Vlado Menkovski bio photo

Vlado Menkovski

Assistant Professor at Eindhoven University of Technology.

Email Google Scholar

n recent years, deep learning became fundamentally disruptive to conventional machine-learning practices and traditional computer vision techniques alike. Specifically, machine-learning practices for high-dimensional data with weak correlations between target variables. The success of deep learning is largely caused by the availability of more data and more powerful computational engines like Graphic Processing Units. This turn of events has led to the recent introduction of affordable deep-learning-based solutions for the business world. However, despite their capability to learn efficient and powerful representations, the broad adoption of deep learning models is hindered because they are difficult to interpret. This can make it challenging to validate applied deep learning in an enterprise environment. This research project presents a comprehensive look at how a classification model based on convolutional neural networks for an uncommon dataset can be developed for the business world, in a comprehensive, transparent and efficient manner.

The primary goal of this thesis is to present efficient ways to solve and optimize a classification problem from a producer of protein microarrays (or protein chips). Several case studies from prior research have verified that deep-learning-based solutions outperform traditional computer vision solutions for this type of problem. Moreover, open-source software libraries such as Tensorflow have simplified the development of complex high-level machine intelligence systems. This sequence of events forms the foundation of this work, for which various state-of-the-art architectures of deep-learning models have been implemented with Tensorflow. A pre-trained and complex deep neural network has been used as a baseline model to compare with a custom network design that was specifically trained on a real-world dataset of the protein-chip producer. The comparison includes an empirical evaluation of their architectures based on a preprocessed and augmented confidential data set.

This work also provides an extensive analysis and comparison of several visualization techniques. Such techniques form an essential part of the proposed model implementation strategy for business cases with similar business project constraints. Visualizations can convey the correctness of a convolutional neural network to an audience that may lack extensive knowledge on deep learning. The techniques suggested in this research helps untrained users discern a well-built convolutional neural network from an inadequate one, even when both networks make an identical prediction. Based on the reasoning from Kindermans et al in “PatternNet and PatternAttribution - Improving the interpretability of neural networks”, Gradient Class Activation Mapping (Grad-CAM) and Guided Grad-CAM have been confirmed as valid visualizations. However, what remains to be done is to show an empirical comparison between these techniques and novel explanation methods, such as PatternNet and PatternAttribution, to determine their ranking compared to the state of the art.

The implementation of the proof-of-concept is provided to CQM as proprietary software for further research and development.