State of the art

State of the art

Sound classification

Hierarchical classification

Neural networks have been used for solving a multitude of classification problems. The hierarchical classification (HC) makes reference to the classification problems that can be organized in a hierarchy. This can be represented with a tree, where the leafs are the labels to classify.

Typically, there are three main different approaches for solving HC: flat classifications, global classifiers or local classifiers.

Flat classification

In this approach, the hierarchical information is discarded, and a normal multi-label classification is used over the leafs of the tree. Its main advantage is its simplicity, and easy implementation. On the other hand, its main disadvantage is the disregard of the hierarchical information, that could allow a better more precise.

Global classifiers

This term reference those classifiers that use one single network to make the classification, where the hierarchical information is incorporated. There are many ways to create these classifiers, some of those are INSERT REFERENCES. The main advantage of this method is that, as only one network is used, it can be trained with all the data, and may offer good results. The disadvantage of these methods are that they are generally complex, require fine-tuning and need to be hand tailored for each problem.

Local classifiers

Local classifiers use a multitude of multi-label classifiers. To traverse the tree, the root classifier is called, and then, depending on the output, other results are called. Among this category, there are three different models: Local classifier per node (LCN), local classifier per parent node (LCPN) and local classifier per level (LCL).

Local classifier per node

Each node in the tree has its own classifier, even the leaves nodes.

Local classifier per parent node

Similar to the former, but in this case, the leaves do not have a classifier.

Local classifier per level

In this case, there is one classifier per level.