Impact of Changing Noise Tensors

We take 7 seeds and permute their noise tensors. The question is whether we can observe movement in p_hat_release from only a re-sampled noise tensor list … the answer is yes !

The plot is structured like this:

• Each color represents a specific seed from 100-106
• The y-axis has the p_hat_cnn for release
• The x-axis is the respective noise tensor id. So x=1 represents the noise tensors from the first seed

We can clearly observe a difference in p_hat_release as the x-axis varies !

Plot 1 - P-Hat Differences with varying noise

• Here we take a seed and observe that as we change the noise tensors (x-axis) the p_hat_relase predictions change for each seed

Plot 2 - Constant Latent Tensors

• Here the x-axis represents a seeds latent tensor, and I vary the noise-tensor input.
• Clearly as we change the noise tensors, we create a distribution of p_hat_release outcomes

Plot 3 - Constant Noise Tensor

• Here we keep the noise tensor constant in the x-axis
• We vary the latent tensor to which the noise is applied
• We still observe a spread for each noise tensor