Periodic Reporting for period 2 - ADOPD (Adaptive Optical Dendrites)
Okres sprawozdawczy: 2021-10-01 do 2023-03-31
The increased demand for computation with low energy consumption requires entirely novel hardware concepts that combine high computational speed with a low energy budget per calculation. In ADOPD, we develop ultra-fast computational units based on optical-fiber technologies exploiting information processing principles used by neurons in their dendritic trees. Neurons can process multiple signals on their dendrites in a highly condensed, local, and parallel manner. Those processing capabilities are adaptive and (often) non-linear, which allows a neuron to perform complex calculations, for example for recognizing spatial-temporal patterns in their incoming signals. These properties are first modeled on conventional computers and in a second step transferred to optical systems consisting of fiber optics as well as other optical components. For the first prototype, ADOPD uses well-established single-mode fiber technology to build an optical-dendritic unit. Similar to a real dendrite, this unit is adaptive using principles of neural plasticity to learn from experience. This allows the system to adapt to changing signal structures and we show that such an optical-dendritic unit can be used for adaptive noise reduction in a technical application. From there, we move on to cutting-edge multi-mode fibers to obtain an all-optical second prototype of a dendritic tree with significantly higher computing power and compactness. This work is complemented by theoretical studies that address how such optical-dendritic units could be combined into larger networks where we quantify the computational efficiency of such multiple, parallel operating devices. Thus, the optical dendritic units created by ADOPD represent a novel, cutting-edge computing hardware for fast, low-power, parallel computing, with the potential to help addressing the rising demands for computation.
This project divides into four workpackages that address R&D aspects. WP1 and WP4 provide theoretical contributions and work now closely together on the aspect of context-dependent input-correlation learning (ICO learning). WP2 and WP3 address the technical implementation of the theory in single mode and multi-mode optical fiber systems.
In detail: WP1 is offering novel biophysical insights on heterosynaptic (“across synapse”) dendritic plasticity based on Calcium diffusion and, in addition, we have investigated a novel dendritic learning rule with learning rate annealing to achieve stability on the network level. This aspect extends the planned work for ADOPD into a complex neuro-biophysical domain. However, in addition to this, and directly linked to the central goals of ADOPD, we report that the core contributions of WP1 on investigating the ICO-learning rule have directly led to an optical implementation in WP2 for single dendritic function on single mode fibers. The context dependent aspects of WP4 are a natural extension to multiple dendrites, which are now opening up another venue for optical implementation. This venue is currently under consideration for the next period of WP2. WP3 has naturally a strong connection to WP2 and it has done significant ground work for abstracting dendritic functions into parallel optical computation via the use of different speckle patterns that arise in multi- or few-mode fibers (MMF, FMF). This is an important conceptual step, as it may allow subsuming the computations of different dendritic branches, investigated in WPs1 and 4, in a holographic manner into the same MMF or FMF. Hereby the preference of ADOPD lies now on FMFs, because of the lower complexity of the resulting speckle pattern and the resulting smaller sensitivity to noise in the read-out process. Accordingly, a new FMF, specific for our purpose, is currently being fabricated and will soon be tested. WP4 has provided deeper insights into context-dependent dendritic processing on multiple branches using different structures and learning rules, where – as mentioned above – the context-dependent ICO rule provides a direct link to WP1 and to the implementation in WP2.
In detail: WP1 is offering novel biophysical insights on heterosynaptic (“across synapse”) dendritic plasticity based on Calcium diffusion and, in addition, we have investigated a novel dendritic learning rule with learning rate annealing to achieve stability on the network level. This aspect extends the planned work for ADOPD into a complex neuro-biophysical domain. However, in addition to this, and directly linked to the central goals of ADOPD, we report that the core contributions of WP1 on investigating the ICO-learning rule have directly led to an optical implementation in WP2 for single dendritic function on single mode fibers. The context dependent aspects of WP4 are a natural extension to multiple dendrites, which are now opening up another venue for optical implementation. This venue is currently under consideration for the next period of WP2. WP3 has naturally a strong connection to WP2 and it has done significant ground work for abstracting dendritic functions into parallel optical computation via the use of different speckle patterns that arise in multi- or few-mode fibers (MMF, FMF). This is an important conceptual step, as it may allow subsuming the computations of different dendritic branches, investigated in WPs1 and 4, in a holographic manner into the same MMF or FMF. Hereby the preference of ADOPD lies now on FMFs, because of the lower complexity of the resulting speckle pattern and the resulting smaller sensitivity to noise in the read-out process. Accordingly, a new FMF, specific for our purpose, is currently being fabricated and will soon be tested. WP4 has provided deeper insights into context-dependent dendritic processing on multiple branches using different structures and learning rules, where – as mentioned above – the context-dependent ICO rule provides a direct link to WP1 and to the implementation in WP2.
