Optimizing Accuracy-Efficiency Tradeoffs in Emerging Neural Workloads