120,711 views
In this video we take a detailed look at several of the activation functions currently used in neural networks. We review how they work, why they are needed, and compare some of their features. I already have courses! Learn to program from scratch with Python: https://www.domestika.org/es/courses/... Learn AI from scratch with Python: https://www.domestika.org/es/courses/... If you want to support me and be part of this project, you can do so in several ways: Patreon: http://bit.ly/patreon-ringatech Youtube Membership: / @ringatech If you liked the video, click like and leave me a comment! = = = CONTENT 0:00 What are they for? 4:18 Step 4:55 Derivatives 6:48 Logistics / Sigmoid 7:52 Gradient Vanishing 9:07 Hyperbolic Tangent and Softsign 11:00 ReLU 12:45 Dead Neurons 13:35 Leaky ReLU 14:00 PReLU and GELU 14:44 Softplus, Maxout, ELU 15:43 Swish / SiLU 16:29 Mish 17:15 Identity and Softmax 19:47 When to use each one 21:24 Oscillatory? = = = Long description for the algorithm: Activation functions are an indispensable component for neural networks to be able to generate learning in order to make predictions. In this video we see functions like the logistic/sigmoid function, the hyperbolic tangent/tanh, Softsign, ReLU, Leaky ReLU, Parametric ReLU, GELU, SILU, ELU, Softplus, Maxout, Swish, Mish, GCU, Softmax and the identity function. Many of the functions allow the backpropagation process to work more efficiently, so we look for activation functions with various properties such as being differentiable, having a low computational cost, and in some cases being centered on zero, or being non-monotonic. In the video we see in an intuitive way how and why they work, with visual aids and review of publications. #machinelearning #neuralnetworks #artificialintelligence