Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
The proposed architecture is novel but limited to the benchmarks. The curriculum learning has been already explored. Moreover, in terms of significance, how improving the performance in this task would be beneficial to real-world problems remains unclear. The presentation and clarity can be also improved please consider followings. - The description of the target task is unclear. - The definition of X_k at L151 is wrong. It confused me a lot and I thought that X_k is a set containing everything. I think the union operation should be replaced with a Cartesian product. - How can you be sure that the performance gain is from mitigating the effects of distracting features (L55)? The use of the curriculum does not directly mean mitigating the effects of distracting features. Please justify this argument as your paper is based on this argument.
Summary: This paper observes that the presence of distracting features is a key factor which causes substantial difficulties in abstract reasoning tasks. Inspired from the sequential nature of the human learning process, this paper proposes a feature robust abstract reasoning (FRAR) model which uses a reinforcement learning-based network as the teacher network and a Logic Embedding Network (LEN) that explicitly enumerates a much larger space of logic reasoning to disentangle abstract reasoning, as the student network. The authors further demonstrate the proposed model can have over 10% improvements over the state-of-the-art models on large-scale benchmark datasets, PGM  and RAVEN , in abstract reasoning. Originality: The main contribution of this paper lies in 1) its observation of one of the abstract reasoning’s fundamental characteristics: the co-presence of useful features for reasoning and distracting. Such observation provides an intuition why previous works on designing various kinds of models are not very effective for more sophisticated abstract reasoning tasks, and 2) following the key observation in 1), a proposed teacher-student framework that allows the teacher to learn to determine the appropriate teaching trajectory to optimize learner’s performance. A question I have is that how the task to distinguish reasoning features from distracting features differs from learning disentangled representations. Since many works have proposed their approaches to tackle the task of learning disentangled representations, it is recommended that the authors should also compare their proposed model with the approaches in previous works on learning disentangled representations (e.g. , , , etc.) under related work section. Quality: In general, this submission is technically sound, and the claims are mostly justified by the experiments. A few concerns are addressed below. The claims in this paper could be greatly strengthened by including a set of extrapolation experiments on the proposed model. Many real-world reasoning problems require the ability to extrapolate from limited training data (e.g., learning addition with 2-digit numbers and extrapolating to addition with 3-digit numbers). According to previous papers on abstract reasoning (e.g. ), the performance usually drops when the model is tested under extrapolation settings. As the performance on extrapolation is one of the key indicators of the model’s abstract reasoning ability and extrapolation can also be treated as one kind of distracting features, a set of experiment on extrapolation will further demonstrate the proposed model’s ability to distinguish distracting features and reasoning features. In addition, it would also be interesting to see how the performances of models other than LEN (e.g. RN, WReN, etc.) as the student networks can benefit from a teacher model. Clarity: The paper is generally well-written and structured clearly. Significance: This paper seems to be a useful contribution to the literature on abstract reasoning, showing a large improvement over the state of the art. =======================Post-rebuttal update============================== I am happy to main my ratings and recommend this work for acceptance after reading through other reviews as well as the author rebuttal, and engaging in the discussions. I look forward to seeing the discussions on disentangled representation, experimental results on extrapolation, and performances of models other than LEN in the camera ready version of this paper.  Adam Santoro, Felix Hill, David Barrett, Ari Morcos, and Timothy Lillicrap. Measuring abstract reasoning in neural networks. In International Conference on Machine Learning, pages 4477–4486, 2018.  Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019  Hoogeboom, Emiel. "Few-Shot Classification by Learning Disentangled Representations." Msc, University of Amsterdam (2017).  Bengio, Yoshua, et al. "A meta-transfer objective for learning to disentangle causal mechanisms." arXiv preprint arXiv:1901.10912 (2019).  Hsu, Kyle, Sergey Levine, and Chelsea Finn. "Unsupervised learning via meta-learning." arXiv preprint arXiv:1810.02334 (2018).
Originality: The model proposed is quite novel. The teacher network uses recent development in reinforcement learning and applies in an innovative way to provide a curriculum for abstract reasoning task. While other curriculum learning works exist, this is one of the first to use it for challenging visual reasoning tasks. The related works regarding both abstract reasoning and curriculum learning are properly discussed and the authors do a good job of putting the proposed work in context of existing work. Quality: The proposed method is well-motivated and technically sound. The experimental setup is also clear and shows that both of the contributions (the LEN model and teacher model) are independently effective for both RAVEN and PGM dataset surpassing existing art by a solid margins (over 10% for both RAVEN and PGM). I also appreciate the authors including several other competing curriculum learning algorithms as well as testing the proposed teacher model on a variety of algorithms. Overall, this a a good quality submission. Clarity: The details are mostly clear. There are some minor writing issues. First, the use of past tense in abstract and body of text is unusual. E.g. "Table 1 illustrated such an idea..." instead of "Table 1 illustrates such idea". While it does not affect clarity, it reads odd. Most of the figures are blurry, I would consider using a vector graphics for revised version for better readability. Significance: As outlined above, the paper has two main algorithmic contributions. The teacher model shows success for a wide variety of models and is likely to be adopted for other tasks as well. Similarly, a large improvement in challenging visual reasoning task is also significant. The authors introduce a number of small task-specific engineering changes, such as the use of separate streams for processing "shape" and "line" features which are also likely to be useful to the community for similar visual reasoning tasks. *** POST REBUTTAL COMMENTS *** After carefully reading all other reviews, the author rebuttal and engaging in detailed discussions with other reviewers, I am happy to recommend this work for acceptance. The main merits of the work is in solving abstract reasoning datasets that have been independently studied (and duly peer-reviewed) to be a good proxy tests for the visual reasoning task. There is definitely valuable discussion to be had for how true that statement exactly is (as noted by R1). However, the authors of this submission should not have to defend the validity of community-established benchmarks. The author rebuttal also shows success in other tasks such as CLEVR. My ratings is unchanged from before: An accept.