Daniel Burfoot

An implementation of a novel method for data compression called the "mirror" technique. Here are some examples that use the technique.

This technique avoids a crucial difficulty in compression, requiring the encoder and decoder to be perfectly in sync. Using this library, developers can easily create new data compression algorithms with just a few lines of code.

To use this tool, a developer first writes an algorithm or software process using a special set of conventions. Then the user can automatically extract a visual diagram describing the algorithm.

The diagram is very useful for documentation purposes; other developers (or the original developer, at a later point in time) can easily understand the way the code works just by looking at the diagram, without needing to dive into the specific details.

At Ozora Research, I built a broad grammar sentence parser without using labeled training data (almost all other work in the area of parsing depends on labeled "treebank" data).

The parser is built in combination with a specialized text compressor which compresses text by using a parse tree. The parser produces the tree that will produce the smallest code length for the given sentence. You can demo the parser at the link provided.

This is a book that I wrote about a new approach to empirical science based on lossless data compression. In this philosophy, a researcher proposes a theory, builds the theory into a data compressor, and measures the quality of the theory by invoking the compressor on a large shared data set. If the theory achieves a lower net code length (including the size of the compressor itself) than previous theories, it is confirmed as the new "champion" theory.

This philosophy guided my work at Ozora Research. In this case, the relevant data set was English newspaper text. To compress this data, I developed theories of grammar and syntax, and build those theories into a data compressor.

My Ph.D. thesis developed an approach to statistical modeling based on the search for randomness deficiencies in an encoded form of the data.

According to algorithmic information theory, if a given model is a perfect fit for a data set, then when you encode the data using the model, the resulting encoded data (typically a bit string) is completely random. This implies that if you have a model—and encode the data using the model and find a randomness deficiency in the encoded data—then there is a flaw in your model. Furthermore, an analysis of the randomness deficiency illustrates a way to improve the model.

The thesis developed a suite of machine learning algorithms that work by using this idea.