Datasets:

MihirRajeshPanchal

/

IITPatnaPeerReview

like 0

The paper is fairly clear and these extensions are reasonable[paper-POS], [CLA-POS]

paper

CLA

POS

POS

. However, I just don't think the focus on 2D grid-based navigation has sufficient interest and impact[null], [IMP-NEG]

IMP

NEG

. It's true that the original VIN paper worked in a grid-navigation domain, but they also had a domain with a fairly different structure; I believe they used the gridworld because it was a convenient initial test case, but not because of its inherent value.[null], [EMP-NEU]

EMP

NEU

So, making improvements to help solve grid-worlds better is not so motivating[null], [IMP-NEG]

IMP

NEG

. It may be possible to motivate and demonstrate the methods of this paper in other domains, however.[methods-NEU], [EMP-NEU]

methods

EMP

NEU

NEU

The work on dynamic environments was an interesting step: it would have been interesting to see how the models learned for the dynamic environments differed from those for static environments.[null], [CMP-POS]

CMP

POS

Superior performance to recent baselines (e.g. EWC) is reported in several cases.[performance-POS], [CMP-POS]

performance

CMP

POS

POS

Unfortunately, the paper does not go beyond the relatively simplistic setup of sequential MNIST, in contrast to some of the methods used as baselines.[null], [CMP-NEU, EMP-NEG]

CMP

EMP

NEU

NEG

The proposed architecture implicitly reduces the continual learning problem to a classical multitask learning (MTL) setting for the LTM, where (in the best case scenario) i.i.d. data from all encountered tasks is available during training. This setting is not ideal, though.[architecture-NEU, setting-NEG], [EMP-NEU]

architecture

setting

EMP

NEU

NEG

NEU

There are several example of successful multitask learning, but it does not follow that a random grouping of several tasks immediately leads to successful MTL.[null], [EMP-NEU]

EMP

NEU

Indeed, there is good reason to doubt this in both supervised and reinforcement learning domains.[null], [EMP-NEU]

EMP

NEU

I agree that problems can be constructed where these assumptions hold, but this core assumption is limiting.[assumption-NEU], [EMP-NEU]

assumption

EMP

NEU

NEU

The requirement of task labels also rules out important use cases such as following a non-stationary objective function, which is important in several realistic domains, including deep RL.[null], [EMP-NEG]

EMP

NEG

Experiments show a clear advantage during learning when compared with a vanilla DQN. [Experiments-POS], [EMP-POS]

Experiments

EMP

POS

POS

Nonetheless, there are some criticisms than can be made of both the method and the evaluations:[method-NEU, evaluations-NEU], [EMP-NEU]

method

evaluations

EMP

NEU

NEU

NEU

The fear radius threshold k_r seems to add yet another hyperparameter that needs tuning.[null], [EMP-NEG]

EMP

NEG

Judging from the description of the experiments this parameter is important to the performance of the method and needs to be set experimentally.[description-NEU, experiments-NEU, parameter-NEU, performance-NEU, method-NEU], [EMP-NEU]

description

experiments

parameter

performance

method

EMP

NEU

NEU

NEU

NEU

NEU

NEU

There seems to be no way of a priori determine a good distance as there is no way to know in advance when a catastrophe becomes unavoidable.[null], [EMP-NEU]

EMP

NEU

No empirical results on the effect of the parameter are given.[empirical results-NEG], [SUB-NEG, EMP-NEG]

empirical results

SUB

EMP

NEG

NEG

NEG

The experimental results support the claim that this technique helps to avoid catastrophic states during initial learning.[experimental results-POS, claim-NEU, technique-POS], [EMP-POS]

experimental results

claim

technique

EMP

POS

NEU

POS

POS

The paper however, also claims to address the longer term problem of revisiting these states once the learner forgets about them, since they are no longer part of the data generated by (close to) optimal policies.[paper-NEU], [IMP-NEU]

paper

IMP

NEU

NEU

This problem does not seem to be really solved by this method.[problem-NEU, method-NEG], [EMP-NEG]

problem

method

EMP

NEU

NEG

NEG

Danger and safe state replay memories are kept, but are only used to train the catastrophe classifier.[null], [EMP-NEG]

EMP

NEG

While the catastrophe classifier can be seen as an additional external memory, it seems that the learner will still drift away from the optimal policy and then need to be reminded by the classifier through penalties.[null], [EMP-NEU]

