THUDM/INFTY

What it solves

INFTY is an optimization engine designed to address core challenges in Continual AI, such as catastrophic forgetting, the stability-plasticity dilemma, and poor generalization. It provides a unified framework to replace fixed optimizer configurations in existing AI pipelines, allowing models to learn new tasks without forgetting previous ones.

How it works

INFTY acts as a plug-and-play wrapper around base optimizers (like SGD). It implements specialized algorithms to manage how models update their parameters across different learning scenarios:

• Generalizability support: Promotes flat loss landscapes to improve adaptation over time.
• BP-Free support: Uses gradient approximation when backpropagation is not possible.
• Multi-objective support: Reduces gradient interference between old and new task objectives by manipulating shared parameters.

It also includes diagnostic tools to visualize optimization behavior, such as loss landscapes, Hessian eigenvalue spectrum density (ESD), and gradient conflict curves.

Who it’s for

It is intended for researchers and developers working on Continual AI, including those focusing on PTM-based continual learning, continual PEFT, continual diffusion models, and continual vision-language models (VLMs).

Highlights

• Broad Compatibility: Supports diverse architectures including ResNet, Transformer, ViT, CLIP, and Diffusion models.
• Lighweight Integration: Plugin-style design allows for easy replacement of optimizers in existing pipelines.
• Diagnostic Suite: Built-in tools for visualizing loss landscapes and quantifying gradient interference.
• Versatile Application: Proven effective in scenarios ranging from typical continual learning to text-to-image diffusion and multi-modal learning.