Why spiking networks?

Artificial Neural Networks (ANNs) excel at pattern recognition and decision-making, but biological brains are fundamentally more capable; and brains solve harder problems while consuming vastly less energy! The key difference lies in how neurons compute.

Biological neurons are hybrid systems: They gather analog information internally through continuous chemical and electrical processes, but communicate digitally via discrete electrical pulses called action potentials (also known as spikes) Neumann, 2012. This combination gives brains - both the precision of digital computation and the efficiency of analog processing.

Spiking Neurons are computational models that capture this dual nature. They integrate incoming signals over time and emit discrete spikes when certain conditions are met; this is unlike the artificial neurons (e.g., $\texttt{ReLU}$). When networks of spiking neurons are deployed on specialized neuromorphic hardware, they achieve:

• Higher computational expressivity: Time-based encoding and processing carries more information
• Greater energy efficiency: Events are processed only when spikes occur, thus, inherently sparser
• Lower latency: Asynchronous processing enables real-time responses, aided by locality of memory and compute

The challenge: SNNs are harder to understand and train than the conventional ANNs. Working with and leveraging the temporal dynamics and discrete spike events require different mathematical tools and training methods. Moreover, adapting and implementing the SNNs on emerging neuromorphic hardware is also not straightforward and standardized; not to mention, designing and building neuromorphic chips/systems is another complex task.

Our goal: Give you the intuition and practical skills to design, build, train, and deploy SNNs on neuromorphic hardware effectively. In this book, we start with the fundamentals and build toward real-world applications with neuromorphic systems.