Learning diverse causally emergent representations from time series data

Authors: David McSharry, Christos Kaplanis, Fernando Rosas, Pedro A.M Mediano

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our results show that causal emergence speeds up learning of more complex features of the data relative to pure mutual information maximisation, and we demonstrate the scalability of our method through an analysis of real-world brain activity data.
Researcher Affiliation Collaboration David Mc Sharry Department of Computing Imperial College London dm2223@ic.ac.uk Christos Kaplanis Google DeepMind kaplanis@google.com Fernando E. Rosas Sussex AI, University of Sussex f.rosas@sussex.ac.uk Pedro A.M. Mediano Department of Computing Imperial College London p.mediano@imperial.ac.uk
Pseudocode Yes For convenience, pseudocode for the algorithm used to train the model is provided in Supp. Sec. A.
Open Source Code Yes Code implementing our proposed architecture and reproducing our key results is available at https: //github.com/Imperial-MIND-lab/causally-emergent-representations
Open Datasets Yes Primate ECoG dataset. We evaluate our method on a dataset of electrocorticography (ECoG) brain activity data from a macaque monkey, originally reported by Chao et al. [10].
Dataset Splits No Glider decoding b) Classification accuracy of the state of the glider (c.f. Supp. Fig. 7) on a held-out test set. Dashed black line represents chance level at 25%.
Hardware Specification Yes All experiments can be run on a single A10G GPU in less than two hours.
Software Dependencies No The paper does not explicitly state specific software dependencies with version numbers.
Experiment Setup Yes Hyperparameters for the models trained on the synthetic and brain datasets can be seen in the below tables. The hyperparameters do not change when learning diverse sets of emergent features. For each training run the clip was fixed at 5 in the SMILE mutual information estimators and the optimizer used was Adam W.