Multi-modal Discriminative Model for Vision-and-Language Navigation

Haoshuo Huang, Vihan Jain, Harsh Mehta, Jason Baldridge, Eugene Ie

TL;DR

This paper introduces a novel discriminator model that learns to assess the alignment between natural language instructions and visual navigation paths. By leveraging this discriminator, the authors demonstrate that only a small fraction of augmented, high-quality data is needed to significantly improve vision-and-language navigation agents, and that initializing agents with discriminator components leads to substantial performance gains.

What Problem Does This Paper Address?

Vision-and-Language Navigation (VLN) is a challenging task where an agent must follow natural language instructions to navigate through a visual environment. A major bottleneck in VLN is the scarcity of high-quality, human-annotated data. While data augmentation techniques exist, they often generate noisy or poorly aligned instruction-path pairs. This paper addresses the problem of effectively utilizing these augmented datasets by proposing a method to discriminate between good and bad instruction-path alignments, thereby enabling more efficient training of VLN agents and improving their generalization to unseen environments.

Methods

The core contribution is a multi-modal discriminative model with a two-tower architecture. Each tower (one for language, one for visual input) uses a bidirectional LSTM to encode sequences into latent representations. An alignment layer then computes a similarity score between these representations using a novel alignment-based similarity metric. This metric involves matrix transpose, softmax pooling, and softmin operations to encourage the model to construct optimal alignments between the instruction and the path.

To train this discriminator, the authors employ a strategy of negative path sampling. Given a positive instruction-path pair from the R2R dataset, they generate negative pairs by either substituting entire paths, partially reordering path segments, or performing random walks. This allows the discriminator to learn what constitutes a poor alignment. They also explore curriculum learning, where the discriminator is first trained on simpler "Path Substitution" negatives and then on more complex "Partial Reordering" and "Random Walk" negatives, leading to improved performance.

Datasets

The primary dataset used is Room-to-Room (R2R) (Anderson et al., 2018b), a photo-realistic dataset for VLN. The authors utilize the R2R training set for positive examples and also use its validation splits (seen and unseen) for evaluation. They also leverage the Fried-Augmented dataset (Fried et al., 2018), which consists of machine-generated instructions for R2R paths, to test their data selection strategy.

Key Insights and Ideas

  • Quality over Quantity: The paper strongly suggests that for augmented VLN datasets, the quality of instruction-path pairs is far more important than sheer quantity. A small fraction of high-scoring pairs from the discriminator can yield significant improvements.
  • Discriminator as a Data Filter: The proposed discriminator acts as a powerful filter, identifying genuinely useful augmented data that would otherwise be discarded or even harm model performance.
  • Multi-modal Alignment is Key: The discriminator's ability to learn meaningful alignments between textual descriptions and visual scenes is crucial. This learned alignment capability can be directly transferred to improve navigation agents.
  • Curriculum Learning for Robustness: Training the discriminator with a curriculum of increasingly complex negative samples (PS -> PR/RW) leads to a more robust model that generalizes better.

Key Results

  • The discriminator effectively distinguishes between high-quality and low-quality instruction-path pairs from the Fried-Augmented dataset.
  • Training a VLN agent with just the top 1% of Fried-Augmented data, as scored by the discriminator, is sufficient to match or exceed benchmark performance on unseen environments.
  • Initializing a navigation agent's instruction and visual encoders with components from the trained discriminator leads to a relative performance improvement of over 10% on unseen environments compared to existing benchmarks.
  • The discriminator trained with curriculum learning achieves higher AUC scores on validation sets containing more complex negative samples (PR, RW) compared to models trained without it.

Limitations and Future Work

  • Validation Seen Performance Drop: While the agent initialized with the discriminator performs well on unseen environments, its performance drops on validation seen environments. The authors suggest further experimentation is needed to improve performance on both.
  • Future Work: The authors propose using the discriminator in conjunction with generative models to filter low-quality augmented data and act as a reward signal. They also suggest segmenting instruction-path pairs into shorter segments for curriculum learning and applying the multi-modal alignment measure to other downstream tasks beyond VLN.

Summary updated

Abstract

Vision-and-Language Navigation (VLN) is a natural language grounding task where agents have to interpret natural language instructions in the context of visual scenes in a dynamic environment to achieve prescribed navigation goals. Successful agents must have the ability to parse natural language of varying linguistic styles, ground them in potentially unfamiliar scenes, plan and react with ambiguous environmental feedback. Generalization ability is limited by the amount of human annotated data. In particular, \emph{paired} vision-language sequence data is expensive to collect. We develop a discriminator that evaluates how well an instruction explains a given path in VLN task using multi-modal alignment. Our study reveals that only a small fraction of the high-quality augmented data from \citet{Fried:2018:Speaker}, as scored by our discriminator, is useful for training VLN agents with similar performance on previously unseen environments. We also show that a VLN agent warm-started with pre-trained components from the discriminator outperforms the benchmark success rates of 35.5 by 10\% relative measure on previously unseen environments.
BibTeX
@inproceedings{huang2019multi,
  author = {Haoshuo Huang and Vihan Jain and Harsh Mehta and Jason Baldridge and Eugene Ie},
  title = {{Multi-modal Discriminative Model for Vision-and-Language Navigation}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
}
Figure