Semantic Communication

Definition:

Joint Source-Channel Coding (JSCC)

JSCC with deep learning maps source directly to channel symbols:

x^=Dϕ(H⋅Eθ(x)+n)\hat{\mathbf{x}} = D_\phi(\mathbf{H} \cdot E_\theta(\mathbf{x}) + \mathbf{n})

where EθE_\theta is a neural encoder and DϕD_\phi is a neural decoder. This bypasses the separation theorem for finite block lengths.

Example: Deep JSCC for Images

Implement a simple deep JSCC system for image transmission.

Solution
Architecture
class DeepJSCC(nn.Module):
    def __init__(self, n_channels):
        super().__init__()
        self.encoder = nn.Sequential(
            nn.Conv2d(3, 64, 3, padding=1), nn.ReLU(),
            nn.Conv2d(64, n_channels, 3, padding=1),
        )
        self.decoder = nn.Sequential(
            nn.Conv2d(n_channels, 64, 3, padding=1), nn.ReLU(),
            nn.Conv2d(64, 3, 3, padding=1), nn.Sigmoid(),
        )
LLMs as Reasoning Engines for System DesignFoundation Models for Wireless