Periodic Reporting for period 2 - ChipAI (Energy-efficient and high-bandwidth neuromorphic nanophotonic Chips for Artificial Intelligence systems)
Okres sprawozdawczy: 2020-03-01 do 2022-11-30
The Internet, the personal computer and the mobile phone have revolutionized our lives. Within the last few decades, the computing power has increased exponentially to sustain this transformation.
Notably, the recent rise of Artificial Intelligence (AI) systems powered by computers that can learn without the need for explicit instructions is transforming our digital economy and our society as a whole. They use computational deep neural network models inspired by signal processing in the human brain. However, today’s computing hardware, based on von Neumann architectures, is inefficient at implementing these neural networks largely because of the high power consumption per unit area required.
Significant efforts are therefore being made in adapting these electronics-based architectures for future artificial neural networks. Nonetheless, the energy consumption of the massively dense wired interconnects needed to emulate an artificial brain represents a major bottleneck for scalable and portable implementations.
Inspired by the brain, ChipAI addresses these major challenges through the development an energy-efficient neuromorphic nanophotonic architecture technology using neuron-like nanoscale non-linear light sources and detectors to realize interconnected high-bandwidth spike-encoded synapses for optical neural networks, see Figure, and hence capable of addressing the predicted future needs of AI systems and computing processors. By addressing significant material science challenges with innovative experimental and theoretical approaches, this goal will be attained through 4 specific objectives:
1) Demonstrate miniaturized nonlinear light-spiking devices on III-V/Si compound semiconductors for efficient light confinement, emission and detection;
2) Develop energy efficient interconnected synaptic links with flexible electro-optical control of synaptic weights;
3) Proof-of-concept implementation of spike-encoding algorithms for pattern information processing tasks, towards validation of the platform for photonic artificial neural networks;
4) Contribute to position Europe as a world leader in the emerging neuromorphic photonics industry and enhance its global industrial competitiveness in the highly important AI and digital technology economy sector.
Notably, the recent rise of Artificial Intelligence (AI) systems powered by computers that can learn without the need for explicit instructions is transforming our digital economy and our society as a whole. They use computational deep neural network models inspired by signal processing in the human brain. However, today’s computing hardware, based on von Neumann architectures, is inefficient at implementing these neural networks largely because of the high power consumption per unit area required.
Significant efforts are therefore being made in adapting these electronics-based architectures for future artificial neural networks. Nonetheless, the energy consumption of the massively dense wired interconnects needed to emulate an artificial brain represents a major bottleneck for scalable and portable implementations.
Inspired by the brain, ChipAI addresses these major challenges through the development an energy-efficient neuromorphic nanophotonic architecture technology using neuron-like nanoscale non-linear light sources and detectors to realize interconnected high-bandwidth spike-encoded synapses for optical neural networks, see Figure, and hence capable of addressing the predicted future needs of AI systems and computing processors. By addressing significant material science challenges with innovative experimental and theoretical approaches, this goal will be attained through 4 specific objectives:
1) Demonstrate miniaturized nonlinear light-spiking devices on III-V/Si compound semiconductors for efficient light confinement, emission and detection;
2) Develop energy efficient interconnected synaptic links with flexible electro-optical control of synaptic weights;
3) Proof-of-concept implementation of spike-encoding algorithms for pattern information processing tasks, towards validation of the platform for photonic artificial neural networks;
4) Contribute to position Europe as a world leader in the emerging neuromorphic photonics industry and enhance its global industrial competitiveness in the highly important AI and digital technology economy sector.
The envisioned neuromorphic nanophotonic architecture technology in ChipAI, see Figure, requires a combination of material science developments with innovative experimental and theoretical approaches to establish the key building blocks of the architecture. These aspects were defined and established in the first year of the project. All major objectives of the first 12 months have been attained on time:
1) Preliminary epilayer material design on GaAs, InP and Si substrates for the following nanophotonic neuromorphic devices and synaptic interconnects was established (MS2, D2.1): i) nanoLEDs, ii) nanolasers, and iii) nanophotodetectors using III-V and III-V/Si active materials, and iv) synaptic interconnections using silicon photonic materials. The provisional designs were sent to proceed with the material growth;
2) Fabrication process flow for all the target nanophotonic neuromorphic devices using III-V and III-V/Si semiconductor materials was established (D3.1);
3) Preliminary fabrication process flow for fabrication of 3D printing and transfer printing methods for flexible interconnects was established (D3.1);
4) Fabrication process flow for synaptic interconnections using III-V photonic structures on silicon photonics was established (D2.1 D3.1);
5) Preliminary studies from first principle theoretical models were realized to provide full detailed simulations of spike-encoded neural-like dynamics (D2.2). A graphical user interface tool was developed to allow the consortium to evaluate the preliminary design of the nanodevices with neural-like dynamics as a function of experimental parameters;
6) Baseline studies on optimum spike-encoding formats and methods according to brain-inspired spiking-based neural networks were identified (D2.2). This will allow to validate spike-encoding algorithms for pattern information processing tasks using ChipAI technology.
