Interconnecting Spiking Neural Networks

The human brain consists of multiple interconnected regions (such as the motor area, visual area and auditory area of the neocortex, etc.), each responsible for different cognitive and physiological functions. When mapped onto Spiking Neural Networks (SNNs), these regions can be represented as clusters of specialized networks communicating through spikes. Interconnecting SNNs is crucial, because, just like different regions of the brain, specialized neural clusters must communicate efficiently to enable complex intelligence. Building on the idea of memory crossbars as routers, we strive to extend the application for large-scale interconnected spiking neural networks.

contact

Sahibia Kaur Vohra

Researcher

Building 02.5 /
Room 200f

+49 2461/61-96916

E-Mail

Interconnecting Spiking Neural Networks in System-on-Chip (SoC) Architectures

Interconnecting processing units of SNNs requires an efficient routing scheme to convey the information from one cluster to another. By employing interconnecting spiking neural network topology many benefits can be achieved, such as enabling efficient distributed, parallel, and event-driven processing, allowing multisensory integration and implementing spiking associative memories for robust retrieval of information from partial data. Detailing this prospect, a system could consist of SNN clusters to process different information, such as audio sensory, visual sensory and olfactory sensory information. This system can recall entire experiences from partial sensory inputs. For instance, hearing a specific sound might trigger the recall of the associated visual scene and smell. Thus, these applications of associative memory are widely explored for multisensory integration in robotics, speech and gesture recognition systems.

Taking inspiration from the cortical structure of the human brain, which follows the local dense and global sparse connectivity, small clusters of SNN can be connected through in-memory routing. The conventional way of using Address-Event-Representation (AER) as a routing mechanism involves hardware complexity and induced delay. Instead, bio-inspired connectivity employs crossbars of memristors to transfer the information in form of spikes from one cluster to another. Incorporating the sparse connectivity through in-memory distributed memristor crossbars envisions the benefits of high-speed signal propagation with optimal silicon footprint. Consequently, this leads to densely connected smaller crossbars along with sparsely connected crossbars for routing instead of using larger crossbars. Also, it overcomes the drawbacks of current sneak paths, parasitic and large read currents associated with the larger crossbars. Memory crossbar can also be used to implement associative memory, steering the spike information from one processing unit to another. This concept of associating digits to alphabetical data has been realized previously in our works (Prior_work_1, Prior_work_2) for one-to-one and multidirectional associative memory.

The focus of implementing the interconnecting SNN is to design a low-latency, area- and energy-efficient on-chip routing scheme to connect each cluster of SNN, but challenges like synchronization, computation, training, and scalability need to be addressed. Advances in neuromorphic hardware, efficient training methods, and modular architectures will help make interconnected SNNs more practical. Together, advancements at the device level, circuit level, and architecture level can pave the way for the implementation of brain-like interconnection of clusters of SNN.

This research topic is connected to Bio-Inspired Networks & Systems and Neuromorphic Edge Computing Systems.