Is automated feature engineering a path to general AI?

Question

I recently came across the featuretools package, which facilitates automated feature engineering. Here's an explanation of the package:

https://towardsdatascience.com/automated-feature-engineering-in-python-99baf11cc219

Automated feature engineering aims to help the data scientist by automatically creating many candidate features out of a dataset from which the best can be selected and used for training.

I only have limited experience with ML/AI techniques, but general AI is something that I'd been thinking about for a while before exploring existing ML techniques. One idea that kept popping up was the idea of analyzing not just raw data for patterns but derivatives of data, not unlike what featuretools can do. Here's an example:

It's not especially difficult to see the figure above as two squares, one that is entirely green and one with a blue/green horizontal gradient. This is true despite the fact that the gradient square is not any one color and its edge is the same color as the green square (i.e., there is no hard boundary).

However, let's say that we calculate the difference between each pixel and the pixel to its immediate left. Ignoring for a moment that RGB is 3 separate values, let's call the difference between each pixel column in the gradient square X. The original figure is then transformed into this, essentially two homogenous blocks of values. We could take it one step further to identify a hard boundary (applying a similar left-to-right transformation again) between the two squares.

Once a transformation is performed, there should be some way to assess the significance of the transformation output. This is a simple and clean example where there are two blocks of homogenous values (i.e., the output is clearly not random). If it's true that our minds use any kind of similar transformation process, the number of transformations that we perform would likely be practically countless, even in brief instances of perception.

Ultimately, this transformation process could facilitate finding the existence of order in data. Within this framework, perhaps "intelligence" could be defined simply as the ability to detect order, which could require applying many transformations in a row, a wide variety of types of transformations, an ability to apply transformations with a high probability of finding something significant, an ability to assess significance, etc.

Just curious if anyone has thoughts on this, if there are similar ideas out there beyond simple automated feature engineering, etc.

Answer 1

Automated feature engineering, if it is part of any aproach towards general intelligence, cannot be the whole solution. The search for features that are meaningful, as opposed to those that simply exist with no utility, needs some guidance.

In machine learning, feature engineering is typically a search for features that improve performance at a specific task, such as classification or regression. The "intelligence" is partially in the manner of searching, and in setting the goals in the first place. Automated feature engineering typically uses fairly crude search algorithms to look for good features, such as random combination, and follows up with raw processing power in order to cover large numbers of options. Automated feature engineering also does not set the goals for any task, or feed back active behaviour from the output of ML to direct the search. An intelligent agent might actively search for data in its environment in order to test an idea (e.g. repeat an action in order to discover whether the same thing happens, or move in order to observe an interesting event better). Feature engineering is a separate issue to this.

There are some theories of general intelligence that are focused on smart pattern matching as a key component. For instance, Jürgen Schmidhuber has long been a proponent of the reward system for intelligence being compression of observations and predictions. In such a system, better pattern matching discovered by an agent is an intrinsic reward signal as it allows for better compression of world models used by the agent.

Marcus Hutter is another well-known AI researcher, who has proposed a framework called AIXI which incorporates similar ideas. An agent operating using AIXI will benefit from discovering features in its observations that improve its own predictions of what will happen next. Some form of automated feature engineering could well be core to such an agent.