EMP

NEU

As such the method wouldn't prevent catastrophic forgetting, it would just prevent the worst consequences by penalizing the agent before it reaches a danger state.[method-NEG], [EMP-NEG]

method

EMP

NEG

NEG

It would therefore be interesting to see some long running experiments and analyse how often catastrophic states (or those close to them) are visited.[experiments-NEU], [EMP-NEU]

experiments

EMP

NEU

NEU

Overall, the current evaluations focus on performance and give little insight into the behaviour of the method.[evaluations-NEG, performance-NEU, method-NEU], [SUB-NEG, EMP-NEG]

evaluations

performance

method

SUB

EMP

NEG

NEU

NEU

NEG

NEG

In general the explanations in the paper often often use confusing and imprecise language, even in formal derivations, e.g. 'if the fear model reaches arbitrarily high accuracy' or 'if the probability is negligible'.[explanations-NEG], [CLA-NEG]

explanations

CLA

NEG

NEG

It is wasn't clear to me that the properties described in Theorem 1 actually hold.[Theorem-NEG], [CLA-NEG]

Theorem

CLA

NEG

NEG

The motivation in the appendix is very informal and no clear derivation is provided.[motivation-NEG], [PNF-NEG]

motivation

PNF

NEG

NEG

The authors seem to indicate that a minimal return can be guaranteed because the optimal policy spends a maximum of epsilon amount of time in the catastrophic states and the alternative policy simply avoids these states.[null], [EMP-NEU]

EMP

NEU

However, as the alternative policy is learnt on a different reward, it can have a very different state distribution, even for the non-catastrophics states.[null], [EMP-NEU]

EMP

NEU

It might attach all its weight to a very poor reward state in an effort to avoid the catastrophe penalty.[null], [EMP-NEU]

EMP

NEU

It is therefore not clear to me that any claims can be made about its performance without additional assumptions.[performance-NEU, assumptions-NEU], [EMP-NEG]

performance

assumptions

EMP

NEU

NEU

NEG

This seems to contradict the theorem.[theorem-NEG], [EMP-NEG]

theorem

EMP

NEG

NEG

It wasn't clear what assumptions the authors make to exclude situations like this.[assumptions-NEG], [EMP-NEG]

assumptions

EMP

NEG

NEG

However, I have the following concerns about the quality and the significance: - The proposed formulation in Equation (2) is questionable.[Equation-NEU], [EMP-NEU]

Equation

EMP

NEU

NEU

Since this approach is not straightforward, more theoretical analysis of the proposed method is desirable.[approach-NEU, theoretical analysis-NEU], [EMP-NEU]

approach

theoretical analysis

EMP

NEU

NEU

NEU

- In addition to the above point, I guess the expectation is needed as the original formulation of GAN.[null], [SUB-NEU]

SUB

NEU

Otherwise the proposed formulation does not make sense as it receives only specific data points and how to accumulate objective values across data points is not defined.[null], [EMP-NEU]

EMP

NEU

- In experiments, although the authors say lots of datasets are used, only two datasets are used, which is not enough to examine the performance of outlier detection methods.[experiments-NEG, datasets-NEG], [SUB-NEG]

experiments

datasets

SUB

NEG

NEG

NEG

Moreover, outliers are artificially generated in these datasets, hence there is no evaluation on pure real-world datasets.[evaluation-NEG, datasets-NEU], [EMP-NEG]

evaluation

datasets

EMP

NEG

NEU

NEG

To achieve the better quality of the paper, I recommend to add more real-world datasets in experiments.[experiments-NEU], [SUB-NEU]

experiments

SUB

NEU

NEU

- As discussed in Section 2, there are already many outlier detection methods, such as distance-based outlier detection methods, but they are not compared in experiments.[Section-NEU, experiments-NEU], [CMP-NEG]

Section

experiments

CMP

NEU

NEU

NEG

Although the authors argue that distance-based outlier detection methods do not work well for high-dimensional data, this is not always correct[null], [EMP-NEG]

EMP

NEG

. Please see the paper: -- Zimek, A., Schubert, E., Kriegel, H.-P., A survey on unsupervised outlier detection in high-dimensional numerical data, Statistical Analysis and Data Mining (2012) This paper shows that the performance gets even better for higher dimensional data if each feature is relevant.[performance-NEU], [CMP-NEU]

performance

CMP

NEU

NEU

I recommend to add some distance-based outlier detection methods as baselines in experiments.[baselines-NEU], [CMP-NEU]

baselines

CMP

NEU

NEU

- Since parameter tuning by cross validation cannot be used due to missing information of outliers, it is important to examine the sensitivity of the proposed method with respect to changes in its parameters (a_new, lambda, and others).[null], [EMP-NEU]

EMP

NEU

Otherwise in practice how to set these parameters to get better results is not obvious.[results-NEU], [EMP-NEU]

results

EMP

NEU

NEU

* The clarity of this paper is not high as the proposed method is not well explained.[clarity-NEG, proposed method-NEG], [CLA-NEG, EMP-NEG]

clarity

proposed method

CLA

EMP

NEG

NEG

NEG

NEG

In particular, please mathematically formulate each proposed technique in Section 4.[Section-NEU], [EMP-NEU]

