CLOSURE: Assessing Systematic Generalization of CLEVR Models

TL;DR

The paper addresses whether high CLEVR accuracy implies systematic generalization by introducing the CLOSURE benchmark, which tests recombination of referring-expression primitives in novel contexts. It evaluates end-to-end models (FiLM, MAC) and modular/symbolic approaches (NS-VQA, NMN variants) and introduces Vector-NMN to improve generalization, analyzing both zero-shot and few-shot transfer. The findings reveal substantial generalization gaps across most models, with Vector-NMN offering notable zero-shot gains and few-shot supervision enabling program-based methods to adapt more effectively. This work emphasizes the need for models that robustly recombine known primitives and points to vector-based modular architectures as a promising direction for more generalizable VQA systems.

Abstract

The CLEVR dataset of natural-looking questions about 3D-rendered scenes has recently received much attention from the research community. A number of models have been proposed for this task, many of which achieved very high accuracies of around 97-99%. In this work, we study how systematic the generalization of such models is, that is to which extent they are capable of handling novel combinations of known linguistic constructs. To this end, we test models' understanding of referring expressions based on matching object properties (such as e.g. "another cube that is the same size as the brown cube") in novel contexts. Our experiments on the thereby constructed CLOSURE benchmark show that state-of-the-art models often do not exhibit systematicity after being trained on CLEVR. Surprisingly, we find that an explicitly compositional Neural Module Network model also generalizes badly on CLOSURE, even when it has access to the ground-truth programs at test time. We improve the NMN's systematic generalization by developing a novel Vector-NMN module architecture with vector-valued inputs and outputs. Lastly, we investigate how much few-shot transfer learning can help models that are pretrained on CLEVR to adapt to CLOSURE. Our few-shot learning experiments contrast the adaptation behavior of the models with intermediate discrete programs with that of the end-to-end continuous models.

Figures (8)

• Figure 1: CLEVR questions (Q1 and Q2) require complex multi-step reasoning about the contents of 3D-rendered images. We construct CLOSURE questions (Q3) by using the referring expressions that rely on matching object properties (e.g. the red fragment in Q1) in novel contexts, such as e.g. comparison questions with two referring expressions (Q2).
• Figure 2: Programs P1, P2, P3 that define the ground-truth meaning for the questions Q1, Q2 and Q3 in Figure \ref{['fig:closure']}. The fragments in red correspond to the matching REs in the respective questions.
• Figure 3: Zero-shot accuracy of all models on the 7 CLOSURE tests. For each model and test, the white bar in the background is the model's accuracy on the closest CLEVR questions. The hatching used for "GT-..." models indicates that we used the ground-truth programs at test time.
• Figure 4: The accuracies for NS-VQA, PG-Vector-NMN and MAC after finetuning on 36 examples from each CLOSURE family. The background white bar is the model's accuracy on the closest CLEVR questions. The yellow horizontal line denotes the model's accuracy before fine-tuning. The hatching indicates the use of ground-truth programs at the fine-tuning stage.
• Figure 5: CLOSURE templates.