The ADOPD project exceeds the state of the art in two main fields which are neuro-theory of dendritic computation as well as advanced multi-branch, adaptive photonic systems for ultra-high-speed computations. Specifically, we report on four aspects that can create direct economic impacts, namely:
1) Dendrite-based computation unit towards a photonic integrated circuit platform: Microscopic optical and electro-optical components enable a high degree of integration at the current state of the art. The photonic integrated circuit (PIC) platforms (Silicone, Siliconenitride) are offering an excellent perspective for massively parallel implementation of analog computing paradigms, while relying on well understood semiconductor technologies. The PIC platform promises energy, space and cost saving solutions for analogue optical computing in future. To address this, at this stage, members of the consortium have made a first round of consultancy with VLC Photonics, where we currently evaluate the possible integration technologies and the supporting fabrication platforms.
2) Adaptive Method for Synchronization of Electrical and Optical Signals (Skew Compensation) - ADSYNC (a patent based on ICO learning): This is a novel invention, based on the ICO learning rule, which is central to the ADOPD project. The transmission of electrical or optical signal groups often leads to different travel times and asynchronous arrival at a detector. This can occur due to multiple reasons including, but not restricted to, differences in production (material properties), thermally induced differences, different transmission delays in the system, etc., even if the signal source is identical. In addition, widening of pulse groups or other distortions can happen. Processing of such signals requires prior temporal synchronization for which ADSYNC offers a powerful novel and adaptive solution, which avoids controller recalibration. It has the potential to create impact in different fields where signal synchronization is required. These fields are manifold especially in wireless transmission systems.
3) Implementation of an extended fading memory optical cell: This memory cell is the first that allows storing numerical values by ways of light. We have investigated different versions of the optical memory cell, with a final assessment that the memory is fading and not static. The extended memory that offers is of high interest, especially in systems that exploit this type of memory for computing (e.g. reservoir computing). This sub-system needs to be tested in the context of an operation that requires these memory attributes especially here into different photonic reservoir prototypes that are available.
4) Design and manufacturing of an all-silica step-index few-mode optical fibre. This offers an interesting approach for increasing the bandwidth of optical networks by using few-mode optical fibers in combination with space-division-multiplexing techniques. We cooperate with external partners (TU Dresden) on the experimental realization of this which should be impactful for the telecom/datacom industry.
1) Dendrite-based computation unit towards a photonic integrated circuit platform: Microscopic optical and electro-optical components enable a high degree of integration at the current state of the art. The photonic integrated circuit (PIC) platforms (Silicone, Siliconenitride) are offering an excellent perspective for massively parallel implementation of analog computing paradigms, while relying on well understood semiconductor technologies. The PIC platform promises energy, space and cost saving solutions for analogue optical computing in future. To address this, at this stage, members of the consortium have made a first round of consultancy with VLC Photonics, where we currently evaluate the possible integration technologies and the supporting fabrication platforms.
2) Adaptive Method for Synchronization of Electrical and Optical Signals (Skew Compensation) - ADSYNC (a patent based on ICO learning): This is a novel invention, based on the ICO learning rule, which is central to the ADOPD project. The transmission of electrical or optical signal groups often leads to different travel times and asynchronous arrival at a detector. This can occur due to multiple reasons including, but not restricted to, differences in production (material properties), thermally induced differences, different transmission delays in the system, etc., even if the signal source is identical. In addition, widening of pulse groups or other distortions can happen. Processing of such signals requires prior temporal synchronization for which ADSYNC offers a powerful novel and adaptive solution, which avoids controller recalibration. It has the potential to create impact in different fields where signal synchronization is required. These fields are manifold especially in wireless transmission systems.
3) Implementation of an extended fading memory optical cell: This memory cell is the first that allows storing numerical values by ways of light. We have investigated different versions of the optical memory cell, with a final assessment that the memory is fading and not static. The extended memory that offers is of high interest, especially in systems that exploit this type of memory for computing (e.g. reservoir computing). This sub-system needs to be tested in the context of an operation that requires these memory attributes especially here into different photonic reservoir prototypes that are available.
4) Design and manufacturing of an all-silica step-index few-mode optical fibre. This offers an interesting approach for increasing the bandwidth of optical networks by using few-mode optical fibers in combination with space-division-multiplexing techniques. We cooperate with external partners (TU Dresden) on the experimental realization of this which should be impactful for the telecom/datacom industry.