Figure 1: Neural Decoder Architecture for EEG-to-Image Reconstruction
The decoder employs a multi-layer perceptron with bottleneck architecture to transform 512-dimensional EEG embeddings into 28×28 pixel reconstructions. Each hidden layer incorporates ReLU activation, dropout regularization (p=0.2), and batch normalization for stable training dynamics.
Input Layer
EEG Embeddings
512D
Enhanced Transformer
Features
→
Hidden Layer 1
Linear + ReLU
1024D
Dropout (0.2)
BatchNorm1D
→
Hidden Layer 2
Linear + ReLU
2048D
Dropout (0.2)
BatchNorm1D
(Bottleneck)
→
Hidden Layer 3
Linear + ReLU
1024D
Dropout (0.2)
BatchNorm1D
→
Output Layer
Linear + Tanh
784D
Reshape to
28×28 Image
Reconstructed Image Output (28×28)
Pixel values ∈ [-1, 1]
Mathematical Formulation:
h₁ = BatchNorm(Dropout(ReLU(W₁e + b₁)))
h₂ = BatchNorm(Dropout(ReLU(W₂h₁ + b₂)))
h₃ = BatchNorm(Dropout(ReLU(W₃h₂ + b₃)))
y = tanh(W₄h₃ + b₄)

Loss Function: L(θ) = (1/N) Σᵢ₌₁ᴺ ||yᵢ - ŷᵢ||₂²
Layer Input Dimension Output Dimension Parameters Activation Regularization
Input 512 512 0 - -
Hidden 1 512 1024 525,312 ReLU Dropout(0.2) + BatchNorm
Hidden 2 1024 2048 2,099,200 ReLU Dropout(0.2) + BatchNorm
Hidden 3 2048 1024 2,098,176 ReLU Dropout(0.2) + BatchNorm
Output 1024 784 803,616 Tanh -
Model Complexity Analysis:
Total Parameters: 5,526,304 (≈ 5.5M)
Computational Complexity: O(d₁d₂ + d₂d₃ + d₃d₄ + d₄d₅)
Memory Footprint: ~22 MB (FP32)
Forward Pass Operations: ~11.4 MFLOPs per sample