Encoder Network

The encoder extracts hierarchical spatio-temporal features through three spiking convolutional layers with interleaved max pooling.

Data Flow

flowchart TD
    IN["Input (B, C_in, H, W)"] --> C1
    C1["Conv1: 5x5, 4 features, LIF (leak 90%)"] --> P1
    P1["Pool1: 2x2 max pool (indices saved)"] --> C2
    C2["Conv2: 5x5, 36 features, LIF (leak 10%)"] --> P2
    P2["Pool2: 2x2 max pool (indices saved)"] --> C3
    C3["Conv3: 7x7, N_class features, IF (no leak)"] --> OUT
    OUT["Classification Spikes (T, B, N_class, H', W')"]

Layer Specifications

Property	Conv1	Conv2	Conv3
Kernel	5x5	5x5	7x7
Output channels	4	36	$n_\text{classes}$
Neuron	LIF (subtractive)	LIF (subtractive)	IF
Threshold $\theta$	10.0	10.0	10.0
Leak $\lambda$	9.0 (90%)	1.0 (10%)	0.0
Trained by	Fixed DoG/Gabor	STDP + WTA	STDP
Weight init	$\mathcal{N}(0.8, 0.01)$	$\mathcal{N}(0.8, 0.01)$	$\mathcal{N}(0.8, 0.01)$

Spiking Convolution

At layer $l$, the input current arriving at neuron $(c, h, w)$ is:

$$I_l[c, h, w] = \sum_{c'} \sum_{k_h} \sum_{k_w} W_l[c, c', k_h, k_w] \cdot S_{l-1}[c', h + k_h, w + k_w]$$

$S_{l-1}$ is the binary spike tensor from the previous layer; weights $W_l \in [0, 1]$.

Pooling

Max pooling selects the highest activation and stores the winning index for the decoder:

$$y[c, h, w],\; \text{idx}[c, h, w] = \max_{(k_h, k_w) \in \mathcal{K}} x[c, h \cdot s + k_h, w \cdot s + k_w]$$

EncoderOutput

The encoder returns an EncoderOutput dataclass:

Field	Type	Description
classification_spikes	Tensor	Spikes from Conv3 (T, B, C, H', W')
pooling_indices	PoolingIndices	Pool1/Pool2 indices and output sizes
layer_spikes	Dict[str, Tensor]	Spikes at each layer (conv1, conv2, conv3)
layer_spike_times	Dict[str, Tensor]	First-spike times per layer
has_spikes	bool	Whether any classification spikes occurred
n_classification_spikes	int	Total classification spike count

Paper Presets

from spikeseg.models import SpikeSEGEncoder
from spikeseg.models.encoder import EncoderConfig

# IGARSS 2023 configuration
config = EncoderConfig.from_paper("igarss2023", n_classes=1)
encoder = SpikeSEGEncoder(config)

See API: Models for all constructor parameters.