Motion recognition based on concept learning

Ma, Xiang; Zhao, Tong; Wen, Ruoshi; Wu, Zhaojun; Wang, Qiang

doi:10.1109/i2mtc.2017.7969730

Cited by 4 publications

(8 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To validate the proposed decomposition, we applied it on simulated traces presenting different noise levels. Our algorithm reached accuracies above 90% with respect to the known dimensionality of the simulations at low noise level and still above 80% at high noise level, while an approach not based on space-scale which we took as baseline (Yang et al, 2016;Ma et al, 2017) was strongly affected by noise ( Figure 1A, S2 and S3; details on the simulation of traces, noise levels and algorithm are available in the Methods section).…”

Section: Computation Of the Local 3d Scale Of Neuronal Traces From Thmentioning

confidence: 93%

“…We selected data from Santos et al (2018) hosted in the NeuroMorpho.Org database which describe the expansion of the dendritic arbors Figure 1: Computation of the 3D local scale of neuronal traces from their multiscale dimensionality decomposition. (A) (left) Variation of the intrinsic dimensionality along a portion of 3D curve; (right) evaluation of the decomposition on simulated trajectories and comparison to a baseline method not based on scale-space (Yang et al, 2016;Ma et al, 2017). (B) Schematic presentation of the processing of 3D curves by the nAdder algorithm.…”

Section: Application Of Nadder To Characterize Neuronal Arbors Duringmentioning

confidence: 99%

“…We either take the one corresponding to the largest accuracy (optimal scale, Figure S2) or one at a fixed scale ( Figure S3), r k = 20 µm, small enough to avoid deforming the simulated curve. Our method is then compared with a baseline method (Yang et al, 2016;Ma et al, 2017) that first iteratively assigns each point on the curve as linear/nonlinear by a collinearity criterion, then characterize nonlinear points as planar/nonplanar by a coplanarity criterion. For both methods, the curve was first denoised based on Czesla et al (2018).…”

Section: Evaluation Of Dimension Decomposition On Simulated Curvementioning

confidence: 99%

See 2 more Smart Citations

nAdder: A scale-space approach for the 3D analysis of neuronal traces

Phan

Matho

Beaurepaire

et al. 2020

Preprint

View full text Add to dashboard Cite

The advent of large-scale microscopy along with advances in automated image analysis algorithms 1 is currently revolutionizing neuroscience. These approaches result in rapidly increasing libraries 2 of neuron reconstructions requiring innovative computational methods to draw biological insight 3 from. Here, we propose a framework from differential geometry based on scale-space representation 4 to extract a quantitative structural readout of neurite traces seen as tridimensional (3D) curves 5 within the anatomical space. We define and propose algorithms to compute a multiscale hierarchical 6 decomposition of traced neurites according to their intrinsic dimensionality, from which we deduce 7 a local 3D scale, i.e. the scale in microns at which the curve is fully 3D as opposed to being 8 embedded in a 2D plane or a 1D line. We applied our scale-space analysis to recently published 9 data including zebrafish whole brain traces to demonstrate the importance of the computed local 3D 10 scale for description and comparison at the single arbor levels and as a local spatialized information 11 characterizing axons populations at the whole brain level. The use of this broadly applicable approach 12 is facilitated through an open-source implementation in Python available through GeNePy3D, a 13 quantitative geometry library.14 Keywords Computational neuroanatomy · scale space · quantitative geometry · python 15 1 Introduction 16 Throughout the evolution of the field of neuroscience, neuroanatomy has played a key role via the analysis of neuronal 17 arbor traces, either at single cell or gross projection levels. The NeuroMorpho.Org database (Ascoli, 2006; Ascoli et al., 18 2007), which collects and indexes neuronal tracing data, currently hosts more than one hundred thousand arbors of 19 diverse neurons from various animal species. Thanks to technological advances in microscopy, through block-face 20 electron microscopy (Helmstaedter et al., 2013) and large-scale fluorescence microscopy (Abdeladim et al., 2019), we 21 are now able to image even larger tissue volumes with ever increasing resolution and contrast modalities. Crucially, the 22 coming of age of computer vision through the advent of deep learning is offering ways to automate the extraction of 23 neurite traces (Magliaro et al., 2019), a process that is both tedious and time consuming to do manually. As a result, we 24 are about to experience an exponential increase in the amount of neuronal traces extracted in diverse species, brain 25 regions, developmental stages and experimental conditions to answer key questions across neuroscience (Meinertzhagen, 26 2018). Methods from quantitative and computational geometry will play a major role in handling and analyzing the 27 A scale-space approach for 3D neuronal traces analysis growing body of neuronal reconstruction data, in its full tridimensional (3D) complexity, a requisite to linking neuronal 28 structure with development and function of brain circuits and systems. 29A vast array of geometric algorithmic met...

