Samsung AI research team develops compact yet high-performance model

In AI research, progress is often equated with larger models. However, a small team at Samsung ‘s Montreal AI Lab (SAL) is showing promising results with a different approach. Their new Tiny Recursive Model (TRM) overturns the assumption that performance is proportional to the number of parameters . Despite having only 7 million parameters, the model demonstrates inference capabilities that are comparable to or even superior to systems thousands of times larger.

In an open-source release on GitHub and a corresponding arXiv paper, the SAIL Lab describes the design of an AI model with a core recursion process, training the network to improve its answers over time.  A Recursive Approach to Model Building Instead of building massive networks, TRM uses recursion. Recursion is like repeatedly asking, “Is my answer good? If not, can I make it better?” The model produces an answer once, then goes back and refines it.

Also Read: CTCSP Corporation unveils “Digital Workplace for AI” solution

This process is repeated several times until the model is satisfied. In the TRM paper, the authors state that they “recursively refine the latent and output states without assuming convergence.” This means that they don’t force the model to settle on a single, fixed answer early on. Furthermore, they use deep supervision, which provides feedback at multiple steps, not just at the end, to aid learning. They also use adaptive halting, so the model decides when to stop refining, rather than running it endlessly. “Through recursion, models with only a few parameters can achieve surprisingly strong performance on inference tasks,” Samsung researcher Alexia Jolicoeur-Martineau and her colleagues write in a paper published in a GitHub repository and on arXiv.

SOURCE: Yahoo