About the research

The paper MaCP: Minimal yet Mighty Adaptation via Hierarchical Cosine Projection introduces a smart way to help large AI models learn faster and more efficiently. Instead of adjusting thousands of parameters every time – a process that is slow and resource-heavy – the new method, MaCP, uses ‘cosine projection’. In simple terms, it translates adjustments into a kind of vibration pattern (like sound waves) and keeps only the most important ones. This makes the model faster, lighter, and smarter.

Why does it matter?

• Efficiency: MaCP shows that advanced AI can achieve strong results without needing vast amounts of memory and computing power.
• Versatility: While tested in language tasks, the method also works for images and video – for example in visual recognition or video analysis.
• Impact: Smarter, more efficient AI benefits everyone. It can make apps respond quicker, improve translation tools, reduce energy use, and make technology more accessible.

Publication details

The authors of this publication are UvA researchers Yixian Shen, Qi Bi, Jia-Hong Huang, Hongyi Zhu, Andy D. Pimentel, and Anuj Pathania.