show abstract

Section: Computation Of the Local 3d Scale Of Neuronal Traces From Thmentioning

confidence: 93%

Section: Application Of Nadder To Characterize Neuronal Arbors Duringmentioning

confidence: 99%

Section: Evaluation Of Dimension Decomposition On Simulated Curvementioning

confidence: 99%

See 1 more Smart Citation

nAdder: A scale-space approach for the 3D analysis of neuronal traces

Phan

Matho

Beaurepaire

et al. 2020

Preprint

View full text Add to dashboard Cite

show abstract

“…To validate the proposed decomposition, we applied it on simulated traces presenting different noise levels. Our algorithm reached accuracies above 90% with respect to the known dimensionality of the simulations at low noise level and still above 80% at high noise level, while an approach not based on space-scale which we took as baseline [37,38] To make use of the local dimensionality decomposition across multiple scales and compute a simpler and more intuitive metric, we define the local 3D scale at a given position as the highest scale at which the trace still remains locally 3D around that position. The trace will then locally transform to 2D or 1D for scales higher than that local 3D scale.…”

Section: Plos Computational Biologymentioning

confidence: 99%

nAdder: A scale-space approach for the 3D analysis of neuronal traces

et al. 2022

View full text Add to dashboard Cite

Tridimensional microscopy and algorithms for automated segmentation and tracing are revolutionizing neuroscience through the generation of growing libraries of neuron reconstructions. Innovative computational methods are needed to analyze these neuronal traces. In particular, means to characterize the geometric properties of traced neurites along their trajectory have been lacking. Here, we propose a local tridimensional (3D) scale metric derived from differential geometry, measuring for each point of a curve the characteristic length where it is fully 3D as opposed to being embedded in a 2D plane or 1D line. The larger this metric is and the more complex the local 3D loops and turns of the curve are. Available through the GeNePy3D open-source Python quantitative geometry library (https://genepy3d.gitlab.io), this approach termed nAdder offers new means of describing and comparing axonal and dendritic arbors. We validate this metric on simulated and real traces. By reanalysing a published zebrafish larva whole brain dataset, we show its ability to characterize different population of commissural axons, distinguish afferent connections to a target region and differentiate portions of axons and dendrites according to their behavior, shedding new light on the stereotypical nature of neurites’ local geometry.

show abstract

“…Its learning and reasoning outperform deep neural network learning. Some motion recognition research [9] is also conducted based on the ideas of concept learning. However, most of the existing works focus on concept learning and pattern recognition in the visual field.…”

Section: Introductionmentioning

confidence: 99%

Bayesian Network Model for Learning Arithmetic Concepts

Lv¹,

Tong²,

Qian³

et al. 2019

Int J Performability Eng

View full text Add to dashboard Cite

An object usually belongs to multiple concepts, but some concepts can be judged directly while other concepts need to be inferred indirectly. To learn some arithmetic concepts from positive integer number sets, we address an arithmetic concept Bayesian network (ACBN) model by taking advantage of Bayesian networks. Specifically, we first give an ACBN model to represent the arithmetic concept knowledge and their direct relationships, and then we design an ACBN model learning algorithm based on domain knowledge. Furthermore, to infer indirectly some arithmetic concepts, we design the learning method of evidence concepts based on the idea of knearest neighbors, and then we propose the inference algorithm of the ACBN model. Finally, the experimental results demonstrate that the ACBN model can effectively learn some daily arithmetic concepts.

show abstract

Motion recognition based on concept learning

Cited by 4 publications

References 20 publications

nAdder: A scale-space approach for the 3D analysis of neuronal traces

nAdder: A scale-space approach for the 3D analysis of neuronal traces

nAdder: A scale-space approach for the 3D analysis of neuronal traces

Bayesian Network Model for Learning Arithmetic Concepts

Contact Info

Product

Resources

About