[Paper Review] CLOSURE: Assessing Systematic Generalization of CLEVR Models

Dzmitry Bahdanau, Harm de Vries+5 more|arXiv (Cornell University)|Dec 12, 2019

Multimodal Machine Learning Applications30 references36 citations

TL;DR

The paper introduces the CLOSURE benchmark to test whether CLEVR-trained models generalize systematically to novel referring expressions. It shows substantial generalization gaps across end-to-end and neural-symbolic models, introduces Vector-NMN to improve compositional generalization, and explores few-shot transfer effects.

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.

Motivation & Objective

• Assess whether CLEVR-trained models exhibit systematic generalization to novel referring expressions in CLOSURE.
• Quantify generalization gaps of end-to-end and symbolic-program-based VQA models on new, CLEVR-like tasks.
• Investigate model variants (including a Vector-NMN) that enhance compositional generalization.

Proposed method

• Construct seven CLOSURE tests by substituting matching referring expressions into CLEVR-like questions to create out-of-distribution, yet semantically compositional, queries.
• Evaluate a range of models on zero-shot generalization: FiLM, MAC, NS-VQA, NMN variants, and the proposed Vector-NMN.
• Compare ground-truth programs versus learned program generators, and assess the impact of providing ground-truth programs at test time (GT-*).
• Propose Vector-NMN with vector-valued inputs/outputs to improve modular generalization, building on FiLM-style conditioning and shared CNN/U filters.
• Explore few-shot transfer by finetuning pretrained models with 36 examples per CLOSURE family, with and without explicit program supervision.

Experimental results

Research questions

• RQ1 How do state-of-the-art CLEVR models generalize to novel combinations of referring expressions (CLOSURE) not seen during CLEVR training?
• RQ2 Do symbolic-program-based approaches (NS-VQA, NMN) exhibit better or worse systematic generalization compared to end-to-end models on CLOSURE?
• RQ3 Can a vector-based NMN module improve cross-compositional generalization, and under what conditions?
• RQ4 What is the impact of few-shot transfer on systematic generalization for both program-based and end-to-end models?

Key findings

• Most CLEVR-trained models show substantial degradation on CLOSURE tests versus their baseline CLEVR-like questions, indicating weak systematic generalization.
• MAC performs better than FiLM but still loses 15–35% of baseline accuracy on several CLOSURE tests, revealing a generalization gap.
• NS-VQA and NMN variants exhibit notable generalization gaps, especially on logical referring expressions, even with ground-truth programs in some cases.
• Vector-NMN generalizes much better than Tensor-NMN, often matching baseline performance across tests, with the exception of and_mat_spa.
• Providing few-shot examples (36 per CLOSURE family) substantially improves CLOSURE performance for most models; program-based models often reach target accuracy with supervision, while end-to-end models benefit from continued training.
• Ground-truth programs at test time dramatically help program-based models, highlighting the reliance on accurate program grounding for compositional tasks.