Happy new academic year!
We will celebrate with our first ‘seminar’. As the standard E2-215 room is needed for an interview next week, we will meet in the CfAO atrium (on the ground floor of this building on Science Hill). We will brainstorm ideas and speakers for the coming year, discuss new papers, and introduce ourselves to one another.
If you have been before, bring a new friend. If you are new, then welcome!
This week we will have 2 presentations from Computational Media:
Speaker 1:
Mahika Dubey, Graduate Student in Computational Media, UCSC Creative Coding Lab
Title:
Data Brushes: Interactive Neural Style Transfer for Data Art
Abstract:
We introduce in-browser applications for the application of data art-based style-transfer brushes onto an image, inviting casual creators and other non-technical users to interact with deep convolutional neural networks to co-create custom artworks. In addition to enabling a novel creative workflow, the process of interactively modifying an image via multiple style transfer neural networks reveals meaningful features encoded within the networks, and provides insight into the effect particular networks have on different images, or even different regions of an image, such as border artifacts and spatial stability or instability. Our data brushes provide new perspectives on the creative use of neural style transfer for data art and enhance intuition for the expressive range of particular style transfer features.
Speaker 2:
Oskar Elek, Postdoctoral Researcher in Computational Media, UCSC Creative Coding Lab
Title:
Learning Patterns in Sample Distributions for Monte Carlo Variance Reduction
Abstract:
This ongoing project investigates the prediction of unknown distributions from stochastic, sparse samples. A relevant problem in many domains (especially forecasting), we address it from the perspective of stochastic Monte Carlo integration in physically based image synthesis (rendering). Because the sample distributions obtained in rendering are complex, chaotic, and do not conform to known statistical distributions, simple estimation methods usually yield results with high amounts of variance. To tackle this issue, we systematically study these sample distributions to understand common patterns and archetypes, and propose to use deep neural networks to learn them. I will present our current results, with an open discussion centered on the main challenge: How can we use the knowledge of characteristic sample patterns to bootstrap the network and get better predictions?
Where/when: E2-215 at 12pm
Presenter: J. Xavier Prochaska (Astronomy & Astrophysics)
Title: Activation Atlas
Abstract: While it remains conventional wisdom that Deep Learning techniques are primarily the result of impenetrable “black boxes”, there is a growing effort to peer into the box. I will describe a few of the initial efforts and then focus on the Activation Atlas built by researchers at Google and Open AI. I have been willing to state out-loud that I find this akin to peering into the brain of the network.
Here is the link to the primary paper: https://distill.pub/2019/activation-atlas/
Pizza will be provided.
E2-215
12-1pm
Ryan Hausen (PhD student, CSE at UCSC)
Title: Morpheus: A Deep Learning Model for Pixel-Level Morphological Classification
Abstract: The majority of astronomy begins with images filled with stars and galaxies of various sizes in arrays that can be hundreds of millions of pixels. From these images, one may study galaxies by their shape yet classifying by morphology can be difficult and often involves an element of subjectivity. Furthermore, doing so with astronomical-scale images is both time and cost-prohibitive if one were to rely on human resources alone. The Computational Astrophysics Research Group at UCSC is leveraging current methods and innovating new techniques in Deep Learning to develop new and more efficient approaches this problem. One such way is Morpheus, a deep learning model that morphologically classifies astronomical images with pixel-level precision. Using Morpheus, astronomers can extract a detailed analysis of the morphological composition galaxies and stars of an image. This kind of information is key to understanding how galaxies form and evolve.
Title: Nanopore Variant Calling using Deep Neural Networks
Kishwar Shafin, Graduate Student in the Computational Genomics Lab
Genomic sequencing of an individual genome produces millions of sequence reads. Once aligned to a reference genome, the reads can be used to identify genetic variations. Variant calling is essential in clinical genomics because genetic variants can be associated with genetic diseases. Next-generation short read sequencing technology is widely used for variant calling, but the short-reads are unable to solve complex regions of the genome. The third-generation long-read sequencing technology produces sequences that can span larger area in the genome which provides a better resolution in the complex areas of the genome. Although the third-generation long read sequencing technology like the Oxford Nanopore has a clear advantage, due to the error rate of the output sequences, the existing variant callers perform poorly. The Computational Genomics Lab (CGL) under the UCSC Genomics Institute is developing deep neural network based modules to enable analysis with noisy long reads. In this presentation, we will discuss various aspects of using deep neural networks to perform variant calling with noisy long read sequences.
Come join the first AAII seminar (formerly SCML group) this
coming Wednesday from 12pm-1pm in E2-215.
This first meeting will be an open discussion of the latest
and greatest in ML and AI. Come with your own thoughts to share.
Free pizza will be provided to the first ~30 attendees.