Past Meetup

Machine Learning #1

This Meetup is past

280 people went

Location image of event venue

Details

Machine Learning Meetup #1 (by ACM Munich (http://munichacm.de/))

Schedule:
18:00 – 18:30: Arrival and registration
18:30 – 18:45: Introduction by the ACM Student Chapter Munich
18:45 – 19:30: Talk by Dr. Tom Sterkenburg
19:30 – 19:35: Pitch on Data Science Lab@LMU
19:35 – 19:50: Break
19:50 – 20:20: Talk by Caner Hazirbas
20:20 – 20:40: Extreme AI Scaling: Breaking the Barriers (Nvidia)
20:40 – 21:30: Pizza and socializing
------------------------------------------------------------------
It’s time for a new Meetup! We’re back with the first Meetup for this year!
This time, we will first explore the epistemological possibilities of Machine Learning and then hear about some current research on Deep Learning for Computer Vision!
We are proud to announce our speakers Dr. Tom Sterkenburg, who has been researching about the theoretical possibilities to induce knowledge from data, and Caner Hazirbaş, who is a PhD student at the chair of Computer Vision and AI at TUM.
--------------------------------------------------------

Dr. Tom Sterkenburg
"Epistemology of Machine Learning, or how to fool your algorithm"

Hume's problem of induction tells us that we cannot give any proper, noncircular reason for the basic procedure that underlies all learning: extrapolating data to more general conclusions. An analogue of sorts to the problem of induction in machine learning is the so-called no-free-lunch theorem, that states that there can be no bias-free, universal learning algorithm. Another way of saying this is that for every conceivable learning algorithm, there exists potential input data on which it will fail miserably. Worse still, such adversarial data can be very easy to fabricate!
"In this talk, I will first discuss an early mathematical argument against the possibility of a universal learning algorithm, due to the philosopher Hilary Putnam. In essence, Putnam employed Gödel's Incompleteness Theorem in a setting of inductive learning. This sounds serious, but, as I will show, this proof really comes down to an extremely simple algorithm that for any input description of another learning algorithm generates an adversarial datastream for this algorithm. I also relate this to Solomonoff's proposed universal learning method (based on Kolmogorov complexity). I then connect the insights gained from this abstract setting to more recent results about adversarial data for neural networks."

Biography
Tom Sterkenburg is a postdoctoral fellow at the Munich Center for Mathematical Philosophy, LMU, where he works on the philosophical foundations of statistics and machine learning. He holds a BSc in Artificial Intelligence, a MSc in Logic, and a MSc in History and Philosophy of Science. In his PhD project, he investigated the theoretical possibility of a universal prediction algorithm – and, sadly, concluded that there can be no such thing.

--------------------------------------------------------
Caner Hazırbaş
In this talk, Caner will give a brief introduction to research problems he has been working on, such as optical flow, semantic scene understanding, image-based camera localisation, depth from focus and more. The focus will be on the major problems/challenges of these research topics and provide an insight on how to address them with deep networks.

Biography
Caner received his Computer Engineering Bachelors Degree at Yildiz Technical University, Istanbul and Informatics Masters Degree at the Technical University, Munich (TU München). During his study, he focused on object recognition and detection in the fields of computer vision and Machine Learning. He is currently pursuing a Ph.D. at the Computer Vision Group, TU Munich under the supervision of Prof. Dr. Daniel Cremers. He is interested in Deep Learning, object detection/recognition, semantic scene understanding, optical flow, camera pose estimation and depth from focus.