Yann's Pages
Blog
Archives
Tags
Search
Reviews For Paper
ACML 2015
Asian Machine Learning Conference 2015
Nov 20-22, 2015, Hong Kong, China
Reviews For Paper
Paper ID
79
Title
Curriculum Learning of Bayesian Network Structures
Masked Reviewer ID:
Assigned_Reviewer_4
Review:
Question
Relevance to ACML
Good
Impact of Results
Average
Application Novelty
Average
Technical Novelty
Good
Presentation Quality
Good
Reproducibility
Average
Overall Rating
Accept
Confidence in Review
Marginal
Detailed Comments
This paper has proposed a heuristic algorithm exploiting the idea of curriculum learning to learn the Bayesian network structure. Theoretical analysis was given that the intermediate Bayesian structures are gradually approaching the target structure. Experiments were designed to compare the proposed algorithm and the state of the art -MMHC
This paper is well-written. Some minor suggestions are :
1. Page1 the last paragraph, PC and IC are the first time to use. Their full name should be given.
Please check the similar problems in the whole paper.
2. Eq.(7), the penalty term is a good solution, however, the author did not give an explanation why this specified form (Eq. 7) has been use rather than other?
Masked Reviewer ID:
Assigned_Reviewer_5
Review:
Question
Relevance to ACML
Excellent
Impact of Results
Good
Application Novelty
Good
Technical Novelty
Good
Presentation Quality
Excellent
Reproducibility
Good
Overall Rating
Accept
Confidence in Review
Average
Detailed Comments
The main concern of this paper is the structure learning of Bayesian Networks. The authors introduced curriculum
learning strategy to this problem and proposed a learning algorithm in which greedy hill climbing technique was
used. The behavior of the learning algorithm was theoretically well investigated. Some numerical experiments shows a
convincing results.
The theorems in section 4 that guarantee a monotonicity of the estimation accuracy in the learning process are
interesting. Comparison to some of the different statistical approaches other than MMHC would strengthen the paper.
Masked Reviewer ID:
Assigned_Reviewer_6
Review:
Question
Relevance to ACML
Good
Impact of Results
Average
Application Novelty
Average
Technical Novelty
Average
Presentation Quality
Good
Reproducibility
Average
Overall Rating
Neutral
Confidence in Review
Average
Detailed Comments
- This paper concentrates on constructing the topology of a BN given a set of random variables. It leverages curriculum learning with score-based approach to drive the search algorithm. The theory is provided to state that the Hamming and distribution distances of an intermediate BN, says G_i, to G_k is further than those of G_j to Gk provided i \leq j \leq k. This theoretical analysis somehow indicates that the intermediate states are gradually expanded and become more complicated. However, it cannot guarantee that the latter BNs gradually converge to the ground-truth BN.
- The paper is overall well-organized and well-presented. Nonetheless, some notations, e.g. the training set, domain for X_(i), and data segments should be explicitly declared and defined for comprehensibility.
The idea of conjoining the curriculum learning with score-based approach is nice and motivated. However, some heuristics that are difficult to verify are used in the paper, e.g. fixing X’(i) to different values, our learning target is actually the same DAG structure G_i but with different parameters (CPDs), use of all the training samples in learning Gi at each stage by revising the scoring function to take into account multiple versions of parameters. Apart from this, the role of curriculum learning is not much appealing. Actually, the proposed algorithm is more similar to part-by-part search than curriculum-based.
Please enable JavaScript to view the
comments powered by Disqus.
0
%