Skip to yearly menu bar Skip to main content


Workshop

Localized Learning: Decentralized Model Updates via Non-Global Objectives

David I. Inouye · Mengye Ren · Mateusz Malinowski · Michael Eickenberg · Gao Huang · Eugene Belilovsky

Meeting Room 310

Sat 29 Jul, noon PDT

Despite being widely used, global end-to-end learning has several key limitations. It requires centralized computation, making it feasible only on a single device or a carefully synchronized cluster. This restricts its use on unreliable or resource-constrained devices, such as commodity hardware clusters or edge computing networks. As the model size increases, synchronized training across devices will impact all types of parallelism. Global learning also requires a large memory footprint, which is costly and limits the learning capability of single devices. Moreover, end-to-end learning updates have high latency, which may prevent their use in real-time applications such as learning on streaming video. Finally, global backpropagation is thought to be biologically implausible, as biological synapses update in a local and asynchronous manner. To overcome these limitations, this workshop will delve into the fundamentals of localized learning, which is broadly defined as any training method that updates model parts through non-global objectives.

Chat is not available.
Timezone: America/Los_Angeles

Schedule