Selected papers on modular neural nets for compositional reasoning.
Neural Module Networks. [NMN]
Andreas, Rohrbach, Darrell and Klein. CVPR 16.
http://arxiv.org/abs/1511.02799
Introduces the general module net framework. Neural module networks are input-specific neural networks built from a discrete library of shallow neural "modules" that are jointly trained to be freely composable (possibly subject to type constraints). In question answering settings, we can think of these question-specific networks as roughly corresponding to formal meaning representations, and use ideas from model-theoretic semantics to suggest the relevant primitives and types. But at a high-level this framework is just about building lots of different network layouts, tying pieces of their parameters together in complex structured ways, and training everything jointly. We know from Socher-style tree RNNs that this works for scoring and representing sentences; what's new with NMNs is that these networks induce not just representations but functions, which compose and generalize for transformations between real-world inputs and outputs.
W/r/t this paper, the good: then-SOTA on VQA. Maybe more interestingly, SHAPES experiments show that it's possible to train on simple questions and evaluate on complex ones without loss in performance; in other words, we can turn the ability to generalize about syntax immediately into the ability to generalize about vision. As far as I know these experiments haven't been repeated with any more recent versions of the framework. The bad: the set of delexicalized high-level structures we use for real vision tasks is tiny. Though VQA results were better than the baseline, they were just barely so, and in hindsight probably for reasons unrelated to the NMN part.
Learning to Compose Neural Networks for Question Answering.
Andreas, Rohrbach, Darrell and Klein, NAACL 16.
http://arxiv.org/abs/1601.01705
The above, repackaged for the semantic parsing crowd. Goes into a little more detail about connections to formal representations of meaning in linguistics; what's new technically is the use of a policy gradient method to learn a little bit of structure scoring, and empirically is the extension of these techniques to database-style QA problems. RL helps a tiny bit on VQA. As before, the real story is in the small-scale experiments: on geography QA, it actually makes a difference in learning whether to insert quantifiers and what branches to prune from prospective network layouts.
Inferring and Executing Programs for Visual Reasoning. [PG+EE]
Johnson, Hariharan, van der Maaten, Hoffman, Fei-Fei, Zitnick and Girshick. ICCV 17.
https://arxiv.org/abs/1705.03633
This paper begins by observing (correctly) that the original NMN work has a weird hacky dependence on a syntactic parser because we don't actually get to observe logical forms in language; it then "fixes" the problem by running only on fake datasets where we have ground truth LFs. The paper bills learning the layout predictor with a seq2seq model as one of the main contributions, but insofar as most modern parsers are basically seq2seq models this isn't actually any different from starting with a pretrained parser that gets run offline. In this respect there's basically no difference between this and the SHAPES experiments in the original CVPR paper. What is new & cool here is the module parameterization, which is convolutional but not attentional---there's much less structure imposed on the latent states in this model than the earlier work, and the fact that things still compose nicely is significant. But it's a lot of free parameters for a small vocabulary, and I'm pretty sure it would be impossible to train this model on any real-world dataset.
Learning to Reason: End-to-End Module Networks for Visual Question Answering. [N2NMN]
Hu, Andreas, Rohrbach, Darrell and Saenko. ICCV 17.
https://arxiv.org/abs/1704.05526
Also tries to relax the tight coupling between the parser and the layout predictor, but with more of an eye towards performance on real-world data. In particular, replaces hard assignment of words to modules with a soft, attentive selection, and uses parse trees only for coarse constituency structure. This finally makes it possible to use much deeper trees than in the original work. Much better performance than the earlier work; worse than other things on CLEVR but mostly because of a weird parameterization of the Transform module (which acts only on spatial attentions without looking at the underlying visual features). As a bonus, adds learned structure prediction with RL at the action level rather than the tree level; this helps a tiny bit on CLEVR but not on real data.
Neural Compositional Denotational Semantics for Question Answering
Gupta and Lewis. EMNLP 2018.
https://arxiv.org/abs/1808.09942
Totally different way to fdoing structure discovery from scratch: builds a complete parse-chart-shaped NN according to the CKY recurrence (sort of like Le and Zuidema, EMNLP 18), with soft gating to pick which descendants actually get used. Works through the kinds of types and operators you need to assign an NMN to every span of the input question (requires a few more pieces than just the attentions ← attentions and attentions ← truth values). Can solve structure discovery from scratch on symbolic world representations with either fake language (CLEVR) or real (GENx), even when the test set features more complex structures than the training set. No experiments on more complex groundings.
