Introduction

As part of my academic research endeavours, I'm undertaking to train myself to analyse research papers with a more methodical and critical eye. 

The particular paper reviewed in this article is "A survey of deep neural network architectures and their applications" by Liu, W. et al. 

The approach I've used to structure my review process is outlined in Research Review Process.

Table of Contents

Research question

What are the prevalent deep learning architectures, and how are they currently being used?

Research aim

This paper draws from research papers to determine what the aggregate applications and types of deep neural networks are. In this respect, it might be considered interpretive research; however, it also contains a large descriptive element because it describes the 4 main types of DNNs are and how they work, in addition to a description of the main applications of DNNs.

There is also an element of predictive research as the authors try to predict the future avenues of research for DNNs. However, on balance, this is likely to be classified as descriptive research or generally a descriptive survey.

Type of research

Qualitative.

The authors read papers and interpret the overall applications for various DNN architectures. There does not however appear to be a explicitly systematic approach. Indeed there is no explicit methodology described in the paper to inform the method used.

Mode of enquiry

Non-Scientific.

There are no quantitative measurements taken, nor experimental research undertaken. Their is no particular strategy/methodology they seem to employ to pick the papers and its possible therefore that they could leave out some important papers or discount the importance of others. The conclusions are qualitative and interpretive which makes it more subjective in nature. There is more risk of subjective opinions in conclusions.

Methodology

Non-empirical, theoretical research based on the gaining a contextual understanding of existing research papers in order to determine what the prevalent architectures and applications are for DNNs.

The main research deliverables are:

1. Determine the different types of widely-used architectures for deep learning.
2. Progress that each architecture has made.
3. Determine the applications of deep learning.

Research Methods

• Survey
• Qualitative narrative review

The research surveys a variety of research papers and describes and identifies the common themes such as the main types and applications of DNNs.

Research techniques & Data analysis

• Thematic review
• Use of secondary material for review and analysis of research papers.

The authors use the data from multiple research papers to establish conclusions about what is required in the future based on what they have found, such as a determination of the main types of architectures and applications found in the papers, and to analyse them to inform their conclusions/predictions. They also use the papers to help explain how the architectures are set up and work, and provide a helpful overview of approaches.

Data

The research papers up until 2017. They appear to be from a wide variety of sources, concentrating on DNNs. See the literature review step for more context on this data.

Information

• RBMs, DBNs, CNNs and AE are the main DNN architectures.
• Speech recognition, pattern recognition and computer vision are the main applications for these DNN architectures.
• DNNs are key applications for processing unlabeled big Data (unsupervised)

Future avenues of research include DNN optimisation, the combination of reinforcement learning and that DNNs use in/complex systems should be more thoroughly researched.

Resource consumption is a concern, especially in low-resource environments such as Mobile.
Stability and consistency of DNN are of concern for the future.

Knowledge

Current state of main DNN architectures, approaches and applications as of 2017

Correlation vs Causation

The aggregation of themes that appear in the surveyed papers suggests their commonality. It's possible however, that a bias in the selection of papers could affect the drawn conclusions, such as which specific architectures are used most often, particularly if a particular architecture was not covered because papers selected did not cover it, or it was biased to a particular field. This does, however, seem unlikely.

Literature review

Referenced papers

Figure 7: Distribution of years of research referenced papers
This paper was published in 2017. See Figure 7  for the chronology of its referenced papers.

Note: only papers from 2000 onwards are included.

Citations

• 2744 citations from Science Direct: https://www.sciencedirect.com/science/article/abs/pii/S0925231216315533
• No representation in ACM or IEEE.

Reasoning method (deduction/induction)

Induction. The research papers are the data, and the outcomes are interpretations/conclusions about that data, such as common themes, observations and future avenues. Also, the papers allowed the authors to group topics and extract meaning from the data, see Figure 8.

