Deep Symbolic Learning: Discovering Symbols and Rules from Perceptions

Authors: Alessandro Daniele, Tommaso Campari, Sagar Malhotra, Luciano Serafini

IJCAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide experimental analysis to substantiate the efficacy of DSL in simultaneously learning perception and symbolic functions. and We evaluate our approach on a set of tasks where a combination of perception and reasoning is essential. in Section 5.
Researcher Affiliation Academia Alessandro Daniele1 , Tommaso Campari1, 2 , Sagar Malhotra1, 3 and Luciano Serafini1 1Fondazione Bruno Kessler, Trento, Italy, 2Universit a degli Studi di Padova, Italy 3TU Wien, Austria
Pseudocode No The paper does not contain any clearly labeled 'Pseudocode' or 'Algorithm' blocks. The methodology is described in prose and through architectural diagrams (Figure 1 and Figure 2).
Open Source Code Yes 1Project webpage: https://dkm-fbk.github.io/projects/dsl.html
Open Datasets Yes Firstly, we test our system on a variant of the MNIST [Le Cun et al., 1998] Sum task proposed in [Manhaeve et al., 2018]
Dataset Splits No The paper mentions a 'training set' and 'test set' for its experiments (e.g., 'we trained our model on the 2-digits sum and we evaluate the learned model on sequences of varying length'), and also mentions using 'optuna to select the best hyperparameters', which implies the use of a validation set. However, it does not provide specific details on the dataset splits (e.g., percentages, sample counts for train/validation/test) or reference a standard predefined split with a citation for all experiments, except partially for specific tasks (e.g., 'sequences of 4 images during training and 20 on the test' for Visual Parity).
Hardware Specification Yes All the experiments were conducted with a machine equipped with an NVIDIA GTX 3070 with 12GB RAM.
Software Dependencies No The paper mentions 'Mad Grad [Defazio and Jelassi, 2022] for optimization and optuna to select the best hyperparameters'. However, it does not provide specific version numbers for these or any other software dependencies, which is required for reproducibility.
Experiment Setup Yes To train the model we use the binary cross entropy loss on the confidence t of the predicted symbol. and In our experiments, we use ϵ-greedy during training, and greedy policy at test time. and We used Mad Grad [Defazio and Jelassi, 2022] for optimization and optuna to select the best hyperparameters for every experiment. and DSL reached an accuracy of 98.7 0.4 in 1000 epochs