[Paper Review] Deep Transparent Prediction through Latent Representation Analysis

Dimitrios Kollias, N. Bouas+7 more|arXiv (Cornell University)|Sep 13, 2020

Generative Adversarial Networks and Image Synthesis33 references39 citations

TL;DR

A novel framework extracts latent representations from trained DNNs, clusters them into concise centers for transparent, unified predictions, and extends to multi-source domain adaptation across Parkinson’s, COVID-19 imaging, and retail packaging quality checks.

ABSTRACT

The paper presents a novel deep learning approach, which extracts latent information from trained Deep Neural Networks (DNNs) and derives concise representations that are analyzed in an effective, unified way for prediction purposes. It is well known that DNNs are capable of analyzing complex data; however, they lack transparency in their decision making, in the sense that it is not straightforward to justify their prediction, or to visualize the features on which the decision was based. Moreover, they generally require large amounts of data in order to learn and become able to adapt to different environments. This makes their use difficult in healthcare, where trust and personalization are key issues. Transparency combined with high prediction accuracy are the targeted goals of the proposed approach. It includes both supervised DNN training and unsupervised learning of latent variables extracted from the trained DNNs. Domain Adaptation from multiple sources is also presented as an extension, where the extracted latent variable representations are used to generate predictions in other, non-annotated, environments. Successful application is illustrated through a large experimental study in various fields: prediction of Parkinson's disease from MRI and DaTScans; prediction of COVID-19 and pneumonia from CT scans and X-rays; optical character verification in retail food packaging.

Motivation & Objective

• Motivate the need for transparent, accurate predictions from DNNs in healthcare and other domains.
• Develop a method to extract latent variables from trained DNNs and convert them into concise, interpretable representations.
• Create a unified prediction model by clustering latent representations and using nearest-centroid decision rules.
• Extend the approach to multi-source domain adaptation to enable predictions on non-annotated, diverse environments.
• Demonstrate the framework on Parkinson’s disease prediction from MRI/DaTScans, COVID-19 detection from CT/x-ray images, and optical character verification in food packaging.

Proposed method

• Train a CNN or CNN-RNN to predict disease status from image data.
• Extract high-level outputs from the last fully connected or hidden layer as latent vectors v.
• Apply k-means clustering to V_s to obtain L clusters and centers c(i) forming a concise representation C.
• Classify new samples by nearest cluster center in C, enabling efficient and transparent predictions with corresponding visual/anotation support.
• Extend to a unified model by training a new DNN on V_s representations from additional datasets and merging C with C' to handle catastrophic forgetting.
• In a multi-source DA extension, minimize feature discrepancy (MMD and CORAL) and class discrepancy across N sources, with CAM visualizations for interpretability.

Experimental results

Research questions

• RQ1 How can latent representations extracted from trained DNNs be transformed into concise, interpretable representations for prediction?
• RQ2 Can a unified, transparent predictive model be built by clustering latent features and using nearest-neighbor decisions across multiple datasets?
• RQ3 How effective is domain adaptation from multiple labeled sources for predicting in non-annotated, new environments?
• RQ4 To what extent do the resulting cluster-centered representations provide actionable, interpretable explanations to domain experts?
• RQ5 What is the impact of the proposed framework on real-world medical imaging tasks like Parkinson’s and COVID-19 detection, and on non-medical tasks such as optical character verification?

Key findings

• On Parkinson’s prediction, the CNN-RNN achieved 98% accuracy on the Greek/PPMI DaTscan+MRI dataset, with five clusters constituting C (two NPD, three PD progression centers).
• The unified representation enabled by merging C and C' maintained high predictive performance when applied across datasets (e.g., PD/NPD classification remained accurate).
• For COVID-19 CT detection, a 32-d representation with five clusters yielded F1 = 0.842 overall (0.855 non-COVID, 0.828 COVID), outperforming the baseline EfficientNet-based result.
• Chest X-ray fusion with COVID-19 datasets using the same approach achieved F1 = 0.96 with five clusters, surpassing prior network performance under similar settings.
• In optical character verification for food packaging, multi-source domain adaptation improved accuracy by over 6 percentage points (90.53% vs 84.14%).
• The paper demonstrates transparent predictions by linking cluster centers to representative medical images and expert annotations, enabling trust and annotation-driven validation.