Zum Inhalt springen

Details

It is time for our next Meetup.
Due to the ongoing Corona crisis we will move into the digital realms as an online meetup.

The format stays the same (2 talks ~40 min in English), except that unfortunately you have to provide your own food and drinks.

Link to the meetup:
https://zoom.us/j/93915726545?pwd=aWZzT1k4SVNuZjNPSUtZQkpWTXArdz09
Password: 678480

Questions for our speakers will be collected in our brand new datageeks slack channel:
https://join.slack.com/t/munich-datageeks/shared_invite/zt-czv02m7b-MDxEj~x0h3HkVwk_bD9s8Q

The Lineup:

Tuan Nguyen - Autonomous Driving for Dummies

In order to teach our cars how to drive, you have to train it with huge amount of data. You let the car drive for hours and receive data as training set. However, do you really need 100% of this data? How to identify video section where “nothing” happens and eliminate them in order to save storage and have more overview of the “good” training set?
I propose an automated process for identifying and classifying driving situations that are related to other vehicles around the car. In the demonstration sensor and video data is extracted during a 30min drive along a highway containing time series.
We then used an algorithm with automatically engineered features to split the video sections into time intervals, each representing a driving situation and finally classified them via hierarchical clustering.

Bio:
I am a curious data enthusiast who love to generate knowledge out of huge amount of data. I then realize the hard truth that in order to do it you have to dig deep, shovel and transform data for 80% of my time. I recognize I like that too and slowly also became a big data engineer. After finishing 2 years distributed master in Germany / Japan / New Zealand in “study and travel” mode, learning new languages and discovering new hobbies, I moved back to Germany and now enjoying my time at NTT Data.

Mina Rezai - Deep Representation Learning from Imbalanced
Data Distribution

In real-world scenarios, data tends to exhibit
imbalanced/Long-tailed distribution. Training of long-tailed datasets
opens new challenges in machine learning where high frequent class tends
to get a concentrated representation of data space. Predicted results by
these networks are not generalized well to new test examples and less
frequent samples. In medical applications, the cost of misclassification
of the less frequent class could be more than the cost of
misclassification of the high frequent class. For example, the risk of
not detecting a tumor could be much higher than referring to a healthy
subject to a doctor. The current talk introduces several deep
learning-based approaches for handling class imbalanced problems for
learning multi-task such as disease classification and semantic
segmentation.

Bio:
Mina Rezaei is Postdoc at the Department of Statistics at the
Ludwig-Maximilians-University Munich (LMU) in the working group
Computational Statistics. Previously, she has been a machine
learning/deep learning researcher at Hasso-Plattner Institute (HPI),
Potsdam University from November 2015 to June 2019 for her Ph.D. studies
under the supervision of Prof. Dr. Christoph Meinel. Mina did her M.Sc.
degree in Artificial Intelligence at the Department of Computer Science,
Shiraz University and Bachelor's degree in Computer Science, Software
Engineering.
During her Ph.D. research, she focused on both technical and
theoretical skills of the machine learning, by developing various
algorithms on deep learning and generative adversarial networks to
handle the class imbalance problem in medical imaging for diseases
diagnosis and segmentation of abnormal tissues as well as body organs
such as voxel-GAN, Recurrent-GAN, Ensemble GANs, Mutual-GAN, and
Bayesian Ensemble GAN.

Das könnte dir auch gefallen