If probabilistic layer gives a distribution as output and sampling from them using model.predict gives non deterministic values then how can we sure that this is the exact value

TensorFlow General Discussion
,
1

In traditional machine learning models, we use predict and it gives the point that always lies on line of best fit. Taking example of regression, but in reality the point could live in variance of line best fit.

Since the output of tensorflow probability model having distribution lambda in output gives the distribution that represents the data coming from previous layers.

So lets say I am customer and devs have used model.predict(x) and it gives me different values for same input, how can I trust whether single output is correct?

3

Hi @tbhaxor ,

Probabilistic models like TensorFlow Probability, predictions output a distribution capturing both the mean (best estimated values) and variance (uncertainty in values). While traditional models provide a single point, probabilistic models reflect variability in predictions.

To trust an output, use the mean for a reliable estimate and assess the variance for confidence. High variance indicates uncertainty, while low variance suggests a more trustworthy prediction, allowing for better decision-making based on uncertainty.

Thank You