In theory, neural networks can learn any computable problem. All they need is the right form, a comprehensive dataset for learning, and the correct parameters for the training algorithm. The mathematics of these simulated neurons is surprisingly simple, and the training algorithm is also based on an intuitive idea.
00:00 The template: biological neurons
00:21 Simplified: simulated neurons
01:20 The activation function: simulated firing
01:50 Complete formula: the mathematics is done
02:50 Pina's example: one neuron, two parameters
04:40 The loss function: automatically measuring learning success
05:44 Gradient descent: The learning algorithm
06:25 Backpropagation: The key innovation
07:52 Try it out in the browser: TensorFlow Playground
08:24 Try it out in the browser: Hart und Trocken
Additional Resources:
- Pina's c't article with the mini-example with only one neuron: www.heise.de/select/ct/2019/2...
- TensorFlow Playground: playground.tensorflow.org
- Hart und Trocken - neural network: www.hartundtrocken.de/my-prod...
This video is part of a series on artificial intelligence (AI) by c't magazine.
The video is presented by Andrea Trinkwalder and Pina Merkert.
You can find more information about c't magazine on the following websites:
► c’t Magazin: ct.de
► c’t auf Mastodon: social.heise.de/@ct_Magazin
► c’t auf Instagram: / ct_magazin
► c’t auf Facebook: / ctmagazin
► c’t auf Bluesky: bsky.app/profile/heisec.bsky....
► c’t auf Threads: www.threads.net/@ct_magazin
► c’t auf X/Twitter: / ctmagazin
► c’t auf Papier: überall wo es Zeitschriften gibt!
#AI #neuralnetworks #maths