FastConceptMapping

Code and data corresponding to work on fast concept mapping (FCM) based on sparse and meaningful representations learned self-supervised through curiosity and embodiment.

Publication:

https://ieeexplore.ieee.org/document/10274870

Cite as:

@ARTICLE{10274870, author={Clay, Viviane and Pipa, Gordon and Kühnberger, Kai-Uwe and König, Peter}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, title={Development of Few-Shot Learning Capabilities in Artificial Neural Networks When Learning Through Self-Supervised Interaction}, year={2024}, volume={46}, number={1}, pages={209-219}, doi={10.1109/TPAMI.2023.3323040}}

Pre-Print:

https://arxiv.org/abs/2102.02153

Data:

Find data in the ./data folder and the bigger trainings and test data sets here: http://dx.doi.org/10.17632/zdh4d5ws2z.2

Project Structure:

The jupyter notebooks in the ./scripts folder are numbered. The two scripts indexed with 0 are for model training of the autoencoder and the classifier. To train the agent I use ml-agents ppo (https://github.com/vkakerbeck/ml-agents-dev). The trained models can also be found in the data folder as well as the corresponding hidden layer activations on the test set of images. These activations were produced with code from 1_GetNetworkActivations.ipynb. From this you can then calculate the performance of fast concept mapping (FCM) on the repersentations using 2_CalculateFCMStatistics.ipynb. To visualize the statistics and further insights into the encodings use 3_AnalyzeResults.ipynb.

If you have any questions or requests feel free to contact me: [email protected]