1) Preliminary epilayer material design on GaAs, InP and Si substrates for the following nanophotonic neuromorphic devices and synaptic interconnects was established (MS2, D2.1): i) nanoLEDs, ii) nanolasers, and iii) nanophotodetectors using III-V and III-V/Si active materials, and iv) synaptic interconnections using silicon photonic materials. The provisional designs were sent to proceed with the material growth;
2) Fabrication process flow for all the target nanophotonic neuromorphic devices using III-V and III-V/Si semiconductor materials was established (D3.1);
3) Preliminary fabrication process flow for fabrication of 3D printing and transfer printing methods for flexible interconnects was established (D3.1);
4) Fabrication process flow for synaptic interconnections using III-V photonic structures on silicon photonics was established (D2.1 D3.1);
5) Preliminary studies from first principle theoretical models were realized to provide full detailed simulations of spike-encoded neural-like dynamics (D2.2). A graphical user interface tool was developed to allow the consortium to evaluate the preliminary design of the nanodevices with neural-like dynamics as a function of experimental parameters;
6) Baseline studies on optimum spike-encoding formats and methods according to brain-inspired spiking-based neural networks were identified (D2.2). This will allow to validate spike-encoding algorithms for pattern information processing tasks using ChipAI technology.
In general the achieved results and their respective impact can be divided into three sections:
1) Routes towards miniaturized nonlinear light-spiking neuromorphic devices:
- First demonstration of highly-efficient vertical-light-emitting nanopillars with submicrometer scales showing potential for high-performance nanoLED sources;
- First generation of nanoscale resonant tunnelling diode devices with submicrometer scales and photodetection properties showing potential to attain the extraordinary low power requirements of equivalent biosystems;
- Novel design of a flipped cavity nanolaser on a membrane platform on silicon for high-performance miniaturized nanolaser diode sources.
2) Interconnect synaptic links:
- First ChipAI synaptic interconnects chip fabricated for incoherent and coherent light that match the design specifications in terms of waveguide losses and directional couplers. These will support the design of a synaptic interconnect for a spiking neural network demonstrator, see Figure, until the end of the project.
3) Validation of spike processing and methods for photonic artificial neural networks:
- A graphical user interface that provides automatic parameter scan and evaluation of spike-encoded neural-like dynamics of the ChipAI building blocks technology. Free software can be provided with user-friendly design for scientists and engineers;
- First pattern recognition tasks using laser neurons implemented using ultrafast spiking processing methods. Protocols developed to be validated in the envisioned ChipAI technology until the end of the project.
1) Routes towards miniaturized nonlinear light-spiking neuromorphic devices:
- First demonstration of highly-efficient vertical-light-emitting nanopillars with submicrometer scales showing potential for high-performance nanoLED sources;
- First generation of nanoscale resonant tunnelling diode devices with submicrometer scales and photodetection properties showing potential to attain the extraordinary low power requirements of equivalent biosystems;
- Novel design of a flipped cavity nanolaser on a membrane platform on silicon for high-performance miniaturized nanolaser diode sources.
2) Interconnect synaptic links:
- First ChipAI synaptic interconnects chip fabricated for incoherent and coherent light that match the design specifications in terms of waveguide losses and directional couplers. These will support the design of a synaptic interconnect for a spiking neural network demonstrator, see Figure, until the end of the project.
3) Validation of spike processing and methods for photonic artificial neural networks:
- A graphical user interface that provides automatic parameter scan and evaluation of spike-encoded neural-like dynamics of the ChipAI building blocks technology. Free software can be provided with user-friendly design for scientists and engineers;
- First pattern recognition tasks using laser neurons implemented using ultrafast spiking processing methods. Protocols developed to be validated in the envisioned ChipAI technology until the end of the project.