A few drawbacks to neural networks
While neural networks are an excellent step forward in automating the document classification process, there are a number of drawbacks that need to be considered. To begin, developing the classification hierarchy is expensive. It requires powerful computers to process the data.

Neural networks build a classification hierarchy from a representative sample of existing domain data. Therefore, if data behavior changes, the change won't be noticed. Instead, it'll be hidden by being placed in the closest-fit classification node, much like a structured data silo or stove piping.

Another issue to consider is that neural networks deliver only one taxonomic view of the data they're classifying. No, "taxonomy" and "taxidermy" aren't the same thing. Throughout this article, I'm going to use the words "taxonomy" and "taxonomic". Although they're certainly not the easiest words to get your arms around, they're much more precise when talking about classification systems than the term "category". But, if it's easier to handle, whenever you see the word "taxonomy", substitute "category" in your head.

In today's complex global marketplace, there's a need to have multiple classification hierarchies applied against the same domain of data, each from the perspective of the interest group, individual, or country.

While neural networks deliver good classification hierarchies, they deliver poor information navigation and discovery environments. When you select a taxonomy node of interest, you may need to find a document collection in the thousands, hundreds of thousands, or millions of documents that are represented in that node.

To get to a document collection that can answer your question, you'll need to apply another discovery method, such as full text retrieval. This can break your concentration when in the discovery mode, often resulting in you losing your train of thought. At this point, the idea you were discovering can dissipate into frustration. Many of us have certainly expressed this frustration using full text retrieval for discovery, when faced with large hit lists of terms that are out-of-context and therefore irrelevant.

Adaptive processing-based classification systems
Where the neural network simulates how the brain classifies information, adaptive clustering technology emulates the way in which the brain classifies information, delivering all the inherent benefits. Unlike neural networks, the brain must sample data as it arrives and do its best to compartmentalize that data, reshuffling or associating it in multiple dimensions as it becomes more abundant. Through successive simplification, the brain divides data into multiple dimensions of simpler representations as it reaches critical mass and the dividing boundaries become clearer.