the photonic neural network, establishing a bidirectional link.
Deep learning algorithms modeled biological network dynamics, initially in an artificial network and later in a hybrid artificial-biological network. Our Reservoir Computing Network (RCN) model accurately predicts network connectivity and simulates responses to stimuli.
Dissemination activities included over 50 presentations at conferences and workshops, 30 published articles, and organizing one winter school and two workshops. Six Master’s theses were completed, a patent was filed internationally, and an ERC-POC grant was approved. Numerous spin-off projects have been submitted for evaluation.
Using optogenetics and patterned light excitation in cultured neurons, we investigated synaptic strengthening and memory formation. We developed an optical platform for single-cell optogenetic experiments in vitro, utilizing multiple wavelengths to induce excitation or inhibition. This platform, coupled with labeling techniques or electrophysiological monitoring, allows selective excitation or inhibition of cells in a neuronal network, aiding molecular, physiological, and morphological characterization.
A photonic chip and neuronal culture have been designed, allowing optical signals to activate specific neurons, while neuronal activity influences the photonic neural network, establishing a bidirectional link.
Deep learning algorithms modeled biological network dynamics, initially in an artificial network and later in a hybrid artificial-biological network. Our Reservoir Computing Network (RCN) model accurately predicts network connectivity and simulates responses to stimuli.
Dissemination activities included over 50 presentations at conferences and workshops, 30 published articles, and organizing one winter school and two workshops. Six Master’s theses were completed, a patent was filed internationally, and an ERC-POC grant was approved. Numerous spin-off projects have been submitted for evaluation.
Neuromorphic computing hardware requiring conventional backpropagation training is hard to scale due to the need for full observability of network states and programmability of parameters. Thus, finding hardware-friendly, biologically-plausible learning schemes and platforms is crucial. We demonstrated a photonic integrated neural network using a massive array of microring resonators with rich recurrent nonlinear dynamics and both short- and long-term plasticity. Scalability is enhanced by processing input and generating output encoded concurrently in temporal, spatial, and wavelength domains.
We showcased a high-performing feed-forward photonic neural network for chromatic dispersion compensation in Intensity Modulation/Direct Detection optical links, achieving effective equalization for a 20 Gbps 4-level Pulse Amplitude Modulated signal up to 125 km. An evolutionary algorithm and a gradient-based approach were compared for training, focusing on repeatability and convergence time. The optimal weights from training were analyzed using the theoretical transfer function of the optical fiber. Simulations showed scalability to larger bandwidths, up to 100 Gbps.
An all-optical design for hybrid neurons and a photonic chip is proposed. Neurons are stimulated by light scattered from the photonic chip, with emissions from chromophores stimulated and collected by the chip, allowing light to write and read from neurons. The interface between neuronal cultures and microring resonators (the artificial neurons) translates the electrical activity of cells into an optical output, potentially connecting the chip's sensing area with biological neurons to a neuromorphic circuit.
Modeling of the neuronal culture was performed using a random walk model of memory, leading to mathematical equations for network dynamics. Simulations revealed three possible memory regimes, with results showing metaplastic memory retrieval even after plastic memory erasure. A reservoir computing model of the neuronal culture was benchmarked for connectivity retrieval using multi-electrode array experiments. Experimental responses to stimuli were compared with model predictions.
The random walk model was applied to a structured network topology to reduce computational costs. This reservoir computing model could study neuronal functionality and circuitry. Using optogenetic strategies, we created an artificial synaptic engram, showing that astrocytes are unnecessary for engram formation. We are developing methods to control epileptic hyperexcitability in neuronal cultures using integrated photonic circuits in a closed-loop system, involving the creation of a hyperexcitable network and the extinction of epileptic foci with patterned light excitation.
Our research on primary neuronal culture responses to convulsant drugs has highlighted integrated photonic circuits as promising for modulating neural activity. Ongoing experiments suggest these circuits could provide precise interventions for neural activity, offering a method to study drug impacts on neural circuits. We designed a closed-loop system to monitor cell activity and use localized optogenetic inhibition to control abnormal activity in specific target electrodes through light stimulation, demonstrating potential in managing epileptic activity.
We showcased a high-performing feed-forward photonic neural network for chromatic dispersion compensation in Intensity Modulation/Direct Detection optical links, achieving effective equalization for a 20 Gbps 4-level Pulse Amplitude Modulated signal up to 125 km. An evolutionary algorithm and a gradient-based approach were compared for training, focusing on repeatability and convergence time. The optimal weights from training were analyzed using the theoretical transfer function of the optical fiber. Simulations showed scalability to larger bandwidths, up to 100 Gbps.
An all-optical design for hybrid neurons and a photonic chip is proposed. Neurons are stimulated by light scattered from the photonic chip, with emissions from chromophores stimulated and collected by the chip, allowing light to write and read from neurons. The interface between neuronal cultures and microring resonators (the artificial neurons) translates the electrical activity of cells into an optical output, potentially connecting the chip's sensing area with biological neurons to a neuromorphic circuit.
Modeling of the neuronal culture was performed using a random walk model of memory, leading to mathematical equations for network dynamics. Simulations revealed three possible memory regimes, with results showing metaplastic memory retrieval even after plastic memory erasure. A reservoir computing model of the neuronal culture was benchmarked for connectivity retrieval using multi-electrode array experiments. Experimental responses to stimuli were compared with model predictions.
The random walk model was applied to a structured network topology to reduce computational costs. This reservoir computing model could study neuronal functionality and circuitry. Using optogenetic strategies, we created an artificial synaptic engram, showing that astrocytes are unnecessary for engram formation. We are developing methods to control epileptic hyperexcitability in neuronal cultures using integrated photonic circuits in a closed-loop system, involving the creation of a hyperexcitable network and the extinction of epileptic foci with patterned light excitation.
Our research on primary neuronal culture responses to convulsant drugs has highlighted integrated photonic circuits as promising for modulating neural activity. Ongoing experiments suggest these circuits could provide precise interventions for neural activity, offering a method to study drug impacts on neural circuits. We designed a closed-loop system to monitor cell activity and use localized optogenetic inhibition to control abnormal activity in specific target electrodes through light stimulation, demonstrating potential in managing epileptic activity.