Section

EMP

NEU

NEU

* Since the proposed formulation is not convincing due to the above reasons and experimental evaluation is not thorough, the originality is not high.[originality-NEG], [NOV-NEU]

originality

NOV

NEG

NEU

Minor comments: - P.1, L.5 in the third paragraph: architexture -> architecture[null], [PNF-NEG]

PNF

NEG

Although the paper has been improved, I keep my rating due to the insufficient experimental evaluation.[rating-NEU, experimental evaluation-NEG], [REC-NEU, EMP-NEG]

rating

experimental evaluation

REC

EMP

NEU

NEG

NEU

NEG

The idea has some novelty and the results on several tasks attempting to prove its effectiveness against systems that handle named entities in a static way.[idea-POS, results-POS], [NOV-POS]

idea

results

NOV

POS

POS

POS

One thing I hope the author could provide more clarification is the use of NER.[null], [EMP-NEU]

EMP

NEU

For example, the experimental result on structured QA task (section 3.1), where it states that the performance different between models of With-NE-Table and W/O-NE-Table is positioned on the OOV NEs not present in the training subset.[experimental result-NEG], [EMP-NEG]

experimental result

EMP

NEG

NEG

To my understanding, because of the presence of the NER in the With-NE-Table model, you could directly do update to the NE embeddings and query from the DB using a combination of embedding and the NE words (as the paper does), whereas the W/O-NE-Table model cannot because of lack of the NER.[null], [EMP-NEU]

EMP

NEU

This seems to prove that an NER is useful for tasks where DB queries are needed, rather than that the dynamic NE-Table construction is useful.[null], [EMP-NEU]

EMP

NEU

You could use an NER for W/O-NE-Table and update the NE embeddings, and it should be as good as With-NE-Table model (and fairer to compare with too).[null], [CMP-NEU, EMP-NEU]

CMP

EMP

NEU

NEU

That said, overall the paper is a nice contribution to dialogue and QA system research by pointing out a simple way of handling named entities by dynamically updating their embeddings.[contribution-POS], [IMP-POS]

contribution

IMP

POS

POS

It would be better if the paper could point out the importance of NER for user utterances, and the fact that using the knowledge of which words are NEs in dialogue models could help in tasks where DB queries are necessary.[null], [EMP-NEU]

EMP

NEU

Although I found the results useful and potentially promising,[results-POS], [EMP-POS]

results

EMP

POS

POS

I did not find much insight in this paper.[insight-NEU], [EMP-NEU]

insight

EMP

NEU

NEU

It was not clear to me why scatter (the way it is defined in the paper) would be a useful performance proxy anywhere but the first classification layer.[null], [EMP-NEG]

EMP

NEG

Once the signals from different windows are intermixed, how do you even define the windows?[null], [EMP-NEU]

EMP

NEU

Minor Second line of Section 2.1: "lesser" -> less or fewer [Second line-NEU, Section-NEU], [PNF-NEU]

Second line

Section

PNF

NEU

NEU

NEU

What I like about the approach is the investigation of the interplay between unsupervised and hierarchical supervised learning in a biological context.[approach-POS], [EMP-POS]

approach

EMP

POS

POS

I agree with the authors that an integrated view of self-organization and learning across layers is presumably required to better understand biological learning.[null], [EMP-POS]

EMP

POS

The general methodology also makes sense to me.[methodology-POS], [EMP-POS]

methodology

EMP

POS

POS

However, I do have concerns including two major concerns: (A) delimitation of results from earlier work; (B) numerical results (especially Tab. 1).[results-NEG, earlier work-NEU], [CMP-NEG]

results

earlier work

CMP

NEG

NEU

NEG

(A) The paper derives the main update equation of W which combines self-organization and label-sensitive learning - Eqn. 15.[paper-NEU, Eqn-NEU], [CMP-NEU]

paper

Eqn

CMP

NEU

NEU

NEU

The paper also states (Secs. 1 and 2) that the the network studied here is based on Hartono et al, 2015, with the main difference of the sigmoidal ouput layer being replaced by a softmax layer.[paper-NEU], [CMP-NEU]

paper

CMP

NEU

NEU

What is missing is a discussion of the differences regarding the later numerical experiments, and a clear delimitation to Hartono et al., 2015, when Eqn. 15 is discussed.[discussion-NEG, numerical experiments-NEU], [SUB-NEG]

discussion

numerical experiments

SUB

NEG

NEU

NEG

What is the major structural difference to their Eqn. 13 which is discussed along very similar lines as Eqn. 15 of this paper.[Eqn-NEU], [CMP-NEU]

Eqn

CMP

NEU

NEU

(B) A further difference to Hartono et al, 2015, are comparisons with multi-layer networks, and the presentation and discussion of this comparison is my strongest concern.[presentation-NEG, discussion-NEG, comparison-NEG], [CMP-NEG, PNF-NEG]

presentation

discussion

comparison

CMP

PNF

NEG

NEG

NEG

NEG

NEG