AI Miscellaneous (1) What models do you like in Ml? What is the future direction of ML development?

Sender: echostate (AI), email area: AI
Question: Which models do you like in Ml? What is the future direction of ML development?
Mail station: Shui muCommunity(Sun JUN 17 01:27:00 2007), Station

I have been idle for a while recently. If you are interested, come and have a chat.
I am only a beginner, and there must be many errors. You are welcome to correct and discuss it.
ML = machine learning in this article, no unreasonable associations ~~~~~

Machine Learning is a cross-discipline that is closely related to many disciplines, such:
-- Statistics (a synonym for ML is statistical learning)
-- AI (Ml is considered a branch of AI, although different from some traditional AI fields)
-- Information Theory (many models use the theoretical basis of information theory, such as entropy and MDL)
-- Brain and Cognitive Sciences (a typical example is that neural networks originated from the brain)
-- Psychology/Social Science (Ml is often inspired by human learning/animla learning, such as ion clusters/ant colony/other traditional evolutionary computing)

Therefore, ML research models/directions are varied, and their respective foundations are quite different.
So what are the most interesting directions in Ml? What do you want ml to do in the future?

Is it a graph model? Will ML be more closely integrated with statistics in the future (after all, the decision-making problem is ultimately the highest posterior probability. ML can't escape this box for supervised learning or unsupervised learning )? In today's ML world, the "Graph Model" flow maintained by Jordan, Kollar, laffery, and others can be described as a huge hill of ML. These people are quite statistical (many are statistical) models developed in the graph model scope, such as the classic Bayesian Networks and Markov Random Field, various directed/undirected graph models designed for different applications are filled with annual nips, icml, jmlr, and JML. But I often wonder if the ml that relies entirely on statistical reasoning is the ml in our hearts? Or is it consistent with the original intention of ML? Is this ml the most attractive?

Another type of model/direction is not from the perspective of pure statistical reasoning, but from the perspective of High-dimen1_space/geometry. Of course, these models will eventually be put under the statistical framework to prove their effectiveness, but their intuition is not statistics, but some ideas based on geometry/high-dimen1_space. For example, K-nn/SVM (kernal-base methods, Max margin methods)/dimension regression ction (PCA/MDS/Manifold Learning)/graph-based semi-supervised learning. I personally think these methods are more intuitive, more attractive, and more intuitive and beautiful. If it is theoretically complete (such as SVM), it can become a classic.

Some directions are based on cognitive science, such as the famous neural networks. The idea of NN's initial understanding is very attractive, so now we have encountered a bottleneck. The current research is also deadlocked. For example, Artificial NN has abandoned its biological significance. It is sometimes puzzling to simply design network architecture as a computation tool.

Some directions are based on information theory. Demo-tree is an example. You can also explain it from other perspectives.

Some directions are based on social science. For example, ion Groups/ant colony/other traditional evolutionary computing, studying the relationship between individual micro-behavior rules and overall behavior results-of course these are strictly not part of ML, but optimization science.

Some directions are based on human learning. An extremely hot direction is ensemble learning. The famous ones such as Bagging and boosting are mostly studying how to train multiple classifiers and how to make them collaborate to produce better results. I personally think that ensemble learning is intuitive and interesting, but it is currently limited to intuitive heuristic research and has not yet developed a theoretical foundation.

There are many more, and I don't know or list them myself. In short, there is a big difference between different directions in ml. The main difference is that each direction has its own intuition source. What is the most interesting direction in Ml? What do you want ml to do in the future?