Membrane Potential
A core state variable in spiking neurons such as the LIF model. It accumulates input current and triggers a spike emission when it exceeds a threshold, carrying the neuron's temporal integration history.
Timeline (last 90 days)
In spiking Transformers, membrane variables are usually as important as weights, because the query, key, and value computations in the attention mechanism all depend on spike sequences, and small deviations in membrane variables propagate and accumulate across time steps, causing firing patterns to deviate.
DAQ constructs an objective function that encourages the quantized membrane variables to remain within the reference firing region along the quantized model's own state trajectory, thereby reducing the disruption of quantization to spike firing patterns.
All Facts (2)
In spiking Transformers, membrane variables are usually as important as weights, because the query, key, and value computations in the attention mechanism all depend on spike sequences, and small deviations in membrane variables propagate and accumulate across time steps, causing firing patterns to deviate.
50%UnverifiedDAQ constructs an objective function that encourages the quantized membrane variables to remain within the reference firing region along the quantized model's own state trajectory, thereby reducing the disruption of quantization to spike firing patterns.
50%