Binary Neural Networks (BNNs) are compact and efficient by using binary weights instead of real-valued weights. Current BNNs use latent real-valued weights during training, where several training hyper-parameters are inherited from real-valued networks. The interpretation of several of these hyperparameters is based on the magnitude of the real-valued weights. For BNNs, however, the magnitude of binary weights is not meaningful, and thus it is unclear what these hyperparameters actually do. One example is weight-decay, which aims to keep the magnitude of real-valued weights small. Other examples are latent weight initialization, the learning rate, and learning rate decay, which influence the magnitude of the real-valued weights. The magnitude is interpretable for real-valued weights, but loses its meaning for binary weights.
In this paper we offer a new interpretation of these magnitude-based hyperparameters based on higher-order gradient filtering during network optimization. Our analysis makes it possible to understand how magnitude-based hyperparameters influence the training of binary networks which allows for new optimization filters specifically designed for binary neural networks that are independent of their real-valued interpretation. Moreover, our improved understanding reduces the number of hyperparameters, which in turn eases the hyperparameter tuning effort which may lead to better hyperparameter values for improved accuracy.

CodeFeedr is a research project at the software engineering division of the Delft University of Technology in collaboration with the Software Improvement Group. The research focuses on a software infrastructure which serves software practitioners in utilizing data-driven decision making. Currently, frameworks like Apache Flink are capable of high-performance data streaming. However, these frameworks have a lot of overhead in setting up, and adding new streaming queries takes a lot of time. They also have several limitations in combining real-time data with historical data and doing aggregations on streams from multiple sources. The developed product is a plug-in framework on top of Apache Flink, that provides a pipelining system for streaming queries. This product includes abstractions for well-known sources like GitHub, TravisCI and Twitter as well as support for historical data in mongoDB. With this framework the users can spend their efforts on actually writing streaming queries instead of setting up environments, input sources and output destinations. The product also includes orchestration tools for running streaming jobs on a distributed system.