The PIXL lunch meets every Monday during the semester at noon in
room 402 of the Computer Science building. To get on the mailing
list to receive announcements, sign up for the "pixl-talks" list at
Monday, March 30, 2020
Monday, April 06, 2020
Monday, April 13, 2020
Monday, April 20, 2020
Monday, April 27, 2020
Monday, May 04, 2020
Monday, May 11, 2020
Monday, May 18, 2020
Monday, February 10, 2020
Monday, February 17, 2020
Data + Art: better science communication
What is Data Art? Can a spreadsheet be transformed into a beautiful, emotional visualization able to communicate knowledge/information more deeply than either the dry spreadsheet, or the traditional charts we use on a daily basis? What additional insight can these complex visualizations provide? Inspired by data visualization and creative coding, we will discuss their main differences as we dwell on interesting (and entertaining) data science research projects such as biodiversity, the strength of collaboration, the Montreux Jazz Festival, Wikipedia and the Star Wars expanded universe. This seminar will show that Data Art can be used as a new communication medium to tell impactful insightful stories and analyses by connecting scientific rigor with creativity.
Kirell Benzi is a data artist, speaker and data visualization lecturer. He holds a Ph.D. in Data Science from EFPL (Ecole Polytechnique Fédérale de Lausanne). His unique work, mixing data visualization and abstract aesthetics, has been shown in museums, newspapers, magazines and on over 100 websites in 10 languages. In 2018, he gave a keynote at a TEDx symposium in Annecy, France on the combination of data and art. He regularly tours the world to inspire people to gain more data literacy by showing the positive outcomes of technology for our society using art.
Monday, March 23, 2020
Encoder-Free 3D Deep Learning for Shape Recognition and Registration
Yi Fang (Assistant Professor of NYU Abu Dhabi and NYU Tandon)
With the availability of both large 3D datasets and unprecedented computational power, researchers have been shifting their focus on applying deep learning to address the challenges in specific tasks such as 3D classification, registration, recognition, and correspondence. The deep learning model often takes input as a grid structured input to effectively exploit discrete convolutions on the data as its fundamental building blocks. However, the irregular Non-Euclidean 3D data representation poses a great challenge for directly applying standard convolutional neural networks to 3D applications such as object recognition from 3D point clouds, 3D shape registration and matching, 3D localization and mapping and so on. In this talk, to address the challenges of learning with irregular 3D data representation, I will discuss our labs recent efforts in the development of an encoder-free design of deep neural network architecture, which we apply to 3D deep learning for shape recognition and registration. The mainstream 3D deep learning efforts require explicitly designed encoders to extract deep shape features and/or spatial-temporal correlation features from irregular 3D data representations. By contrast, we acknowledge the challenges in designing an explicit form of an encoder to extract deep features from unstructured 3D data. As a result, we propose our novel approaches to work around this issue by implicitly extracting the deep features towards various 3D tasks. Our key novelty is that we present a novel unified concept of task-specific latent code (TSLC). The TSLC can take on different forms depending on the nature of the task. It represents a 3D shape descriptor for shape recognition, a 3D shape spatial correlation tensor for shape alignment, or a spatial-temporal descriptor for 3D group registration. The TSLC captures the geometric information from unstructured 3D data essential to each task and is used as input to a task-specific decoder to produce the desired output. Particularly, our approach starts with a randomly initialized TSLC. Next, at training time, we jointly optimize the latent shape code and update the neural network decoders weights towards the minimization of a task-specific loss, while at inferencing time we hold the decoders weights fixed and only optimize the TSLC. Our novel encoder-free approach brings forth two unique advantages: 1) it avoids the inclusion of an explicit 3D feature encoder for irregular 3D data representation. 2) It enhances the flexibility of feature learning for unseen data. The new design centers around the combination of optimization and learning, enabling further fine-tuning on the test data for better generalization abilities. By contrast, the conventional neural network does not have the flexibility in fine-tuning at the testing phase. We conducted experiments on a variety of tasks, including the unsupervised learning of 3D registration, 3D correspondence, and 3D recognition. Qualitative and quantitative comparisons on these experiments demonstrate that our proposed method achieves superior performance over existing methods.