Новый метод позволяет масштабировать AI-модели через экспертные блоки без переобучения

A team of computer vision specialists has developed a technology that allows scaling already-trained AI models to larger versions — by adding specialized expert modules and without launching training from scratch.

What is the essence of the new method?

The new technology breaks the familiar link between model power and the need to train it from scratch. The traditional approach looks like this: if a larger neural network is needed — a new architecture must be designed, a dataset collected, and a full training cycle launched, which can take weeks or months of computation. The method developed by the team allows taking an existing pre-trained model and expanding it by adding specialized expert modules on top of the already accumulated knowledge.

At the core of the idea is the desire to preserve what already works. A pre-trained model has already seen billions of data examples and has formed valuable internal representations of the world. Discarding them for the sake of scaling is wasteful. The new method uses these weights as a stable foundation and builds specialized components on top.

This approach is conceptually close to the Mixture of Experts (MoE) architecture, where different "experts" within the network specialize in different types of data or tasks. The fundamental difference of the new solution is that it allows growing such a structure from a ready-made base, rather than building MoE from scratch.

"The technology allows expanding pre-trained AI models into larger systems with specialized expert modules without training from scratch," the authors describe the development according to

TechXplore data.

Why does this matter for AI developers?

The cost of training frontier neural networks is one of the main barriers in the industry. According to estimates from Epoch AI analysts, the costs of training leading models in 2024–2025 reached tens of millions of dollars. Teams that already have a well-trained base model face a difficult choice: work with the limited version or spend comparable sums on retraining from scratch.

The expert module scaling method offers a third path. Existing models can be developed incrementally: adding new specialized blocks as tasks and requirements grow, without losing the accumulated weights and internal representations of the base network.

• Technology created by a team of computer vision specialists
• Allows expanding pre-trained AI models without full retraining from scratch
• Uses specialized expert modules embedded on top of the ready-made base
• Preserves weights and accumulated representations of the original model during scaling

Where is this applicable?

Computer vision tasks are particularly sensitive to model scale. Medical imaging recognition, industrial quality control, and autonomous driving systems simultaneously require a broad general knowledge base and deep specialization for a specific domain. The expert module expansion method allows resolving this trade-off: taking a powerful general model and building specialized blocks on top for a specific task.

For corporate teams that have already-trained models on proprietary data, this means the ability to upgrade the system without losing corporate specificity and without having to go through the full cycle of data collection and training again.

What this means

If the expert module expansion method proves practical effectiveness, it could change the logic of corporate AI system development: instead of costly restarts every few years — incremental growth, in which each model version becomes the foundation for the next.