E2-215,   12pm

David Reiman, PhD student of Physics, will present on “AstroGANs: Deep Generative Models for Astrophysics and Galaxy Evolution”.

Here is his abstract:

Deep learning has revolutionized big data—from outperforming doctors in skin cancer diagnoses to precisely forecasting earthquake aftershocks. In astronomy, deep discriminative models have been applied with great success to problems like galaxy classification and exoplanet identification. On the other hand, applications of powerful generative models are scarce. Here, we apply generative adversarial networks (GANs), a model composed of two dueling neural networks, to a variety of open problems in galaxy evolution and cosmology, namely: (1) deblending superpositions of distant galaxies to salvage galaxy images captured in the densest regions of the universe by near-future surveys like LSST, (2) super-resolving optical Suprime-Cam galaxy images from the COSMOS field to near Hubble quality to recover useful features for improved study of galaxy morphology and evolution, and (3) inferring the Lyman-alpha emission of high-redshift quasars given their redward spectrum to extract information about the early universe intergalactic medium.

Rob Currie will moderate a policy discussion on AI related to the medical industry:

A turning point occurred last year when the FDA approved the first ‘Fully Autonomous AI Driven Clinical Diagnostic’. The IDx system diagnoses diabetic retinopathy, requires a high school level of education to operate, replaces the ophthalmologist, and itself has an AMA billing code and malpractice insurance. I’ve attached a slide deck from IDx discussing the ~8 year process they went through with the FDA to validate and approve this system. First question from the audience: ‘How do you approve updating the weights based on all the new data collected?’ – ‘We are working on that with the FDA and it’s not clear how without repeating the full clinical trial’. This is at the nexus of building a learning health system and AI. As further background read a recent survey article by Dr. Eric Topol, High-performance medicine: the convergence of human and artificial intelligence (also author of the excellent ‘The Patient Will See You Now’). These are both on the ‘for’ side of the debate, with some important details on the ‘how’. Welcome discussion here or at the meeting on ‘why we shouldn’t’ or why this is ‘too early’.

Postdoctoral Scholar Mohammad Jafari from the group of Professor Marcella Gomez (Applied Math) will present “On the Machine Learning-Based Prediction and Control of Biological Systems”.

Abstract:  Due to the complex nonlinear nature of biological systems, it is challenging to predict their behavior and/or control their response. It is known that no model will be able to capture all the details of the system and large datasets are not always available a priori. Therefore, employing methodologies with no dependency on large-scale datasets or a priori knowledge of
dynamical models is very important.

In this talk, I will present the key concepts behind online Machine Learning based prediction and control of dynamical systems. I will present the concept of online training vs. offline training and their pros and cons. I will share strategies for tuning neural network parameters and present the latest results of applying online ML-based prediction and control techniques to biological systems by the Gomez Lab at UC Santa Cruz.

Professor Adam Smith (Computational Media) will present on “Deep Learning and the Future of Search: Objects, Apps, and Beyond”.    Here is his abstract:

Web search has matured over the past 20 years, but it still leans on the traditions of textual document retrieval. You give me a textual query, and I’ll give you some textual documents that are related to it. Most audio and visual search technologies, just 10 years younger, rely on hand-engineered feature extraction systems to replace the role of words in documents. Recent sweeping advances in perceptual artificial intelligence are now making it possible for search systems to radically extend the space of queries and documents. We can now search by photograph or by voice to find physical objects or moments reachable within interactive media.

In this talk, you will learn about the key concepts behind cutting-edge, cross-modality search engines. You will learn how to train semantic embeddings which map items onto vectors in an abstract space where distances and directions are significant. You will learn strategies for turning piles of data into a useful index that can be efficiently queried. Examples shown will draw on interactive media indexing and retrieval systems produced by the Design Reasoning Lab at UC Santa Cruz which map out the space of interesting content within apps and games by direct interacting with them.

Members of the group will take ~5min to present on their latest activities, an inspiring new paper, or any advance in machine learning that has caught their attention.

We will also discuss goals for SCML in the New Year.

Angus Forbes will discuss recent trends in the field of Creative AI and survey applications of machine learning for computer graphics. He’ll also present an overview of Deep Illumination, a technique to approximate global illumination for real-time rendering using a conditional generative adversarial network (cGAN).

Angus Forbes is an assistant professor in the Computational Media department at UC Santa Cruz, where he directs the UCSC Creative Coding research lab.

More information about his work can be found at https://creativecoding.soe.ucsc.edu or http://angusforbes.com.

12-1pm in E2-215

David Parks will present on “Model compression techniques & CNN model architecture advances”.

Here is his description:

At next week’s SCML meeting I will be presenting primarily on model compression techniques, focused in detail on quantization / binarization, otherwise known as: How to run large neural network models in the real world – on devices with limited compute, memory, and power (such as mobile devices, raspberry pies, low power accelerators, etc).

To introduce the topic, I will also be covering some of the recent advances in model architecture design used in convolutional neural networks which haven’t been discussed previously.