SCAN: Learning Hierarchical Compositional Visual Concepts

Authors: Irina Higgins, Nicolas Sonnerat, Loic Matthey, Arka Pal, Christopher P Burgess, Matko Bošnjak, Murray Shanahan, Matthew Botvinick, Demis Hassabis, Alexander Lerchner

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the performance of SCAN on a dataset of visual frames and corresponding symbolic descriptions collected within the Deep Mind Lab environment (Beattie et al., 2016).
Researcher Affiliation Industry Deep Mind, London, UK {irinah,sonnerat,lmatthey,arkap,cpburgess, matko,botvinick,demishassabis,lerchner}@google.com
Pseudocode No The paper contains architectural diagrams (e.g., Figure 2A, Figure 3A) but no formal pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement or link for the open-source code of the described methodology.
Open Datasets Yes Environment We evaluate the performance of SCAN on a dataset of visual frames and corresponding symbolic descriptions collected within the Deep Mind Lab environment (Beattie et al., 2016). We ran additional experiments on a more realistic dataset Celeb A (Liu et al., 2015)... In this section we describe additional experiments testing SCAN on the d Sprites (Matthey et al., 2017) dataset.
Dataset Splits Yes We split the dataset into two subsets. One was used for training the models, while the other one contained a held out set of 300 four-gram concepts that were never seen during training, either visually or symbolically. The accuracy and diversity metrics were calculated on two sets of sym2img samples: 1) train, corresponding to the 133 symbols used to train the models; and 2) test (symbols), corresponding to a held out set of 50 symbols.
Hardware Specification No The paper mentions using TensorFlow but does not provide specific details on the hardware used, such as GPU or CPU models, or memory specifications.
Software Dependencies No The padding algorithm used was SAME in Tensor Flow (Abadi et al., 2015). The optimiser used was Adam (Kingma & Ba, 2015) with a learning rate of 1e 3 and ϵ = 1e 8.
Experiment Setup Yes We set βy = 1 for all experiments, and λ = 10. We trained the model using Adam optimiser with learning rate of 1e 4 and batch size 16.