Dashboard Filtering

The Dashboarding system allows a user to quickly filter out specific subsets to review. Instead of examining the entire dataset correlation in a line plot or visualizing it in the population exploration, the user can quickly filter for a specific, interesting slice of the dataset. Common examples are:

• Examining high-loss or errornous slices

• Filtering images that have a specific object/class of interest

• Examinig a single set of the data (train/val/test/unlabeled)

Local And Global Filters

The Dashboard support two type of filters:

• Local filters - that effect only the behaviour of the current dashlet and related assets

• Global filters - that effects every dashlet in the dashboard

How to add a filter

To add a filter in the dashboard, you use the (+) icon, usually located at the top of each dashlet, and at the top of the dashboard itself.

After clicking the + sign, a drop-down menu opens, with the filter conditions.

Filling in the field to filter on, the operator to use (Equal, Greater than, etc.), and the value to evaluate the operator would create a filter.

After a filter is added, it can be seen next to the filter icon in the dashlet on which it was created.

Edit and remove an existing filter

Hovering over an existing filter allows to perform several actions:

• Edit a filter - allows to change the settings of the filter

• Pin a filter - makes a filter shared between users

• Disable a filter - disables the current filter but does not remove it

• invert a filter - invert the filter condition

• remove a filter - remove the filter.

Apllying a local filter Globally

At any point, the set of local filters applied at a dahlet level could be moved to global scope to be applied to the entire dashboard. This is done by clicking the green "up" button that is adjacent to the local filters.

A common use-case for this global applying of filters is if one dashlets shown an interesting property on a subpopulations - which we would like to now view throughout the dashboard.

Thus for example, seeing a correlation between high brightness of the image and the loss in a line plot, could be used to find specific brightness level where the loss is high and creating a local filter from it - and then applying it globally and examining the population exploration to qualitatively examine which samples are affiliated with this slice.