Submitted by
Assigned_Reviewer_1
Q1: Comments to author(s).
First provide a summary of the paper, and then address the following
criteria: Quality, clarity, originality and significance. (For detailed
reviewing guidelines, see
http://nips.cc/PaperInformation/ReviewerInstructions)
This paper presents a fast ICA algorithm that works
best under Gaussian noise. This is demonstrated with components simulated
from different univariate distributions and variable Gaussian noise.
The writing is clear. The paper is incremental in the sense that
it builds on ideas from (Belkin et. al, 2013) but focuses on speeding up
and improving their cumulant-based approach. This is achieved via
1) a Hessian expansion of the cumulant-tensor-based
quasi-orthogonalization. 2) gradient-based iterations that preserve
quasi-orthogonalization of the latent factors (noised case) as well as
whitening in the noiseless case.
Significance: In this manner
the authors manage to deliver a noise-invariant estimator of the mixing
matrix A and sources s, while improving significantly on the
dimensionality complexity.
Other notes: - As you point out,
your method weaknesses come out (inferior, unstable and slower) in the
small data regime (less than 1000), due to
quasi-orthogonalization. Q2: Please summarize your review
in 1-2 sentences
The authors present a new robust ICA cumulant-based
approach. Results are significant under Gaussian noise-corrupted data and
would be of interest to the ICA community.
I've read the author's
rebuttal. Submitted by
Assigned_Reviewer_4
Q1: Comments to author(s).
First provide a summary of the paper, and then address the following
criteria: Quality, clarity, originality and significance. (For detailed
reviewing guidelines, see
http://nips.cc/PaperInformation/ReviewerInstructions)
This paper proposes a cumulant based independent
component analysis (ICA) algorithm for source separation in the presence
of additive Gaussian noise. The algorithm is somewhat incremental building
upon Refs [2] and [3], but appear technically correct with experimental
results confirming the claims made. The algorithms used for benchmarking
assume no additive noise but is like InfoMax often quite robust to
addition of noise.
The ICA literature is huge and many algorithms
exist for ICA with additive noise for example quite a lot Bayesian using
deterministic approximative inference as well as MCMC. For algorithms in
the former category from around year 2000 see for example
http://www.mitpressjournals.org/doi/pdf/10.1162/089976601753196003,
http://www.mitpressjournals.org/doi/pdf/10.1162/089976602317319009,
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.37.597 Another
field where the studied model is used all time is in factor analysis with
identifiable source models. It is therefore doubtful that the proposed
algorithm will really fill a void for ICA in the presence of noise.
Q2: Please summarize your review in 1-2 sentences
This paper proposes a cumulant based independent
component analysis algorithm for source separation in the presence of
additive Gaussian noise. The algorithm is somewhat incremental building
upon Refs [2] and [3], but appear technically correct with experimental
results confirming the claims made. Submitted by
Assigned_Reviewer_6
Q1: Comments to author(s).
First provide a summary of the paper, and then address the following
criteria: Quality, clarity, originality and significance. (For detailed
reviewing guidelines, see
http://nips.cc/PaperInformation/ReviewerInstructions)
The authors propose new algorithms for determining
independent components, when the data is corrupted by Gaussian noise. The
results of alternative methods are affected by noise. The algorithm
improves on the runtime of a method based on the 4th cumulant.
The
problem is relevant. Quality of the paper is high. The approach seems
novel and the contributions significant.
The experimental results
are not totally convincing. In synthetic experiments most of the variants
of ICA perform similarly. For a small number of samples, the
quasi-orthogonalization based method performs significantly worse than all
other methods. For a large number of samples, though, the method performs
at least comparable, even in the noiseless setting. A benefit is
demonstrated for a high Gaussian noise setting, if sufficient data is
provided.
It would be interesting how the different ICA methods
compare in various real world applications, where the noise distribution
is unknown a priori. Runtime experiments are provided only for the
method at hand. To set these into context, the runtime of alternative
methods should be included.
Q2: Please summarize
your review in 1-2 sentences
New methods for the relevant problem of recovering
independent components from data corrupted by Gaussian noise are proposed.
General quality of the paper is high, the approach seems novel and the
contributions significant.
Q1:Author
rebuttal: Please respond to any concerns raised in the reviews. There are
no constraints on how you want to argue your case, except for the fact
that your text should be limited to a maximum of 6000 characters. Note
however that reviewers and area chairs are very busy and may not read long
vague rebuttals. It is in your own interest to be concise and to the
point.
We would like to thank the reviewers for their useful
comments. We believe that our approach to noise-invariant ICA is different
from existing methods in the extensive ICA literature. We hope that our
method will be useful in practice.
|