Does "fusion" in "feature fusion" has any formal definition?

Question

I encountered the phrase "fusing features" several times in the literature. I am providing an excerpt from a research paper to provide context for usage of the word fusion.

The reason is that the signals measured by multiple sensors are disordered and correlated with multiple sources. Those methods that are proposed with an attempt to use multiple data sources are called data fusion techniques. Upon the position where the fusion operation is conducted, there are three general approaches: signal-level fusion, feature-level fusion, and decision-level fusion.

I am guessing that "fusing features" refers to an act of combining several features, from different domains, and then generating new features that serves the purpose of fusing.

If yes, the word "fusion" here refers to its common English usage

The process or result of joining two or more things together to form a single entity.

That is,we need to combine multiple features in any manner and then coming up with new features that are good enough to perform our AI task.

Or does it have any formal definition and requirements based on the input or output features? Is there any formal definition for fusion operator?

Answer 1

With this link I could read the paper. Thanks.

So there is this discipline called sensor fusion. It is very sounded in the field of Autonomous Vehicles where in order to take one decision (whether to break or not) you have to take into account information for multiple sources: car mounted cameras, LIDAR, ultrasound, radar...

So the term "fusion" refers to the operation of aggregating the information from multiple sources (that has its problems as the paper says: the signals measured by multiple sensors are disordered and correlated with multiple sources). In order to perform this fusion or aggregation you can aggregate the information in different levels of the processing pipeline: close to raw data (signal fusion), close to high level information (decision fusion) or something in the middle (feature fusion).

Normally when you have a signal (image, radar, electromagenic...) you proces it somehow (normally using filters). The output of those filters is a feature map (in case of 2D images) or feature vectors (in case of 1D signals). Usually when you have those signals processed you use a decision module to extract high level information (a classification head, a SVM, a regressor...).

Basically you can aggregate information at those 3 levels. The authors refer as feature fusion as to aggregate information in the middle step. You do not have raw data, but you do not have refined data either. They do so expecting they remove the noisy part of the signals (filters) or the non relevant parts (PCA) but without removing the nuances lost when using the decision modules.

The name "feature fusion" comes from the deep learning terminology in which when you have a signal and you process it somehow the output of it is a feature. When reading a paper on image detection / classification you would see "feature maps" but when reading a paper on sensor or audio you would read "feature vector" hence then name