Spike-based neuromorphic hardware with a long short-term memory for AI applications
The promise of spike-based neuromorphic technology is that it will provide more energy efficient implementations for Deep Neural Networks than standard hardware, such as GPUs. This requires understanding how DNNs are emulated using an event-based sparse-firing regime. Otherwise, the energy advantage is lost. DNNs which solve sequence-processing tasks usually use Long Short Term Memory (LSTM), but these are difficult to emulate when there are few spikes. We show that a facet of many biological neurons, slow after-hyperpolarizing (AHP) currents after each spike, provides an efficient solution. AHP-currents are easily implemented in neuromorphic chips that support multi-compartment models of neurons, such as Intel’s Loihi Chip. The filter approximation hypothesis explains how AHP-neurons are able to emulate the functions of LSTM units.