Figure 8: Paper’s inductive process

Subjectivity/Objectivity & Threats to validity

Construct Validity

- No obvious flaws

Internal Validity

Research Correctness

Objectivity

1. The paper does cite others' research to convince us that it draws not only from the author's potentially subjective interpretations but that they are objectively based on other research.

Subjectivity/Specificity

1. The papers reviewed aren't systematically selected by any criteria and so the conclusions are limited to the particular papers that the survey/authors decided to use. 
2. The authors don't explicitly mention a methodology they take for constructing the paper, suggesting that a poor/no strategy was taken to select papers.

Research technique

Objectivity:

1. The quality of the papers is likely to be good/objective, but should be supported by the use of other corroborating sources to be more objective about the main DNN architectures and applications.

Subjectivity/Specificity

1. The techniques are appropriate and involve reviewing papers, referring to them in explanations and using them to aggregate and extract common themes from the reviewed papers. The techniques are, however are qualitative and subjective by their nature.

Research techniques vs research question

Objectivity:

1. Surveying research papers to determine common themes and patterns supports the aim to determine the main DNN architectures and their applications.

Subjectivity/Specificity:

1. In order to determine what the prevalent deep learning architectures are and how they are currently being used, the authors have selected a subset of areas/disciplines. This means the prevalent architectures in areas they do not mention are discounted or missing. 

Conclusion vs methods

Objectivity:

1. By aggregating the common architectures in the papers they authors have reviewed, they have objectively determined that from them RBMs, DBNs, CNNs and AE are the main DNN architectures. These are limited to the papers they have reviewed. 
2. Similarly, the papers they chose has lead them to suggest that Speech recognition, pattern recognition and computer vision are the main applications for these DNN architectures.

Subjectivity/Specificity:

1. The conclusions are supported by the papers they reviewed; however, the paper selection criteria are not known.
2. Predictions are subjective. 
3. It's not certain that the authors have reviewed enough papers to suggest all the main architectures and applications. Indeed, this would be difficult to do, but its likely to the conclusions are based on papers reviewed and authors' opinions. 

External Validity

Objectivity:

1. It's likely that of the papers that were reviewed, the proposed main DNN architectures are aggregated/compiled correctly by reviewing the similarity and variances of the architectures they reviewed. This at least means these are the main architectures based on their aggregation results and research.

Subjectivity:

1. Only the main architectures and applications that are representative of the papers reviewed. 

Data Validity

Data subjectivity (specificity/narrowness)

Objectivity:

1. The is a large variety of sources (papers) to derive their conclusions from. 

Subjectivity/Specificity:

1. The research data are research papers, but the methodology of the paper selection criteria is missing, making the selection possibly biased. The data is subject to the selected papers.

Data vs Research Question

Objectivity:

1. The data (research papers) is appropriate and supports the aim to find the main DNN architectures and applications based on the survey of the papers selected in this research.

Subjectivity/Specificity:

1. The main architectures and applications are limited to the papers reviewed. The predictions are subjective.

Summary of general risks to validity

Credibility concerns

Objectivity:

1. A large concentration of recent research over the last 10 years, before the paper being published, suggests methods are indeed state of the art, and therefore, this is likely a good survey. Need to check if there are any 2017 papers, as none appear in the survey, i.e the year the survey was done.
   2744 citations from ScienceDirect suggest that this paper is popular and is likely well-received by researchers.

 Subjectivity/Specificity:

1. The research is 8 years old - applications and architectures are likely to have improved, and some may have become obsolete and depreciated in favour of newer architectures and the applicability to more domains.
2. No representation in ACM or IEEE

Relevance, Contribution, Originality and Novelty

Implications & Contributions

1. Research provides a helpful summary and good balance of technical detail required to understand the main deep neural network architectures.
   Very good for establishing a base understanding of DNNS.
   Helps to indicate the gaps in current applications of DNNs.

Opinion

The research is subject to the papers reviewed; however, the scope of reviewed papers is large, encompassing mostly recent papers, suggesting that it is helpful in suggesting the state of the art near or around the time of publishing.

• Research Review
• Deep Learning