Insights
Extracting automatic insights using Tensorleap
Last updated
Extracting automatic insights using Tensorleap
Last updated
Tensorleap insights could be viewed via the insights panel which is shown once you click on the dashboard view at the top. In order to see insights, a model must be loaded and evaluated or trained using the Tensorleap system.
Insights are essentially clusters in your model's latent space, that have unique properties that should be considered for the sake of model & dataset analysis.
Each Cluster is composed of a group of samples, that have some shared features, and a unique behaviour as detected by the TensorLeap platform.
We detect 4 types of clusters:
High-Loss Cluster - a collection of similar samples that the model performs poorly on.
Overfitting Cluster - a collection of samples on which the model scored significantly better on the train subset than on the validation subset.
Repetitive Cluster - a collection of samples, that have very low variance in features, compared to the rest of the samples.
Underrepresentation Cluster - a collection of samples that are composed of an uneven representation from the validation and train subsets, i.e. the cluster has a significantly higher number of samples that are from the training subset than the validation subset or vice versa.
On the top-right of the insights panel the number of insights, per type, is shown. Clicking the Display button on the insights will refer to the display of the specific cluster.
Once we filter the relevant samples that make out an interesting cluster we can use Tensorleap Sample Analysis to better understand each sample and analyze the root cause for the cluster's behaviour.