Modeling Relationships in Referential Expressions with Compositional Modular Networks. [CMN]
Hu, Rohrbach, Andreas, Darrell and Saenko. CVPR 17.
https://arxiv.org/abs/1611.09978
There's an (informal) duality in formal semantics between variable-binding /
Translating logical forms into NMNs, this distinction corresponds directly to the distinction between energy-based models and ordinary feedforward networks. While most of the NMN work is in the feedforward framework, the CMN paper (though it's not presented this way) is essentially the energy-based version: we resolve referring expressions by optimizing over input groundings to maximize a score at the output.
Using Syntax to Ground Referring Expressions in Natural Images.
Cirik, Berg-Kirkpatrick and Morency. AAAI 18.
https://www.cs.cmu.edu/~vcirik/assets/pdf/AAAI2018_cirik_using.pdf
Referring expression grounding, but in feed-forward mode. Gets basically the same results as the CMN paper, but identifies all the other referents in the image more accurately. Also improves dramatically on the hacky parse-tree-to-NMN-layout heuristic from the original NMN paper.
MAttNet: Modular Attention Network for Referring Expression Comprehension
Yu, Lin, Shen, Yang, Lu, Bansal and Berg. CVPR 18.
https://arxiv.org/abs/1801.08186
Quite similar to CMN paper; the main distinction here is that pieces of the predicate structure can themselves be weighted (which basically corresponds to a soft selection of all the combinatorially many substructures that can be formed from the fixed initial layout).
Transparency by Design: Closing the Gap Between Performance and Interpretability in Visual Reasoning. [TbD]
Mascharka, Tran, Solaski and Majumdar. CVPR 18.
https://arxiv.org/abs/1803.05268
Improved performance on CLEVR using attentional modules similar to those used in the N2NMN paper. (But hard, supervised assignment of words to structure pieces as in PG+EE.) A little over-engineered around CLEVR, but a nice discussion of how the transparency / debuggability of the NMN framework makes it easy to iterate on modeling and learning experiments.
Systematic Generalization: What is Required and Can It Be Learned?
Bahdanau et al. ICLR 19.
https://arxiv.org/abs/1811.12889
Experiments on synthetic data with NMNs and other VQA architectures about who's
robust to shifts between the training and test distributions. Looking only at
questions of the form
There's something that feels slightly wrong about the evaluation that's I can't
quite pin down, but related to the fact that we're looking at robustness to
shifts in
Modular Generative Adversarial Networks.
Zhao, Chang, Jie and Sigal.
https://arxiv.org/abs/1804.03343
Module nets for generative modeling. Implements a conditional GAN as an encoder, a sequence of modular transformations on the latent code, and a decoder. The encoded representation preserves spatial information, so we can use an attention mechanism like the one in the original question answering papers to localize the transformation to particular regions of the input image. Synthesized images look comparable to SOTA; numerical results are better but I don't really know how GAN evaluations are supposed to work. In principle you should be able to use this to learn sequences of transformations that don't commute, but all the experiments in the paper are just adding / removing attributes.
Neural modular control for embodied question answering.
Das, Gkioxari, Lee, Parikh and Batra. CoRL 18.
https://arxiv.org/abs/1810.11181
An options model with
policy-sketch-style supervision
and modular option parameterizations (which factor out verb--argument
combinations). Every option is a "flat" function indexed by an embedding that
takes the current world state onto a next action, but could presumably
incorporate more internal structure by expressing go-to[sofa] with a VQA-style
find module feeding into a controller. Modules are trained with subgoal
supervision.
Visual coreference resolution in visual dialogue using neural module networks.
Kottur, Moura, Parikh, Batra and Rohrbach. ECCV 2018.
https://arxiv.org/abs/1809.01816
N2NMN-style training with supervised attentions and a richer set of primitives
to support visual dialogue. A refer module can grab bounding boxes around
regions that have previously been attended to. Handles questions like "what else
is red" by combining this with negation and primitives from previous work. Does
a pretty good job with off-the-shelf parser layouts, getting then-SOTA results
on the visual dialogue challenge.