Visualizing the Hidden Activity of Artificial Neural Networks

Research output: Journal contributionsJournal articlesResearchpeer-review

Standard

Visualizing the Hidden Activity of Artificial Neural Networks. / Rauber, Paulo E.; Fadel, Samuel G.; Falcão, Alexandre X. et al.
In: IEEE Transactions on Visualization and Computer Graphics, Vol. 23, No. 1, 7539329, 01.2017, p. 101-110.

Research output: Journal contributionsJournal articlesResearchpeer-review

Harvard

APA

Vancouver

Rauber PE, Fadel SG, Falcão AX, Telea AC. Visualizing the Hidden Activity of Artificial Neural Networks. IEEE Transactions on Visualization and Computer Graphics. 2017 Jan;23(1):101-110. 7539329. doi: 10.1109/TVCG.2016.2598838

Bibtex

@article{65c0ec990e9241a0b69a8ffd4695998f,
title = "Visualizing the Hidden Activity of Artificial Neural Networks",
abstract = "In machine learning, pattern classification assigns high-dimensional vectors (observations) to classes based on generalization from examples. Artificial neural networks currently achieve state-of-the-art results in this task. Although such networks are typically used as black-boxes, they are also widely believed to learn (high-dimensional) higher-level representations of the original observations. In this paper, we propose using dimensionality reduction for two tasks: visualizing the relationships between learned representations of observations, and visualizing the relationships between artificial neurons. Through experiments conducted in three traditional image classification benchmark datasets, we show how visualization can provide highly valuable feedback for network designers. For instance, our discoveries in one of these datasets (SVHN) include the presence of interpretable clusters of learned representations, and the partitioning of artificial neurons into groups with apparently related discriminative roles.",
keywords = "algorithm understanding, Artificial neural networks, dimensionality reduction, Informatics, Business informatics",
author = "Rauber, {Paulo E.} and Fadel, {Samuel G.} and Falc{\~a}o, {Alexandre X.} and Telea, {Alexandru C.}",
year = "2017",
month = jan,
doi = "10.1109/TVCG.2016.2598838",
language = "English",
volume = "23",
pages = "101--110",
journal = "IEEE Transactions on Visualization and Computer Graphics",
issn = "1077-2626",
publisher = "IEEE - Institute of Electrical and Electronics Engineers Inc.",
number = "1",

}

RIS

TY - JOUR

T1 - Visualizing the Hidden Activity of Artificial Neural Networks

AU - Rauber, Paulo E.

AU - Fadel, Samuel G.

AU - Falcão, Alexandre X.

AU - Telea, Alexandru C.

PY - 2017/1

Y1 - 2017/1

N2 - In machine learning, pattern classification assigns high-dimensional vectors (observations) to classes based on generalization from examples. Artificial neural networks currently achieve state-of-the-art results in this task. Although such networks are typically used as black-boxes, they are also widely believed to learn (high-dimensional) higher-level representations of the original observations. In this paper, we propose using dimensionality reduction for two tasks: visualizing the relationships between learned representations of observations, and visualizing the relationships between artificial neurons. Through experiments conducted in three traditional image classification benchmark datasets, we show how visualization can provide highly valuable feedback for network designers. For instance, our discoveries in one of these datasets (SVHN) include the presence of interpretable clusters of learned representations, and the partitioning of artificial neurons into groups with apparently related discriminative roles.

AB - In machine learning, pattern classification assigns high-dimensional vectors (observations) to classes based on generalization from examples. Artificial neural networks currently achieve state-of-the-art results in this task. Although such networks are typically used as black-boxes, they are also widely believed to learn (high-dimensional) higher-level representations of the original observations. In this paper, we propose using dimensionality reduction for two tasks: visualizing the relationships between learned representations of observations, and visualizing the relationships between artificial neurons. Through experiments conducted in three traditional image classification benchmark datasets, we show how visualization can provide highly valuable feedback for network designers. For instance, our discoveries in one of these datasets (SVHN) include the presence of interpretable clusters of learned representations, and the partitioning of artificial neurons into groups with apparently related discriminative roles.

KW - algorithm understanding

KW - Artificial neural networks

KW - dimensionality reduction

KW - Informatics

KW - Business informatics

UR - http://www.scopus.com/inward/record.url?scp=84998953869&partnerID=8YFLogxK

U2 - 10.1109/TVCG.2016.2598838

DO - 10.1109/TVCG.2016.2598838

M3 - Journal articles

C2 - 27875137

AN - SCOPUS:84998953869

VL - 23

SP - 101

EP - 110

JO - IEEE Transactions on Visualization and Computer Graphics

JF - IEEE Transactions on Visualization and Computer Graphics

SN - 1077-2626

IS - 1

M1 - 7539329

ER -

Recently viewed

Publications

  1. Partitioned beta diversity patterns of plants across sharp and distinct boundaries of quartz habitat islands
  2. What can conservation strategies learn from the ecosystem services approach?
  3. An Adaptive and Optimized Switching Observer for Sensorless Control of an Electromagnetic Valve Actuator in Camless Internal Combustion Engines
  4. ASSESS — automatic self-assessment using linked data
  5. Wavelet functions for rejecting spurious values
  6. Preventive Diagnostics for cardiovascular diseases based on probabilistic methods and description logic
  7. Knowledge-Enhanced Language Models Are Not Bias-Proof
  8. An analytical approach to evaluating monotonic functions of fuzzy numbers
  9. Self-regulation in error management training: emotion control and metacognition as mediators of performance effects
  10. Spaces for challenging experiences, indeterminacy, and experimentation
  11. Commitment to grand challenges in fluid forms of organizing
  12. A structural property of the wavelet packet transform method to localise incoherency of a signal
  13. Quantum Computing and the Analog/Digital Distinction
  14. A Multimethod Latent State-Trait Model for Structurally Different and Interchangeable Methods
  15. Factor structure and measurement invariance of the Students’ Self-report Checklist of Social and Learning Behaviour (SSL)
  16. Mechanism of dynamic recrystallization and evolution of texture in the hot working domains of the processing map for Mg-4Al-2Ba-2Ca Alloy
  17. A comparison of ML, WLSMV and Bayesian methods for multilevel structural equation models in small samples: A simulation study
  18. AGDISTIS - Graph-based disambiguation of named entities using linked data
  19. Species constancy depends on plot size - A problem for vegetation classification and how it can be solved
  20. A Cross-Classified CFA-MTMM Model for Structurally Different and Nonindependent Interchangeable Methods
  21. Using Conjoint Analysis to Elicit Preferences for Occupational Health Services in Small and Microenterprises
  22. Combining multiple investigative approaches to unravel functional responses to global change in the understorey of temperate forests