Causality in Machine Learning

回到主页

1. issuses with machine learning ---- generalization

Current machine learning approaches tend to overfit the data. They try to learn the past perfectly, instead of learning the causl relationships in the data, which is actually the most important part. Means they can't draw inferences about other cases from one instance. correlation is not causation

Deep Learning has focused too much on the correlation without causation.

2. Solutions

- 1. Causal Bayesian Network

This method estimates the relationships between all variables in a data set and can be considered as a true discovery method. It enables the discovery of multiple causal relationships at the same time.

Basically, it results in an intuitive visual map showing which variables influence each other, as well as the extent of their influence. Indeed, Causal graphic models make it possible to simulate many possible interventions simultaneously.

This approach allows for the incorporation of expert knowledge to counter the possible limitations of a purely data-driven approach. expert can help:

  • Place conditions on the model to improve its accuracy,
  • Determine which variables should go into the model
  • Help understand counterintuitive results.

(ref https://towardsdatascience.com/introduction-to-causality-in-machine-learning-4cee9467f06f)

- 2. Structural Causal Models (SCM)

Developed by Judea Pearl which enables us to infer causality via Directed Acyclic Graphs (DAG), which is also the basic of Causal Bayesian Network.

- 3. Rubin Causal Model (RCM) or potential outcomes

If we want to no the potenial outcomes Y, we can do exeriments on different level of X.

(ref https://towardsdatascience.com/causality-in-machine-learning-101-for-dummies-like-me-f7f161e7383e)

- 4. About Reinforcement Learning

Reinforcement Learning is not actually a kind of causality learning, however it's better than the most deep learning. Reinforement learning can learn from the rewards, therefore the only causal is between the input and the rewards. Of coures, it's good at competitive games, because the only thing the competitive cares about is win. Reinforement learning could only know how to win a game, but he don't know why he can win the game. So, RL is not this kind of causal machine learning even he do much better than CNN or RNN.

For example, CNN or RNN is a student who has been done the answers of a lot of questions, therefore, they can figure out the very similar questions they have been done. But if the question change a little, they don't know how to solve it. RL first don't know the answers of these questions, however, he knows if he figure out one question, he can achieve high score. So, RL know how to figure out the questions, but just like the Test-oriented education, he don't understand the knowledge. He maybe know how to solve the problem using F=ma, but he don't know what is F=ma.

Perhaps, causal machine learning can do better.

本文章使用limfx的vsocde插件快速发布