Skip to article frontmatterSkip to article content
ϕ=σ(ω(wx+b))\phi_\ell = \sigma\left( \omega_\ell\left(\mathbf{wx} + \mathbf{b} \right) \right)

where σ is an arbitrary activation function.

h(x;θ(x))\boldsymbol{h}(\mathbf{x};\boldsymbol{\theta}(\mathbf{x}))

For example we have the hypernetwork:

ϕ(x;θ)=wθx+bθ[wθ,bθ]=h(z;θ)\begin{aligned} \phi_\ell(\mathbf{x};\theta) &= \mathbf{w}_\theta\mathbf{x} + \mathbf{b}_\theta \\ [\mathbf{w}_\theta,\mathbf{b}_\theta] &= \boldsymbol{h}(\mathbf{z};\boldsymbol{\theta}) \end{aligned}

We have the additive transformation

ϕ(x;θ)=wx+b+sθsθ=h(z;θ)\begin{aligned} \phi_\ell(\mathbf{x};\theta) &= \mathbf{w}_\ell\mathbf{x} + \mathbf{b}_\ell + \mathbf{s}_\theta \\ \mathbf{s}_\theta &= \boldsymbol{h}(\mathbf{z};\boldsymbol{\theta}) \end{aligned}

We have the affine transformation

ϕ(x;θ)=(wx+b)sθ+aθ[sθ,aθ]=h(z;θ)\begin{aligned} \phi_\ell(\mathbf{x};\theta) &= \left(\mathbf{w}_\ell\mathbf{x} + \mathbf{b}_\ell\right)\mathbf{s}_\theta + \mathbf{a}_\theta \\ [\mathbf{s}_\theta, \mathbf{a}_\theta] &= \boldsymbol{h}(\mathbf{z};\boldsymbol{\theta}) \end{aligned}

Example: 2D Sea Surface Height

In this example, we will train the same SIREN model from before but we will add a modulation extension.