Researchers have found a way to make one neural network act like many, resizing itself in real time to fit whatever a device can afford.
The framework, called Elastoformer, converts an ordinary deep neural network into what its creators call an elastic one. Instead of shipping several separate versions of a model for different battery levels or memory budgets, a method known as the bag-of-models approach, the network switches between multiple internal operating modes at runtime. The team tested it on both Vision Transformers and CNNs, suggesting the trick is not tied to one particular architecture. In their experiments, the elastic version cut computation by as much as 85 percent, latency by half, and memory overhead by 76 percent compared to running fixed models.
That matters because edge devices, think security cameras, drones, or phones doing on-device vision, rarely run under constant conditions. Battery drains, thermal limits kick in, other apps compete for memory. Maintaining a separate model for every scenario burns storage and engineering time, and none of those separate models can react mid-task if conditions shift. A single elastic model sidesteps both problems.
Dynamic inference is not a new pitch. Early-exit and slimmable networks chased similar goals for years with uneven real-world adoption, since paper savings don't always survive contact with actual silicon and scheduling overhead. Elastoformer's numbers so far come from the authors' own benchmarks, and the code is open on GitHub for anyone who wants to check whether it holds up outside a lab.