Can Transformers Smell Like Humans?

Authors: Farzaneh Taleb, Miguel Vasco, Antonio Ribeiro, Mårten Björkman, Danica Kragic

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

Reproducibility Variable Result LLM Response
Research Type Experimental We use multiple datasets and different types of perceptual representations to show that the representations encoded by transformer models are able to predict: (i) labels associated with odorants provided by experts; (ii) continuous ratings provided by human participants with respect to pre-defined descriptors; and (iii) similarity ratings between odorants provided by human participants. Our experiments do not require significant computational resources: we mostly train linear models that do not involve GPU usage or models that can be trained on a single commercially available GPU under one hour.
Researcher Affiliation Academia Farzaneh Taleb Department of Intelligent Systems KTH Royal Institute of Technology Stockholm, Sweden farzantn@kth.se Miguel Vasco Department of Intelligent Systems KTH Royal Institute of Technology Stockholm, Sweden miguelsv@kth.se Antônio H. Ribeiro Department of Information Technology Uppsala University Uppsala, Sweden antonio.horta.ribeiro@it.uu.se Mårten Björkman Department of Intelligent Systems KTH Royal Institute of Technology Stockholm, Sweden celle@kth.se Danica Kragic Department of Intelligent Systems KTH Royal Institute of Technology Stockholm, Sweden dani@kth.se
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes All computational code to reproduce our results is available at https://github.com/Farzaneh-Taleb/transformer-olfactory-alignment
Open Datasets Yes We use the publicly available version of the following datasets provided by Pyrfume repository [31]. Leffingwell-Goodscent (GS-LF) [27, 28]. ... Sagar [33]. ... Keller [34]. ... Ravia [17]. ... Snitz [18].
Dataset Splits Yes We use nested 5-fold cross-validation to tune the hyper-parameters of the linear models and assess evaluation metrics on the test set that was held out during the training phase using an 80%-20% train-test split.
Hardware Specification No Our experiments do not require significant computational resources: we mostly train linear models that do not involve GPU usage or models that can be trained on a single commercially available GPU under one hour.
Software Dependencies No The paper mentions software and models used (e.g., Mo LFormer [25], Open-POM [32], Alva Desc [35]) but does not provide specific version numbers for these or for any other general software dependencies (e.g., programming languages, libraries).
Experiment Setup Yes We use nested 5-fold cross-validation to tune the hyper-parameters of the linear models and assess evaluation metrics on the test set that was held out during the training phase using an 80%-20% train-test split. This process is repeated 30 times using 30 different train-test splits. ... First, the dimensionality of the extracted representations is reduced to 20 using PCA, followed by z-scoring of each feature. Then, we train individual logistic regression models for each descriptor.