compressed data
inferred codes
uncompressed inferred codes
inferred latent variables
Denoised digits
Denoised digits
Corresponding codes
The joint density factorises in terms of the form p(layer|parents)
With \({\mathbf{z}^{[0]}_n = \mathbf{x}_n}\) and \({L^{[0]} = D}\), that is $$ p(\mathbf{Z}^{[0:K]},\mathbf{U}^{[1:K]},\lambda) = p(\mathbf{Z}^{[K]}) \prod_{k=0}^{K-1} p(\mathbf{Z}^{[k]}|\mathbf{Z}^{[k{+}1]},\mathbf{U}^{[k{+}1]},\lambda^{[k{+}1]})\, p(\mathbf{U}^{[k{+}1]})\, p(\lambda^{[k{+}1]}) $$
Joint \({p(\mathbf{X},\mathbf{Z}_1,\mathbf{Z}_2|\mathbf{U},\lambda_0)}\)
Data layer \({p(\mathbf{X}|\mathbf{Z}_1,\mathbf{U},\lambda_1)}\)
Data layer \({p(\mathbf{Z}_1|\mathbf{Z}_2,\mathbf{U},\lambda_2)}\)
Data layer \({p(\mathbf{Z}_2|\mathbf{Z}_3,\mathbf{U},\lambda_3)}\)
$$ p(u_{ld}=1|\text{rest}) = \sigma \left[\color{red}{ \log\left( \frac{ p(u_{ld}) }{ 1 - p(u_{ld}) } \right)} - \tilde{u}_{ld} \sum\limits_n \left\{ \gamma_{\lambda}(a_{nd}) - \gamma_{\lambda}(a_{nd} - \tilde{u}_{ld}\,\tilde{x}_{nd} (\tilde{s}_{mn} + 1))\right\} \right] $$
E.g. step-exp prior
$$ p(u = 1) = \tfrac{1}{2} \mathrm{H}( 1 - q ) + \tfrac{1}{2} \mathrm{H}(q-1) e^{-a(q-1)} $$
compressed data
inferred codes
reconstruction of codes
latent variables
codes for \({L{=}3}\)
color legend
From the corresponding representations in layer 1 (left) and layer 2 (right)