Dana Drachsler Cohen - SPECIAL GUEST LECTURE
Thursday, 28.12.2017, 14:30
Many software systems rely on data-driven models to make decisions.
Examples include self-driving cars, malware detection and aircraft
collision avoidance detection. Unfortunately, data-driven models
often do not generalize well on unseen examples, despite showing
high accuracy on test sets. This was demonstrated by showing how to
fool these models using adversarial examples. Such adversarial
examples may result in disastrous consequences in safety-critical
systems that rely on these models. It becomes clear that high
accuracy is insufficient in these cases, and exactness is a desired
In this talk, I will discuss a new approach which recovers exactness
in data-driven models. This approach involves interaction with a
user to classify examples, a crucial aspect is minimizing the
number of questions posed to the user. I will then present two
algorithms that guarantee exactness in the setting of program
synthesis from examples. I will also show experimental results that
support the importance of exactness in practice.
Dana Drachsler is an ETH Postdoctoral Fellow in the department of
Computer Science at ETH Zurich.
Her research focuses on applying programming languages techniques to
bring rigor to other areas including deep learning models,
blockchains, and computer networks.
She obtained her PhD from the Computer Science Department at the
Technion in 2017.