The papers of the participants are available in the proceedings of ICGI'23: the one of the winning team (EdiMuskardin), the one of the team that finished second (NeuralChecker), and the one of the Redouane Hakimi team.
This paper, that will appear in the proceedings of the 16th International Conference on Grammatical Inference, describes the competition. It provides info on the data used, the trained models, etc. Following our Creatives Commons Licence, please cite it if you are using the TAYSIR models in your work.
The trained Neural Nets and the datasets of the competition are now available in this archive. We also added several transformers we trained on our datasets but did not used for the competition. We hope that this TAYSIR Benchmark will be used from now on to compare approached on distillation of Neural Networks trained on sequantial symbolic data
The Competition is over! The team from Edi Muskardin won the competition. Congrats! Thanks everyone for your participation.
|March 22nd: |
|March 6th: |
The competition is in beta-testing: you will find the link to access it in our Manual
We created a place on discord for discussions about the competition. Feel free to join us there
We faced some difficulties (tackled!) and we are running late on our schedule. We plan to open the competition on February 22nd.
Our first call for participation is now available. Check it out!!
Transformers+RNN: Algorithms to Yield Simple and Interpretable Representations (TAYSIR, the arab word for 'simple') competition was an on-line challenge on extracting simpler models from already trained neural networks. These neural nets are trained on sequential categorial (=symbolic) data. Some of these data are artificial, some come from real world problems (NLP, Bioinformatics, Software Engineering, etc.)
The quality of the extracted models are evaluated following two directions:
► Approximation of the original model
► Simplicity of the extracted model
Two tracks are proposed:
► Neural nets trained for binary classification, that is, a language in the Formal Language Theory sense
► Neural nets trained for regression, that is, a function assigning a real number to any finite sequence of symbols (e.g the density estimation of Language modelling RNN) Each track consists of roughly 15 trained models.
The trained models are in pytorch but available in a MLFlow format for compatibility with other frameworks.
The Competition took place during Spring 2023, between mid-February and April.
Half a day will be dedicated to the competition results during the 16th International Conference on Grammatical Inference to be held in Morocco in July 2023.
Participants in TAYSIR will be encouraged to attend ICGI 2023 and to submit an extended abstract presenting their work (2 to 4 pages, including appendices) which will be appended to the proceedings of ICGI (Publisher: PMLR) in a track dedicated to the competition. The deadline for this abstract is May 15th, 2 weeks after the end of the competition - note that it is different than the ICGI paper submission deadline which is at the beginning of March. These abstracts will be peer-reviewed primarily for clarity of presentation.
|February 2023:||Beginning of the competition|
|April 30th 2023:||End of the competition|
|May 15th 2023:||Submission deadline for the extended abstracts presenting your work|
|July 10th-13th 2023:||ICGI 2023, including a dedicated session about TAYSIR with presentations from some participants|
Everything, including trained Neural Nets, datasets, script for evaluation, manual, scientific article, and bonus Transformer models can be found here.
The first Track (Binary Classification) was at here. The second Track (Language Modeling/Density Estimation) was at here. There you needed to register in order to download the already trained model. The manual, where main elements are described, is a good start.
After running your extraction algorithm, we will expect participants to upload on our website their extracted model as an archive containing a MLFlow pyfunc. We provide a tool kit that can save any python function in the desirated format. The detail of its (simple) usage can be found in the starter kit given with the models on the submission website.
We used 2 types of metrics: one to evaluate the quality of the approximation, the other to evaluate the simplicity of the submitted model.
For Track 1 (Neural Nets trained on binary classification tasks), the quality is avaluated using the accuracy, that is, the proportion on the sequences of the test set (unkown to participants) on which your model and the neural net agree.
For Track 2 (Neural Nets trained on language modelling tasks), the quality is avaluated using the mean square error. Models are seen as density estimators: they can assign a probability to each sequence of the test set: we compute the square of the difference between your model and the original neural net on each sequence and look at the mean over the whole test set.
We use two metrics to evaluate the simplicity of your submitted models: Memory usage: The memory footprint of your model during a prediction, in megabytes. CPU time: The time spent in CPU of your model during a prediction, in milliseconds.
Notice that when you first submit to the competition, we are creating a virtual environement
Hosei University, Japan
Université Jean Monnet, Saint-Etienne, France
Stony Brook University, New York, USA
Université Jean Monnet, Saint-Etienne, France
DiLySS - IRISA
MLDM Mater program
MLDM Master program
Universitat Politècnica de Catalunya
Yale University, USA