SSPictR

Official repository for SSPictR - A Neuro-symbolic Scene Representation for Human-aligned and Resource-efficient Computer Vision, in submission (2026).

Method

SSPictR is a biologically plausible image representation based on spatial semantic pointers (SSPs) [see Eliasmith 2013, Komer et al. 2019, Dumont and Eliasmith 2020]. SSPictR encodes semantic labels of objects and their spatial locations extracted from segmentation maps. It only requires a single vector to capture a compressed but fully decodable neuro-symbolic representation of an image.

For a demo of SSPictR as presented at the Artificial Intelligence and Cognition (AIC) workshop 2025 see SSPictR-Demo.ipynb.

Experiments

We evaluate human alignment of SSPictR on BOLD5000 (Chang et al. 2019): BOLD5000_RepresentationSimilarityAnalysis.ipynb

Further, we evaluate SceneReconstruction and OpenVocabulary.

As downstream classification task, we train a simple MLP model on an indoor subset of Places365 (Zhou et al. 2018). Training and testing can be found in Places365_SceneRecognition.ipynb.

We show out-of-distribution generalization of our scene recognition model on ADE20K (Zhou et al. 2017): Places365_SceneRecognition_ADE20K.ipynb

Citation

Please cite the following papers if you use SSPictR or parts of this code in your research:

@article{penzkofer26_sspictr,
  author = {Penzkofer, Anna and Habashy, Karim and Eliasmith, Chris and Bulling, Andreas},
  title = {{A Neuro-symbolic Scene Representation for Human-aligned and Resource-efficient Computer Vision}},
  year = {2026},
  doi = {}
}
@inproceedings{penzkofer25_aic,
  author = {Penzkofer, Anna and Habashy, Karim and Eliasmith, Chris and Bulling, Andreas},
  title = {SSPictR: A Biologically-plausible Image Representation},
  booktitle = {ECAI Workshop on Artificial Intelligence and Cognition (AIC)},
  year = {2025},
  pages = {1--13},
  doi = {}
}