Resource-Constrained and Hardware-Accelerated Machine Learning

Abstract - The resource and energy consumption of machine learning is the major topic of the collaborative research center. We are often concerned with the runtime and resource consumption of model training, but little focus is set on the application of trained ML models. However, the continuous application of ML models can quickly outgrow the resources required for its initial training and also inference accuracy. This seminar presents the recent research activities in the A1 project in the context of resource-constrained and hardware-accelerated machine learning. It consists of three parts contributed by Sebastian Buschjaeger (est. 30 min), Christian Hakert (est. 15 min), and Mikail Yayla (est. 15 min).

Talks:

FastInference - Applying Large Models on Small Devices
Speaker: Sebastian Buschjaeger
Abstract: In the first half of my talk I will discuss ensemble pruning and leaf-refinement as approaches to improve the accuracy-resource trade-off of Random Forests. In the second half I will discuss the FastInference tool which combines these optimizations with the execution of models into a single framework.

Gardening Random Forests: Planting, Shaping, BLOwing, Pruning, and Ennobling
Speaker: Christian Hakert
Abstract: While keeping the tree structure untouched, we reshape the memory layout of random forest ensembles. By exploiting architectural properties, as for instance CPU registers, caches or NVM latencies, we multiply the speed for random forest inference without changing their accuracy.

Error Resilient and Efficient BNNs on the Cutting Edge
Speaker: Mikail Yayla
Abstract: BNNs can be optimized for high error resilience. We explore how this can be exploited in the design of efficient hardware for BNNs, by using emerging computing paradigms, such as in-memory and approximate computing.