Stop Doing Deep Learning

STOP DOING DEEP LEARNING • PERCEPTRONS WERE ONLY EVER MEANT TO BE FULLY CONNECTED • THOUSANDS OF PAPERS yet NO REAL-WORLD USE FOUND for going deeper than ONE LAYER • Wanted to add more nonlinearity anyway for a laugh? We had a tool for that: It was called “KERNEL METHODS” • “Yes please give me a network that can PAY ATTENTION TO ITSELF. Please give me PRETRAINED WEIGHTS for my YOLO-9000” - Statements dreamed up by the utterly Deranged LOOK at what “Research Scientists” have been demanding your Respect for all this time, with the statistical methods & optimization algorithms we built for them (This is REAL "Deep Learning", done by REAL “ML Engineers”): ????? ???????????????????? “Hello I would like to learn 1.6 TRILLION parameters please” They have played us for absolute fools

Текст высмеивает современное глубокое обучение, противопоставляя простые перцептроны и классические методы огромным нейросетям с триллионами параметров. Шутка в саркастическом представлении deep learning как бессмысленного усложнения: исследователи якобы требуют всё более глубокие сети, attention и pretrained weights, хотя инженерные методы уже решали практические задачи. Мем используют как реакцию на критику нейросетевого хайпа и чрезмерную сложность современных моделей.

• @profunctor_io / 6350 2021-02-25
• @right_wing_dev_squads / 10465 2024-09-12