Supervisor: Wilfried Gansterer, Torsten Möller
The main objective is to design and implement a good and efficient way of visually investigating the resilience of deep neural networks against silent data corruption (bit flips) based on given empirical measurements. There are many possible causes for such faults (e.g., cosmic radiation, increasing density in chips, lower voltage which implies lower signal charge, etc.), and their "incidence" is expected to increase with current trends in chip architecture.
Starting point for the project is a given example data set which contains information about the relationship between single bit flips across various locations of a certain neural network (which layer, which neuron, which weight, which position within the floating-point representation of a real number, etc.) and the resulting accuracy of the network.
The task is to develop a tool which supports answering various questions about the influence of a bit flip on the resulting accuracy.
Examples for interesting questions are the following:
- (empirical) distribution of the influence of a bit flip on the resulting accuracy over the positions in the floating-point representation
- (empirical) distribution of the influence of a bit flip on the resulting accuracy over the layers in the network architecture
- (empirical) distribution of the influence of a bit flip on the resulting accuracy over the weights in a given layer in the network architecture
In order to answer these questions, an iterative design process is required to
- start with a requirement analysis (task & data analysis)
- low-fi prototypes
- high-fi prototypes
- refinement
- constant evaluation of the visual analysis tool.
The data set, the problem setting and the details of the requirements are provided by Prof. Gansterer, the supervision in visual analysis aspects is provided by Prof. Möller.
Prerequisites: VIS, FDA
Contact: Wilfried Gansterer, Torsten Möller