Self-Organizing Nebulous Growths for Robust and Incremental Data Visualization

Visualization of Incrementally Learned Projection Trajectories for Longitudinal Data

“…The pseudo code for the incremental learning algorithm is as below. The functionality of SONG [14] algorithm can be summarised in three primary iterative steps. 1) Vector Quantization, 2) Self-Organization and 3) Dimensionality Reduction.…”

“…Nonlinear dimensionality reduction techniques such as t-Distributed Stochastic Nonlinear Embedding (t-SNE)[12] and Uniform Manifold Approximation and Projection (UMAP)[13] are designed as static dimensionality reduction techniques, often failing to insert new data points (an increment) to an already learned model without distorting the structure of the existing representation[14]. This makes it difficult to relate visualizations generated at different timepoints of the same experiment, particularly in longitudinal studies (Supplementary Fig.1).…”

“…This makes it difficult to relate visualizations generated at different timepoints of the same experiment, particularly in longitudinal studies (Supplementary Fig.1). Recently, a new nonlinear dimensionality reduction method, Self-Organizing Nebulous Growths (SONG)[14] was proposed that showed superior performance in incremental dimensionality reduction by preserving the topology of previous increments when adapting visualizations to new data. Therefore, we used SONG as the dimensionality reduction technique in our proposed incremental pipeline.…”

Interpretability and accessibility of machine learning in selected food processing, agriculture and health applications

“…Most ML research on increasing interpretability of ML systems on the Y-axis and accessibility through increased AI model design automation on the X-axis (Figure 1) lies along or close to the axes. Research into interpretable neural networks designed with minimal expert intervention (FAIR AI) will increasingly close a significant knowledge gap, also informed by relevant studies for example, ML with continuous and life-long learning capability (Senanayake, Wang, Naik, & Halgamuge, 2021). This paper is organised with an example of scientific interpretation capability of ML models using differential equations followed by relevant applications of ML in three areas of importance namely Food Processing, Agriculture and Health: Examples of uninterpretable ML models applied to food drying, partially interpretable Convolutional Neural Networks (CNN) and XAI in plant disease detection including in rice cultivation, safety of taking multiple pharmaceutical 4 drugs and reuse of existing drugs for new diseases using semi-automated unsupervised ML model construction and shedding some light into yet largely unexplored world of microbes including viruses using semi-automated ML model construction.…”

On the solidification of the manifold of the t-distributed stochastic neighbour embedding for condition classification of machine tools

“…The parametric mapping of the local structure has been learnt by parametric t-SNE in the latent space [35]. The self-organizing nebulous growths have been adopted to support incremental data for t-SNE, which adds new data in the previous data distribution [36]. However it has the disadvantage that the local linear approximation ignores the global distribution of data, while the global linear approximation might lose detailed information.…”