NeurIPS 2020

Multifaceted Uncertainty Estimation for Label-Efficient Deep Learning

Review 1

Summary and Contributions: The authors propose to do active learning by using uncertainty estimates, derived from a Dirichlet Prior Network model, decomposed into vacuity and dissonance to do active learning in a low-data regime. The proposed approach outperforms competing methods.

Strengths: The current work is relevant to the NeurIPS community. The proposed approach is rooted in firmly established Dempster-Shafer theory and Dirichlet Prior Network models, and outperforms alternative methods. The paper is clear and well written. Also, I believe this is the first paper, that I've seen at least, which examines active learning using a Dirichlet Prior Network style model.

Weaknesses: I have several concerns regarding this work. 1. Firstly, I'm not entirely convinced by the need to introduce an evidence-based Dempster Schafer / Subjective logic framework. The proposed decomposition into vacuity and dissonance is essentially the same thing as decomposing into epistemic uncertainty and aleatoric uncertainty. Why not consider the tractable closed form mutual information decomposition into total, epistemic and aleatoric uncertainties which was derived in previous work? I believe that decomposition would, broadly speaking, have many of the same quantities. Similarly, the definition for vacuity (W/S) is almost the same as for expected pairwise KL-divergence introduce in Malinin et al ( (W-1)/S ). Given a DPN trained with the proposed loss, I suspect that mutual information (vacuity) and expected entropy (dissonance) would enable many of the same behaviours. Furthermore, it would allow comparing directly to approaches such as BALD (which just use mutual information, derived via dropout). 2. It would have been really good to see active learning experiments of a greater scale, using for example SVHN/CIFAR-100/TinyImageNet. While the proposed method outperforms the baselines in this small-scale setting (MNIST/notMNIST/CIFAR-10), it is important to see how well this scales to larger and more challenging datasets, and therefore active learning scenarios. I am also surprised by your choice of architecture. Would there be difficulties in getting it to work on more modern architectures, such as WideResNets and DenseNets?

Correctness: The proposed method seems to be mathematically correctly, and is rooted in firmly established theory.

Clarity: Paper is clear and well written.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback:

Review 2

Summary and Contributions: This paper offers a new take on data uncertainty under the guise of evidence-based uncertainty which it decomposes into vacuity and dissonance, under the framework of subjective logic (ref). This decomposition enables the design of a vacuity-aware regularizer that is claimed to guide the model convergence towards the correct decision boundaries, even in over-parametrized models with little data. In the context of active learning, the authors propose a sampling function that is the weighted sum of dissonance and vacuity where vacuity is annealed for the purpose of “balancing” as the authors call it.

Strengths: > The evidence-based entropy decomposition, while not novel or surprising, streamlines and formalizes the quantification of vacuity and dissonance that are important to the practical deployment of decision making, especially in the context of out-of-distribution test data. > The proposed decomposition allows the design of a vacuity-aware regularizer, based on a set of "anchor samples" which represent OOD areas in the data space. Optimizing this loss ensures the model predicts high vacuity (lack of evidence) in these areas. This is reminiscent of aspects of contrastive learning which as been found to help the calibration of uncertainty estimation. > A weighted sum of vacuity and dissonance is used as a new sampling function for active learning. The effect of dissonance is increased over training iterations on the assumption that the decision boundary is properly avoiding OOD areas and can be fine-tuned on high data density regions. > Empirical evaluation and ablation studies seem to support the validity of this approach as it outperforms the baselines.

Weaknesses: > Writing: > > The manuscript is quite verbose and the writing is awkward in certain parts. Everything prior to the experimental section could be rewritten in a more concise matter and certainly with less repetition. > > It should be made clear what "evidence-based entropy" is especially when claiming that entropy can be decomposed into vacuity and dissonance. While it is clear that those two quantities affect entropy, it is not clear how that happens analytically. A reader can be misled to think that entropy is equal to the sum of dissonance and vacuity given the abundant repetition of the phrase "entropy decomposition into vacuity and dissonance" throughout the text. > This work is not well placed with respect to prior literature in that it omits work on data uncertainty such as Malinin & Gales, 2018 and Hafner et al., 2018. > Empirical evaluation: I would have preferred to see more details in the main text instead of the appendix. This should be possible if the first 6 pages are shrunk down to 4. -----------POST-REBUTTAL UPDATE----------------- While the feedback has addressed some of my concerns, I still find both the lacking empirical evidence and the verbose presentation of this work's novelty work (as well as its placement with regards to prior work) insufficient to increase this score beyond a "marginally above the acceptance threshold".

Correctness: The claims and method are generally correct.

Clarity: The clarity of this paper is lacking. The writing is verbose and at times repetitive.

Relation to Prior Work: Differences to previous contributions are not well discussed as this papers omits prior work as Noise-Contrastive Priors (Hafner et al., 2018) and Prior Networks (Malinin & Gales, 2018) which tackle data uncertainty. Calling data uncertainty evidence-based uncertainty obfuscates such connections and might confuse the reader. In fact, noise-contrastive priors leverage the idea of training a model to output high uncertainty for data points outside of the training distribution. This is similar to the vacuity regularization term based on anchor samples. On the other hand, prior networks analyze the effects of vacuity and dissonance (without calling them as such) on OOD detection.

Reproducibility: Yes

Additional Feedback:

Review 3

Summary and Contributions: The paper proposes an uncertainty-aware active learning strategy for a deep learner inspired by subjective logic. Similar to prior networks, the deep learner is designed to regress evidence for a Dirichlet distribution such that areas far from the labeled training data will exhibit high vacuity and areas near class label boundaries will exhibit high dissonance. The active learning part is annealed so that it initially looks for high vacuity samples to label and over time switches to searching for high dissonance samples to label.

Strengths: The idea to incorporate vacuity and dissonance in an active learning framework seems highly novel and potentially very useful. The actual design of the evidence-based deep learner in how it biases high vacuity for samples far from labeled samples is very clever. The experiments do demonstrate the utility of the approach against BALD and a strawman uncertainty-aware approach derived from EDL.

Weaknesses: EDL does not necessarily exhibit high vacuity (or epistemic uncertainty) far from the labeled training data as it is not trained to do so. It seems that a more fitting uncertainty-aware strawman to compare against is prior networks (see ref [11] in the paper). Once the proposed deep learner is trained, it seems that the deep leaner is no longer uncertainty (or evidential) aware. Certainly, the experiments only demonstrate the classification accuracy of the deep learner and not its ability to properly capture vacuity. [After reading the Author Feedback] The authors have addressed my main concerns. If possible, it would be nice to see the Prior Network results in the paper.

Correctness: From what I can tell all the equations are correct. There might be some conditions where dissonance is undefined, e.g., when u = 1.

Clarity: The paper is well written, but there are some points that are not clear. It seems that after each sample is labeled, the parameters for the evidential deep model must be updated. Some words about the computational complexity for this update seems necessary.

Relation to Prior Work: The paper does an adequate job discussing prior related work.

Reproducibility: Yes

Additional Feedback: Of minor note: Line 97: Should DU be EU? Around eqn. (5) it would help the reader to understand what is a reasonable value for W. Line 189: K -> \infty, should K actually be C? Line 277: It seems that Beta can go negative. Is that true?