Population Exploration

This describes the population exploration dashlet

AThe Population Exploration dashlet

Population Exploration Dashlet

The Population Exploration dashlet visualizes the dataset in a model-specific latent space that Tensorleap automatically extracts and optimizes. This high-dimensional representation is then reduced to 2D using techniques like t-SNE or PCA for interactive exploration. Each point represents a sample, where spatial proximity reflects semantic similarity—closer points are likely to share meaningful features. The selected samples for visualization is a subset of the currently filtered population. The size of the visualized subset appears in the top left side of the dahlet, and is controllable via the settings.

Point properties such as color, size, and hover behavior can be customized using dataset metadata or Tensorleap’s computed metrics.

In this guide, we’ll explore the dashlet’s functionality, properties, and configuration settings.

The top bar

The Top Panel of the population exploration, controls dashlet settings, dot color shape and size, filtering and enables a run of a visualization and locating specific samples.

Changing dot color, size, and hover behaviour

At the top of the Population Exploration dashlet, three dropdown menus control the visualization behavior:

• Hover: Sets the information displayed when hovering over a point.

• Color: Determines the color of each point based on a selected metric or metadata.

• Size: Controls the size of each point using a chosen value.

These controls make it easy to explore and highlight different correlations in the dataset by dynamically mapping model insights or metadata to visual properties.

Filtering

Through the top bar we can also add a filter to the population exploration dashlet (+ icon), controlling which subset of the data do we want to visualize (High loss samples, Only a specific geographic area, a specific label, etc.).

The insights menu is coupled to the currently examined population. Filtering a specific population would only show insights relevant to that population.

Visualizing Samples

Visualized samples (highlighted in green) has a white outline, unvisualized samples (highlighted in red) does not.

On the right hand size of the top panel there's a "Visualize" Button. This button allows to run a process that would compute all of the visualization connected to the current evaluated model via its mapping and show them. Depending on the current selection, clicking in this button may allow any or all of the following options:

• Visualize Rest - This visualizes all of the currently unvisualized samples.

• Visualize Selected - This visualize only the currently selected samples

• Revisualize All - This deletes all previous visualization of the current run, and computes new visualization only for the currently presented population

Visualized samples can be seen with dots with a white outline.

To create a seamless exprience, the Tesnorleap platform does not automatically compute sample visualization for the entire dataset. Instead it only visualizes the currently viewed subset of the data.

After each model evaluate the platform automatically visualizes the default selected population, but applying filters to the dashlet may select new, previously unvisualized, samples.

Saving an existing model after changing the associated visualizer in the integration script or after creating a new mapping that changed the connected visualizer will cause new visualization in its respective model runs to be computed and shown. To prevent confusion and avoid multiple versions of the visualizers being presented within the population exploration, we recommend using the "Revisualize All" option in such cases

Locating specific samples

Locating specific samples within a view. in the top left - a count of the located samples, on the right - the selection controller. Clicking the icon next to counter would select all of theses samples.

Sometimes, we would like to locate a specific sample to review it within the current view of the population exploration. For that we can use the "locate sample" icon in the right side of the bar (magnifying glass).

After selecting the categorical metadata of the sample to select for, this highlight the samples (yellow outline) and allows a selection of all relevant samples by clicking the icon in the top-left part of the dashlet, next to their count.

Filtering a specific sample is possible by, for example, selecting its sample id.

Selecting and Reviewing samples

Selecting samples to review

The population exploration dashlet support a click-to-select interface & a drag to select interface.

• Clicking a sample opens the sample review panel

  • Ctrl + clicking other samples would iteratively add multiple samples to the review panel

• Holding shift + selecting an area selects a collection of samples to review multiple samples at once.

The sample review panel

The sample review panel sections: (1) sample list, (2) sample visualization grid, (3) visualization controller, (4) grid settings, (5) action bar

The Sample review panel is divided into: (1) - A list of selected samples. The list shows sample ID and an abbreviation of the subset of data the sample was taken from. The top buttons in this section allow to: (a) filter selected samples and (b) select all samples (2) - Sample visualization grid. This shows the difference visualizations connected to the model by mapping. Each sample has its sample ID listed above it and an abbreviation of the set it belongs to (train, validation, test)

(3) - A dropdown selection that controls the selected visualizer. (4) - Grid Options - with the option to overlay important metadata on each visualized samples, change grid layout, or move from a shared controller to an image-specific controller.

(5) - Action bar, that allows to Fetch Similar samples or run Sample analysis.

Metadata Tags

Adding a metadata tag in sample visualization. In this example, a high loss (right image) and a low loss (left image) is selected from population exploration. Overlaying them with GT label shows a clear case of mislabelling for the right image, labelled 6.

The metadata tags option allows to quickly overlay important sample property over the selected visualization. This allow the sample review process to be efficient, with all the relevant information needed for the analysis of a specific sample - presented in one dashlet.

Changing Grid size and visualizer resolution

Resizing the review panel and changing grid size

Visualizer resolution would be set by the size of the sample review window and the selected grid. This is fully customizable and can be adjusted according to need.

Global contoller vs. controller per sample

Changing the settings from a shared controller (effects all samples together) to a sample-by-sample contoller (granular control)

Some of the visualization have a controller that allows to dynamically set different properties of the presented objects. In the object detection controller, for example, we can control to present only the bounding boxes of a specific class, or all bounding boxes over a specific confidence.

These attributes could be controlled in a global controller that controls the settings of all the reviewed images together, or by a per-sample controller that allows a more granular control on a sample-by-sample basis.

The default controller is a global one. to move to a sample-by-sample controller, click the settings icon on the top bar of the visualizer section.

To export the current selection into .csv you can choose the metadata visualizer and export the data into .csv

The visualization selector dropdown would show all visualizers connected by mapping. Visualizers that are connected to the input would also showcase an attention map in the form of a heatmap, showcasing the common features shared by the concept association of each sample. To further understand available visualizers and their respective heatmaps - refer to their code integration implementation or how they look in the platform.