Big data naturally rescaled

Stoop, Ruedi; Kanders, Karlis; Lorimer, Tom; Held, Jenny; Albert, Carlo

doi:10.1016/j.chaos.2016.02.035

Cited by 9 publications

(8 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…6b), where the set of all avalanche vectors spans a 59-dimensional "feature space". To see how the avalanches are organized in this space, we used an unsupervised Hebbian learning clustering (HLC) developed by us [48][49][50][51] . This approach finds clusters of arbitrary shapes, without prior knowledge of the number of clusters or requiring a data dimensionality reduction step that generally distorts and biases the distances between data points 52 .…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

Fingerprints of a second order critical line in developing neural networks

et al. 2020

Self Cite

View full text Add to dashboard Cite

Patterns of biological activity with properties similar to critical states of statistical mechanics have received much attention, as they were mostly seen as indicators of computational optimality. Commonly, a single regime around an isolated critical point is expected. Our experimental data and our network simulations of developing neural cultures indicate the possibility of transitions between different critical regimes. In the latter, the addition of further fundamental neurophysiological principles to the standard neurodynamics branching model generates steeper power laws that have been observed in various experiments. Our analysis exhibits two populations of neurons, each composed of inhibitory and excitatory sites, that have distinct dynamical and topological properties. This generates a line of second order critical points, similar to what is known from the thermodynamics of two-component alloys. An analysis of two major critical regimes found in the experiments suggests that different critical regimes may express distinct computational roles.

show abstract

Section: Discussionmentioning

confidence: 99%

“…Hebbian learning clustering [48][49][50][51] (HLC) recasts the data (in our application: the avalanche vectors) as the nodes on a k-nearest neighbor graph (using Euclidean distance). The distances between data points furnish the initial edge weights of the graph.…”

Section: Methodsmentioning

confidence: 99%

Fingerprints of a second order critical line in developing neural networks

et al. 2020

Self Cite

View full text Add to dashboard Cite

show abstract

“…3) Many clustering algorithms have difficulties in identifying clusters that are highly nonlinear in multiple dimensions. Unsupervised clustering is an active area of research and future advances in clustering methods—especially density-based and neural network-based methods—may alleviate some of these problems [ 36 ]. However, every such algorithm has its own weaknesses and such improvements may come at the cost of other features (e.g.…”

Section: Results and Assessmentmentioning

confidence: 99%

Quantifying cell densities and biovolumes of phytoplankton communities and functional groups using scanning flow cytometry, machine learning and unsupervised clustering

et al. 2018

View full text Add to dashboard Cite

Scanning flow cytometry (SFCM) is characterized by the measurement of time-resolved pulses of fluorescence and scattering, enabling the high-throughput quantification of phytoplankton morphology and pigmentation. Quantifying variation at the single cell and colony level improves our ability to understand dynamics in natural communities. Automated high-frequency monitoring of these communities is presently limited by the absence of repeatable, rapid protocols to analyse SFCM datasets, where images of individual particles are not available. Here we demonstrate a repeatable, semi-automated method to (1) rapidly clean SFCM data from a phytoplankton community by removing signals that do not belong to live phytoplankton cells, (2) classify individual cells into trait clusters that correspond to functional groups, and (3) quantify the biovolumes of individual cells, the total biovolume of the whole community and the total biovolumes of the major functional groups. Our method involves the development of training datasets using lab cultures, the use of an unsupervised clustering algorithm to identify trait clusters, and machine learning tools (random forests) to (1) evaluate variable importance, (2) classify data points, and (3) estimate biovolumes of individual cells. We provide example datasets and R code for our analytical approach that can be adapted for analysis of datasets from other flow cytometers or scanning flow cytometers.

show abstract

“…From simple reasoning, most elements of a cluster will be the result of the same generating process, obtained via slightly changed parameters of the generative process. Such generative processes lead to distributions that strongly differ from Gaussians [3,26]: whenever two or more parameters are involved in a nonlinear generative process, a shrimp-shaped domain collects the items that have similar properties (figure 4).…”

Section: (C) Sampling Model Genericitymentioning

confidence: 99%

“…In the following, we compare and discuss two promising approaches that are fully based on local k-nearest neighbour information and are therefore promising candidates for unbiased clustering: the Phenograph approach (a recently published leading algorithm in the clustering of mass cytometry data with a view to medical application [29]); and a current implementation of our previously described Hebbian learning clustering (HLC) [26,30] . Shrimp-shaped clusters of similar dynamical behaviour (orange, period 1; blue, period 2; white, higher periodic or divergent behaviour) in the (a, b) parameter space of the Hénon map, which is the prototype for all generic properties of nonlinear processes [18].…”

Section: Searching For Minimally Biased Clustersmentioning

confidence: 99%

Clustering: how much bias do we need?

Lorimer

Held

Stoop

2017

Phil. Trans. R. Soc. A.

Self Cite

View full text Add to dashboard Cite

Scientific investigations in medicine and beyond increasingly require observations to be described by more features than can be simultaneously visualized. Simply reducing the dimensionality by projections destroys essential relationships in the data. Similarly, traditional clustering algorithms introduce data bias that prevents detection of natural structures expected from generic nonlinear processes. We examine how these problems can best be addressed, where in particular we focus on two recent clustering approaches, Phenograph and Hebbian learning clustering, applied to synthetic and natural data examples. Our results reveal that already for very basic questions, minimizing clustering bias is essential, but that results can benefit further from biased post-processing.This article is part of the themed issue 'Mathematical methods in medicine: neuroscience, cardiology and pathology'.

show abstract

Big data naturally rescaled

Cited by 9 publications

References 46 publications

Fingerprints of a second order critical line in developing neural networks

Fingerprints of a second order critical line in developing neural networks

Quantifying cell densities and biovolumes of phytoplankton communities and functional groups using scanning flow cytometry, machine learning and unsupervised clustering

Clustering: how much bias do we need?

Contact Info

Product

Resources

About