BryoBrain – Introduction

The computer program „BryoBrain“ is a “neural classifier”, a program that allows the determination of all 73 European (incl. Macaronesian) foliose liverwort genera (Jungermanniales and Haplomitriales) with the help of “artificial neural networks“ (ANN) using different morphological and biometric data. The program gives an impression of how the determination of plants will work in the future and which methods will replace the dichotomous keys of the “book age”. It works in a similar way as many programs for OCR, speech-, fingerprint- or face-recognition, stock market prediction, industrial quality control and many other things. Neural networks permit the use of chemical or molecular data and will be very important tools and show their capability when these data will be generally available for identification or with image recognition of species.

The ANN used in BryoBrain are multilayer perceptrons (MLP) trained with backpropagation (for explanation see the “Introductions to ANN” on the link-page).

The objectives for the development of BryoBrain were first to demonstrate that ANN are capable of identifying bryophytes, and second to test how many and which characters are required for their identification. For a few other species groups classification with ANN was already tested. However, to my knowledge BryoBrain is the world’s first independently executable and distributable program for the determination of plants or animals.

Neural networks are capable of learning (and can thus be considered „artificial intelligence“), although the distributed version of BryoBrain contains networks that have already been trained to recognize liverwort genera and the learning capability has been omitted.

Compared to database solutions, ANN have several advantages, especially:

• ANN are to a certain degree error-tolerant. If a character is slightly outside of its normal range within the genus, the ANN will in most cases find the correct result, while the database solution would fail.

•   Even if one character is entered completely wrong, the ANN in many cases gives the correct result or shows the correct genus  in the list of genera with lower weights

• Unknown species of a known genus are in most cases classified correctly.

• ANN are capable of detecting relationships between e.

One disadvantage is that the development of an ANN solution different inputs while database solutions simply compare the inputs with the values in the databasis more time-consuming and the ANN have to re-trained any time a change is made to the training data or network topology.

Compared to the last version (0,6) the following features have been improved or added (see also history):

• all 73 European genera are recognized

• reliability of determination is significantly improved

• quality and quantity of the training data and test data were significantly improved

• neural nets partly were improved by genetic algorithms

• new databases can now be created

• all records in a database can be classified with one mouse click

• for each genus a list of European species is shown

• for some genera illustrations are shown (will be extended)

• some characters are explained by illustrations (will be extended)

• Internet picture search now integrated

• update check via internet

In version 0.7 the recognition rate could be increased, this despite the significantly higher number of genera. This is particularly due to the better training data, which have been improved in quantity and quality, as well to improved networks.

The current version (0,7) is still a beta version and possesses some shortcomings. BryoBrain currently classifies >99% of the testing data correctly. In spite of the high recognition rate you should not expect the program to classify all specimens correctly, as the overall variability of many genera is not yet included in the training and testing data.

At present BryoBrain has some problems with genera possessing small, variable underleaves. This issue will require a better definition of the characters in the training data as well as some changes in the network topology in the next version.

The development of BryoBrain will be continued. I’m currently testing if all North American genera can be included. In addition, neural nets for other areas (South America) and other species groups are planned. With every change the nets have to be trained again, which is time-consuming. Therefore, the development of the program will continue slowly.

BryoBrain uses in the current release 21-31 characters for the recognition of genera, which must unfortunately still be entered by hand. The data must be entered completely in order to avoid  mis-classifications, although in some cases the ANN give the correct result already with a small set of characters. All characters are preset to the most common conditions. Therefore several characters, like the presence of ocelli or a vitta, in most cases do not have to be changed.

Currently BryoBrain uses two perianth characters. Since perianths are not always available, two different networks are included:

• one which uses up to 31 characters (including the perianth characters); and

• one which uses only up to 29 characters (excluding the perianth characters).

The net that uses up to 31 characters is somewhat more reliable, but the differences are small. In order to use all 31 characters, the box „Use perianth characters“ must be checked. 

Some users asked for a version that requires fewer characters. Therefore, a net was included that uses only 12 to 14 characters (which are recognizable with the hand lens) for classifying. This fast test is somewhat less reliable (recognition rate ~ 94%).

BryoBrain is now completely bilingual (English/German). The language can be changed with a single mouseclick. It includes an English and German introduction and description.

BryoBrain has database functionality and can store the entered data in a database. I have added a database with about 100 records giving the species name in the ‘Observation-field’, which are not known to the ANN. Therefore, you can play around, change the data and test the program and observe the behaviour of the ANN without entering your own data. The feel of such a program is very different from the use of a dichotomous key in a book and one has to get used to it. Of course it is also possible add your own records to the database and feed the ANN with it.

Version 0.7 includes some Internet capabilities. The program can start a picture search with a search engine using the standard web browser and it can check the BryoBrain homepage to see if a newer version is available.

There is a more detailed description of BryoBrain on my homepage (http://www.drehwald.info) where the program can be downloaded free of charge.

Requires MS Windows® 9x, ME, NT, 2000 or XP.