Pulsformer

Pulsformer is a term used in machine learning research to describe a class of neural network architectures that integrate pulse-based or event-driven input representations with Transformer-style attention mechanisms. The goal is to model temporally sparse or irregular data efficiently by encoding input events as discrete pulses with precise timestamps and then processing them with attention to capture long-range dependencies.

In typical pulsforms, an input module converts streams of events into a sequence of pulses that carry both feature values and time information. A temporal encoding layer may apply time decay, relative time embeddings, or continuous-time attention to reflect the distances between pulses. The core attention mechanism resembles the standard Transformer but is adapted for irregular sequences, often using causal masks or time-aware position encodings to preserve sequential order and enable online processing. Some designs incorporate spiking-neuron-inspired dynamics to control pulse amplitudes or gating between layers, yielding a hybrid between neuromorphic and transformer paradigms.

Variants include encoder-only pulsformers for classification, encoder-decoder pulsformers for sequence-to-sequence tasks, and lightweight versions optimized for low-latency deployment on event streams. Training often leverages standard transformer objectives with additional loss terms that encourage faithful temporal reconstruction or energy-efficient representations.

Applications span event-based vision from neuromorphic cameras, processing of neural time-series data (such as EEG and MEG), sensor networks with irregular sampling, and other domains where inputs arrive asynchronously. The pulsformer concept emerged in the late 2010s and early 2020s as researchers explored combining event-based encoding with transformer attention. It remains a developing area with multiple implementations and no single canonical architecture.