Embedded Files
The Vanishing
Unstill Life with the Vanishing Game
Unstill Life with the Vanishing Game
Inspired by Juan Sánchez Cotán’s Still Life with Game, Vegetables and Fruit, this series explores how patterns can disappear when data is not treated with care. It reflects on the process of normalization of the training and unseen data, where using a different reference can distort the original structure. Like the absent game in this composition, meaningful insights may vanish when scaling breaks the link to what once held them in view.
Original image source: Museo Nacional del Prado
Original Observations
Original Observations
Observations with variables such as pixel coordinates and color value are given for different layers of the above painting. A machine learning algorithm is trained on a portion of the observations such as the background layer and the objects except the Game which is designated as unseen data.
Properly Scaled Data
Properly Scaled Data
Upon training the model, statistics from the training set are extracted and are used to scale variables of both the training and the unseen set. With the correct approach, the unseen data is scaled appropriately with respect to the learned information.
Improperly Scaled Data
Improperly Scaled Data
When the unseen dataset is scaled with respect to its own statistics and not to the training ones, the original relative structure gets distorted and patterns such as an object on the canvas—like the Game—vanish.
Brief Paper - The Vanishing Effect of Inconsistently Scaled Data.pdfBrief Documentation of the Experiments
Brief Documentation of the Experiments
Two brief experiments related to the painting are documented, titled as The Vanishing Effect of Inconsistently Scaled Data.
Page updated
Google Sites
Report abuse