Scalable Theory-Driven Regularization of Scene Graph Generation Models
Authors: Davide Buffelli, Efthymia Tsamoura
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that our technique can improve the accuracy of state-of-the-art SGG models, by up to 33%. Our empirical comparison confirms that NGP: improves the accuracy of state-of-the-art SGG models, namely IMP (Xu et al. 2017), MOTIFS (Zellers et al. 2018) and VCTree (Tang et al. 2019), by up to 33%; scales to theories including approximately 1M ICs sizes no prior symbolic-based regularization technique supports (Donadello, Serafini, and d Avila Garcez 2017); is particularly effective when applied in conjunction with TDE (Tang et al. 2020), a technique that tackles the bias in the data, improving the performance of IMP, MOTIFS and VCTree by up to 16 percentile units; outperforms GLAT (Zareian et al. 2020) and LENSR (Xie et al. 2019), two state-of-the-art regularization techniques that maintain the knowledge in subsymbolic form, by up to 18% and 15%; improves the accuracy of SGG models by up to six times when restricting the availability of ground-truth facts. |
| Researcher Affiliation | Collaboration | Davide Buffelli* 1, Efthymia Tsamoura 2 1 University of Padova, Via Gradenigo 6/b 35131, Padova, Italy 2 Samsung AI, 50-60 Station Road CB1 2JH, Cambridge, United Kingdom |
| Pseudocode | Yes | Algorithm 1: NGP(I, F, T, nt) ... Algorithm 2: GREEDY(I, ρ, T, nt) |
| Open Source Code | Yes | The sources and the data to reproduce our empirical analysis are in https://github.com/tsamoura/ngp. |
| Open Datasets | Yes | Following previous works, e.g., (Zareian, Karaman, and Chang 2020; Li et al. 2021), we use Visual Genome (VG) (Krishna et al. 2017) with the same split adopted by (Tang et al. 2020), and the Open Images v6 (OIv6) benchmark (Kuznetsova et al. 2020) with the same split adopted by (Li et al. 2021). |
| Dataset Splits | Yes | Following previous works, e.g., (Zareian, Karaman, and Chang 2020; Li et al. 2021), we use Visual Genome (VG) (Krishna et al. 2017) with the same split adopted by (Tang et al. 2020), and the Open Images v6 (OIv6) benchmark (Kuznetsova et al. 2020) with the same split adopted by (Li et al. 2021). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory) used for running the experiments. It mentions 'computational resources' in general terms. |
| Software Dependencies | No | The paper does not provide specific software dependencies or library versions (e.g., Python, PyTorch, TensorFlow versions) used for the implementation. |
| Experiment Setup | Yes | We set the number of constraints (i.e., ρ in Eq. 3) to ρ = 3. We found that this value adds minimum computational overhead while improving m R and zs R. We considered the loss functions DL2 (Fischer et al. 2019) (fuzzy logic) and SL (Xu et al. 2018) (probabilistic logic). ... β1 and β2 are hyperparameters setting the importance of each component of the loss. In our empirical evaluation, those hyperparameters are computed in an automated fashion using (Kendall, Gal, and Cipolla 2018). |