NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1052
Title:Meta-Learning Representations for Continual Learning

Reviewer 1

* Writing and clarity - Some descriptions are unclear and hard to follow, e.g., L125-126. - The abbreviation MRCL is used before its definition (Abstract, Fig.3, L171). - How do you define the marginal distribution \mu in L73? Please clearly define -- the number of samples for a given X_t? Then, it's not a probability distribution. - In section 2, the authors assumed "a variety of correlated sequences" (L74). However, the rationale of OML in Eqn.3 is not sufficient to support the "correlated sequences". Even though consulting with Appendix B, the rationale is weak to persuade. Do you assume that the k-step online update in Eqn.3 can give optimal loss for RLN although k is smaller than the length of each session? And, what do you mean by "finding a model initialization and learning a fixed representation such that starting from the learned representation it has xyz properties (Appendix L379-380)"? - I strongly recommend the authors to move Algorithm 1 from Appendix to the paper after polishing Section 2 & 3. * Related works Missing related works make it hard to assess the novelty and significance of the proposed method. If it is possible, please do report the controlled experiments to compare state-of-the-art. Some parts of the comparison with the other methods should be mentioned in the paper to shape the position among the related works. The following papers (not exhaustively listed) are recommended to consider: 1) Meta-learning with Latent Embedding Optimization, Rusu et al., 2018 2) TADAM: Task Dependent Adaptive Metric for Improved Few-shot Learning, Oreshkin et al., 2018 3) Task-Agnostic Meta-Learning for Few-shot Learning, Jamal et al., 2019 * Evaluations - L209-210 says that training accuracy is the indicator for the amount of forgetting. However, Fig.3 shows the limited aspects of analysis on the catastrophic forgetting and how the models persistently maintain the information. Could you plot the progressive results of the accuracies? - There are only two experiments using sine waves and Split-Omniglot, which are both too simple to confirm its significance. More complex and realistic datasets should be tested, e.g., CIFAR100 (Krizhevsky, 2009), Mini-Imagenet (Vinyals et al., 2016), and Tiered-ImageNet (Ren et al., 2018).

Reviewer 2

Originality: There are previous works that fixes the feature representation while only adapts the task learning part for new tasks. AFAIK there's no previous work using meta learning for this. Quality and Clarity: The idea is simple and well illustrated in the paper, however the experimental protocol is less clear. 1. Does the baseline pretraining mean that the model is trained with iid pretraining and then go over the training set again with online setting? The pretraining baseline having the same name as the pretraining set makes it a bit hard for the reader. (If I understand correctly, the pretraining set means the meta-training set.) 2. In split omniglot (a), is the random batch sampled from the entire pretraining set or just those that are already observed? (b) says sampling a single class, is the training phase a one pass through a trajectory? or does it allow revisiting of the data? It needs to the clarified. 3. Note that class id is implicitly used during omniglot training where a batch is sampled from a single class, which is in conflict with the claim that no task id is used. 4. I assume the first figure of Figure 4 is not very necessary because the training phase use a random batch for meta training, so it is implicitly an iid method, it is not very interesting to see its comparison to the other methods, could even be misleading that the MRCL is performing so close to Oracle. Limitations: 1. The evaluation with only Omniglot as the real dataset is limited, because Omniglot is known to show a favorable result for meta learning. 2. This method has a training phase and testing phase, in the training phase random sampling from previous data is required, thus the training phase is implicitly an IID method with cross validation. Only the test phase is under valid continual learning setting. This seems to assume that we could learn on a fixed training set and fixed the representation for all future distribution we encounter. This is not true, the reason why meta learning is successful is that the meta-training set should have a distribution that is similar to the meta-test set. This is incompatible with the continual learning goal that we want to dynamically integrate new knowledge into the network. Overall, I think the contribution of this paper is to use meta-learning (which basically means learning by cross validation to me) to find a representation that is potentially more disentangled and less prone to forgetting when adapting the prediction layer. This is a valid problem, but limited for continual learning.

Reviewer 3

Overall, this is a well-written paper that describes an interesting approach to continual learning (namely, meta-learning good initial representation to facilitate online learning). The paper draws connections to the literature well and provides interesting analysis of the representations learnt in this fashion. One shortcoming is that the only two datasets/tasks are considered and these are not large-scale enough to completely evaluate the performance of the method. There’s an open question of how this method will scale to larger tasks and the feasibility of this approach. For example, what do you pre-train on if you want to do continual learning on Atari games? Nonetheless it should provide a good starting point for further research. The rest of this review provides detailed comments as well clarifications that are required. This paper deals with the important issue of continual learning (CL). Broadly, the objective of continual learning is to allow a network to learn a series of tasks online, without forgetting previously learnt tasks (catastrophic forgetting). There are a range of different ways to tackle this problem, and this paper starts by providing a clear motivation and a comprehensive literature review. Most current methods are largery optimisation based, or memory/replay based. The approach (MRCL) this paper takes is to learn a representation such that when used in a CL setting, it allows for greater transfer and less interference. This is similar to a MAML-style method for meta-learning, where the idea is to find an initialization that allows for rapid adaptation. Instead here, we want a representation that allows for continual learning. (Another work to reference that is missing from the paper is “Memory-based parameter adaptation” that does a similar MAML style update albeit from a memory-buffer). The method here involves splitting the network into two sets of weights: a representation learning network (RLN, weights W) and a task learning network (TLN, weights T). The TLN will be trained to maximise reward/minimise loss on the task at hand, with fixed representation from the RLN. After doing this for k steps, we will have a set of weights W_k. We will then sample a fresh batch and update the initial W_0 and T_0 wrt to the loss on this current batch with parameters W_k and T_0, to give us W’ and T’. Thus, we meta-learn an initial representation that allows for good online learning. There is good intuition provided as to what such a representation looks like wrt the solution manifold. * Please clarify in equation (3) what the expectation is taken with respect to. Further please consider moving the diagram and pseudocode from the appendix to the main paper for clarity. The first task considered is an incremental sine wave task where there are K sine waves to be regressed. At pre-training, the RLN is meta-trained in the fashion above, and at the evaluation time is fixed as a feature extractor. This is compared to a pure pre-training method and another SR-NN method. Please provide more details (possibly in the appendix) for how this sparse SR-NN method works. We find the MRCL allows less interference than other method. This approach is also repeated for split-omniglot. The paper also provides good insight into sparsity that results from training in this fashion However, it would be helpful if the authors could provide a concrete definition of “instance sparsity”. This section could also be shortened somewhat. One question is how more traditional methods like EWC or memory-based methods would do in this regime. This method uses a modified version of the standard CL setup: ie allowing pre training before the tasks. This is evaluated in the last section where MRCL is viewed as orthogonal to these methods and combined as a pre-training method. Could the authors however please clarify the numbers in Table 1 -- are these average over all tasks? How do these related to figure 4? The numbers for EWC are surprisingly low, to me. Could you provide some intuition as to why? How converged were the models when the fisher matrix is calculated? One possible extension would also be to run this on permuted MNIST (where you pre-train on another set of permutations) and compare to doing standard EWC with/without pre-training. This would allow the numbers to be compared to more standard benchmark tasks in the literature. Overall, this paper is a very useful contribution to the field and will provide a significant starting point for future research. I recommend acceptance.