Visualization for Machine Learning
## Color  Vasas et al, PLOS Biology, 2024 ::: footer Image from [link](https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002444) ::: ## Visible Spectrum  ## Light * Visible range: 390-700nm * Luminance has a large dynamic range: – 0.00003 -- Moonless overcast night sky – 30 -- Sky on overcast day – 3000 -- Sky on clear day – 16,000 -- Snowy ground in full sunlight * Colors result from spectral curves – dominant wavelength, hue – brightness, lightness – purity, saturation ## Physiology of the Eye  ## The Retina  ## Photoreceptors * Discrete sensors that measure energy – Adaptation * Rods - active at low light levels (scotopic vision) - only one wavelength-sensitivity function * Cones - active at normal light levels (photoptic) - three types: sensitivity functions with different peaks ## Cone Sensitivity  ::: footer Image from [link](https://wrfranklin.org/Teaching/graphics-f2019/files/stone_colors.pdf) ::: ## Density of Cones  ## Cones and Rods  ## Color Stimulus  ## Color Matching Experiments  ## Color in Visualization * Trichromacy: Humans perceive colors according to three channels * Most usable and useful way to describe colors (especially for visualization): - Hue, Saturation, Luminance ::: footer Material from Enrico Bertini ::: ## How do we use color in visualization? * Quantify * Label ::: footer Material from Enrico Bertini ::: ## Quantify  ::: footer Material from Enrico Bertini ::: ## Label  ::: footer Material from Enrico Bertini ::: ## Quantitative Color Scales * Desired Properties of Quantitative Color Scale: - Uniformity (value difference = perceived difference) - Discriminability (as many distinct values as possible) ::: footer Material from Enrico Bertini ::: ## Quantitative Color Scales * Desired Properties of Quantitative Color Scale: - Uniformity (value difference = perceived difference) - Discriminability (as many distinct values as possible) ::: footer Material from Enrico Bertini ::: ## Single Hue Sequential Scales * Choose one hue * Map value to luminance  ::: footer Material from Enrico Bertini ::: ## Categorical Color Scales * Properties: - Uniformity (uniform saliency / nothing stands out) - Discriminability (as many distinct values as possible)  ::: footer Material from Enrico Bertini ::: ## Categorical Color Scales * How many distinct values can one perceive? * how many can you use in a visualization? * Estimates are between 5-10 distinct codes. * Healey, Christopher G. "Choosing effective colours for data visualization." Proceedings of IEEE Visualization'96, 1996. ::: footer Material from Enrico Bertini ::: ## Diverging Color Scales * Sometime useful/necessary to distinguish values above and below a threshold. ::: columns ::: {.column width="50%"}  ::: ::: {.column width="50%"}  ::: ::: Created using these data: [link](https://github.com/tonmcg/County_Level_Election_Results_12-16) ::: footer Material from Enrico Bertini ::: ## Color Blindness * Missing or defective photoreceptors: - 10% male and 1% female have some color deficiencies  Oliveira, Manuel. "Towards More Accessible Visualizations for Color-Vision-Deficient Individuals." Comput. Sci. Eng., 2013. ::: footer Material from Enrico Bertini :::
