## [[Transformational Abstraction]] #thought - from [[From Deep Learning to Rational Machines]] - Abstraction as a power connoisseurs possess - relevant to the [[The Environmental Aesthetics of Generative AI]] #### Key Aspects of [[Transformational Abstraction]]: ## [[Transformational Abstraction]] #thought - from [[From Deep Learning to Rational Machines]] - Abstraction as a power connoisseurs possess - relevant to the [[The Environmental Aesthetics of Generative AI]] #### Key Aspects of [[Transformational Abstraction]]: - **Systematic Transformations**: The core idea is that by applying systematic transformations to input data, one can identify invariant properties that lead to abstract concepts. - **Handling Nuisance Variables**: Nuisance variables are irrelevant factors that cause variations in sensory input (e.g., lighting, orientation). [[Transformational abstraction]] helps in controlling for these variables to focus on essential features. - **Feature Space and Manifolds**: In high-dimensional feature spaces, data points form structures called manifolds. Transformations help in "untangling" these manifolds, making different categories more separable. - **Unifying Traditional Approaches**: [[Transformational abstraction]] integrates elements from the four traditional methods: - **Abstraction-as-Subtraction**: By filtering out irrelevant details through transformations. - **Abstraction-as-Composition**: By combining simpler features into more complex ones during [[the transformation]] process. - **Abstraction-as-Representation**: By using transformed instances as representatives of general concepts. - **Abstraction-as-Invariance**: By identifying properties that remain constant across transformations.