“Recently, we’ve been witnessing an explosive increase in machine learning applications in modern life,” states Professor Manya Ghobadi of MIT CSAIL. Ghobadi, who holds a PhD in computer science from the University of Toronto , and has worked at Google and Microsoft Research before joining MIT in 2018.
Her current research aims to rethink the basic paradigm for scaling distributed machine learning training workloads. Her team at MIT is building works to build a data center architecture s that can instantaneously train machine learning workloads. Ghobadi’s project reduces the amount of time needed to train machine learning models by distributing training tasks across data centers. As soon as a subtask is completed by a center, the model being trained can be quickly updated across the entire data center. The network therefore does not waste any time waiting for models to be delivered betw een centers.
According to Ghobadi, this project benefits both consumers and the environment because efficiently utilizing data center resources to train machine learning models will enable users to develop machine learning applications faster, preserving tremendous amounts of energy.
This high performance cloud infrastructure could also save a considerable amount of money for businesses.
“It costs about a billion dollars to build a data center and several million dollars per month to keep it operational,” Ghobadi said. “My research is directly involved with improving the performance of these data centers and hence bringing down the cost of online services.”
“With hundreds of data centers around the world, there’s an enormous amount of money and resources th at are concentrated behind the scenes of applications we use every day , ” says Ghobadi. The low cost and efficiency of the cloud architecture she is designing thus “opens up new possibilities of designing models that we haven’t been able to train yet, and i t will realize the true potential of machine learning in modern life.”