exp1_gaussian_mixture_high_density-larger-beta-sweep Results

Discrete Latent Summary

Beta	NMI	Accuracy	Recon	KL final
`0.10`	`0.762`	`0.729`	`6.488`	`1.020`
`0.25`	`0.762`	`0.729`	`6.548`	`0.895`
`0.50`	`0.735`	`0.717`	`6.679`	`0.670`
`1.00`	`0.720`	`0.704`	`7.103`	`0.359`
`2.00`	`0.720`	`0.704`	`7.788`	`0.035`
`5.00`	`0.690`	`0.700`	`7.901`	`0.001`

Continuous Latent Summary

Beta	Probe accuracy	Recon	KL final	Mean radius
`0.10`	`1.000`	`3.272`	`5.125`	`2.935`
`0.25`	`1.000`	`3.521`	`3.482`	`2.451`
`0.50`	`1.000`	`3.909`	`2.477`	`2.062`
`1.00`	`1.000`	`4.895`	`1.322`	`1.533`
`2.00`	`1.000`	`6.505`	`0.421`	`0.875`
`5.00`	`0.979`	`7.827`	`0.047`	`0.271`

Figures

Caption: Discrete training history metrics across the beta sweep.

Caption: Continuous training history metrics across the beta sweep.

Discrete metric heatmap and confusion matrices. Caption: Discrete latent summary metrics and confusion matrices across beta.

Reconstruction versus KL tradeoff across latent type and beta. Caption: Reconstruction-versus-KL tradeoff comparing discrete and continuous latent families.