Advancements in ML for ML
Advancements in Machine Learning (ML) have enabled machines to do more than ever before, such as understand natural language, generate images, and more. ML models use programming frameworks like TensorFlow and PyTorch to simplify programming and training. These libraries handle complex operations like matrix multiplication and neural networks, and automatically optimize models for efficient hardware use.
Improving ML Workloads
ML compilers convert user-written programs into instructions for hardware, and using heuristics to optimize performance. But, these optimizations can be suboptimal. Recent research shows that we can use ML to improve the performance of ML programs.
Introducing TpuGraphs
We have released TpuGraphs, a dataset for improving ML compiler cost models. It includes large computational graphs from popular ML programs like ResNet and EfficientNet. This dataset provides the scale needed to explore graph-level prediction tasks on large graphs.
Improving Large Graph Property Prediction
We use a Graph Neural Network (GNN) to scale training for large graphs. Our method, Graph Segment Training (GST), efficiently handles large graphs for prediction tasks.
Improving ML for ML
With the advancements in ML for ML, we can improve the efficiency of ML workloads. This can help make ML programs run faster and more smoothly, benefiting researchers and developers alike.