DeltaEncoder

Contents

DeltaEncoder#

class braintools.spike.DeltaEncoder(threshold=0.1, positive_only=False, absolute=False, normalize=True)#

Encode a time series by send-on-delta (level-crossing) modulation.

The encoder keeps a reference level per feature, starting at the first sample. Whenever the signal has moved at least threshold away from the reference, a spike is emitted and the reference is reset to the current value. Slow drifts are therefore encoded as well as sudden jumps: a ramp of total height \(h\) produces about \(h / \text{threshold}\) spikes.

Parameters:

  • threshold (float) – Change relative to the reference that triggers a spike.

  • positive_only (bool) – Emit spikes only for increases (output 0/1). The reference still follows decreases.

  • absolute (bool) – If False, increases give +1 and decreases -1 (ON/OFF events). If True, every crossing gives 1.

  • normalize (bool) – Rescale each feature to [0, 1] over time first, so threshold is a fraction of the feature’s range.

See also

StepCurrentEncoder

Encodes inputs as constant currents.

Examples

>>> import jax.numpy as jnp
>>> import braintools
>>> encoder = braintools.spike.DeltaEncoder(threshold=0.25, normalize=False)
>>> encoder(jnp.array([0.0, 0.1, 0.2, 0.3, 0.3, 0.0]))
Array([ 0.,  0.,  0.,  1.,  0., -1.], dtype=float32)
>>>
>>> absolute = braintools.spike.DeltaEncoder(threshold=0.25, normalize=False, absolute=True)
>>> absolute(jnp.array([0.0, 0.1, 0.2, 0.3, 0.3, 0.0]))
Array([0., 0., 0., 1., 0., 1.], dtype=float32)
Contents