The most common mistake people do is trying to learn some complex cutting-edge technology from scratch. Cutting-edge technology is just combining old technology in new ways to solve new problems. To learn it, first learn existing technology. Existing technology is here to stay, it's well-explained, and it's usually much simpler to understand. Then the rest will just click.

This is exactly my js experience.

Hmm… not sure if it also applies to ML/AI, blockchain, cryptography or computer graphics.

@Lensflare Sure it does. For example, basic deep neural networks are just a bunch of perceptrons glued together with non-linear functions like softmax, softplus, logit, etc. A single perceptron with logistic function is essentially logistic regression, which is what the most basic classification models used for decades. And logistic regression is just an algorithm to find some vector w, such that logit(w*data) = prediction, which is just you basic linear algebra.

Going in the other direction, convolution neural networks used in state-of-the-art image recognition simply replace the aforementioned logit(w*data) operation by convolution and connect them into more complex topologies. NeuroEvolution on the other hand just utilises the age-old idea of genetic algorithms and modern GPU power to automatically build and test hundreds of neural network topologies.

@Lensflare Git is essentially a relaxed blockchain protocol which allows for merging forks (because conflicts can be manually resolved without damage), and doesn't require mandatory signatures or consensus algorithm. But again, most blockchains are built from some existing consensus algorithm, some existing cryptographic signature, and some "proof" operation which is hard to scale, like computing a hash of random data.

@hitko ok but in order to not make this mistake (learning from scratch), how is one supposed to know about all of this old basic tech?

@Lensflare Instead of searching for "deep neural image classification", search for "simple ml classification"; in other words, learn to ask basic questions first.

Most Wikipedia articles also mention what technologies something builds on, read that first unless you're already familiar, it'll make understanding the main technology much easier.

Alternatively, and I know that's a really unpopular opinion around here, stay in school. A machine learning class will teach you everything from the most basic classification to implementing your own neural networks from scratch in a single semester, and then you can build on that.

@hitko I've yet to see a network that incorporates subnets subject to neurevolution.

You always see either neuroevolution as a pre or post step, or else networks that dont use any sort of GA at all.

Looe, the human brain is changing connection topology *regularly*, why shouldn't AI be able to change and adapt portions itself on the fly?