2017 Workshop on Visualization for Deep Learning
Image credit: Deep Dream Generator, and Google Inceptionism
Deep networks have had profound impact across machine learning research and in many application areas. DNNs are complex to design and train. They are non-linear systems that almost always have many local optima and are often sensitive to training parameter settings and initial state. Systematic optimization of structure and hyperparameters is possible e.g. with Bayesian optimization, but hampered by the expense of training each design on realistic datasets. Exploration is still ongoing for best design principles. We argue that visualization can play an essential role in understanding DNNs and in developing new design principles. With rich tools for visual exploration of networks during training and inference, one should be able to form closer ties between theory and practice: validating expected behaviors, and exposing the unexpected which can lead to new insights.
The workshop has been held last year. This year, with the rise of generative modeling and reinforcement learning, more interesting directions like understanding and visualization of generative models, visual explanation for driving policy could be explored as well.
Here is a list of reference papers related to this topics in the past. Let us know if you can recommend more!
Early July, 2017
To be annouced
Workshop: Aug 10, 2017 at Sydney, Australia, ICML 2017
Location: C4.3 International Convention Centre
Schedule is available!
A Survey for the panel discussion
Dhruv Batra, Georgia Tech and Facebook AI Research Visual Explanations from Deep Networks
Pierre Andrews, Facebook ActiVis: Visual Exploration of Industry-Scale Deep Neural Network Models
Phillip Isola, UC Berkeley Becoming friends with every pixel
Cinjon Resnick, Google Understanding Generative Models in Google Brain Magenta
Bolei Zhou, MIT Quantifying the Interpretability of Deep Visual Representations
Nathan Hodas, Pacific Northwest National Lab Self-supervised attention for Deep Learning
John Canny (UC Berkeley, firstname.lastname@example.org)
Polo Chau (Georgia Tech, email@example.com)
Xiangmin Fan (Chinese Academy of Sciences, firstname.lastname@example.org)
Biye Jiang (UC Berkeley, email@example.com)
Jun-Yan Zhu (UC Berkeley, firstname.lastname@example.org)
Jeff Clune, University of Wyoming
Wojciech Samek, Heinrich Hertz Institute
Yingnian Wu, University of California, Los Angeles
Jason Yosinski, Uber AI Labs
Devi Parikh, Georgia Tech
Dhruv Batra, Georgia Tech
Thomas Brox, University of Freiburg
Anh Nguyen, Auburn University
Antonio Torralba, Massachusetts Institute of Technology
Phillip Isola, UC Berkeley
Bolei Zhou, MIT
Likely topics for the workshop include, but are not limited to:
• Directly visualizing the activations and parameters in intuitive aggregates
• Visualizing weights as features
• Visualizing gradient aggregates during training
• Improving interpretability of networks:
• Localizing “responsbility” in the network for particular outputs
• Sensitivity/stability of network behavior
• Visualizing loss function geometry and the trajectory of the gradient descent process
• Visual representation of the input-output mapping of the network
• Visualizing alternative structures and their performance
• Monitoring/debugging the training process, i.e to detect saddle points or local optima, saturation units
• Visualizing distributed training methods across a cluster
• Using animation in network visualization
• Interactive visualizations for exploration or parameter tuning
• Software architectures for effective visualization
• Visualization and interaction user interfaces
• Understanding and visualization of generative models
• Visual explanation for driving policy
Call for papers:
We accept many kinds of papers, such as (and not limited to): Extended abstract, short papers, work-in-progress papers and demo papers. You can also submit your published work as 2-page summarization.
We encourage submissions of relevant work that has been previously published, or is to be presented at the main conference. The accepted papers will be posted on the workshop website and will not appear in the official ICML proceedings.
Submissions must be in PDF, written in English, no more than 8 pages long — shorter papers are welcome — and formatted according to the standard ICML 2017 guideline. Submissions do not have to be anonymized.
For accepted papers, at least one author must attend the workshop to present the work.
Submission website: https://cmt3.research.microsoft.com/VDL2017