One of the most important annual programs in the field of Artificial Intelligence this week in Singapore: International conference on teaching representationAs usual, the chip was a major presence at the veteran NVidia conference, which presented more than 70 research papers from his team.
Too: NVIDIA launched NEMO Software Tool to help make enterprises custom AI agents
The papers cover the topics from generating music to 3D-revision videos, robot training tasks and the ability to generate many big language models on a button push.
Not just a chip company
In an interview with ZDNET, NVDia’s Head of Applied Learning Research said, “People often think Nvidia as a chip company, which makes terrible chips, and of course, we are really proud of it.” “But the story that I think the most matters is the most, to make those terrible chips for us, we have to do research in this way, because it teaches us how to make all those systems.”
Too: NVIDIA General AI dominates in benchmark, 2 rival AI Chips Clobball
The papers presented this week, most of which were published in the last one year either on Arxiv preprint server, ranging from pure research to programs that provide immediately usable equipment.
In the former category, for example, A project called Lalamplex Improves the task of generating many big language models from single parents. It is common in the “student” LLM to “distil” a single, very large LLM, inherited the ability of “teacher”, but takes low memory storage.
Nvidia researcher Russian Caqi and his team noticed that the method of distillation could be improved, which they call “elastic preteering”. A single, large pre-educated LLM-In this case, the lama 3.18B of meta platforms 3.18B-They placed it through an additional training phase with 60 billion new training tokens. The result is an algorithm called “router” that can automatically output any number of different -sized offspring on any button push.
Fugatto Foundation Model
In the category of more tangible programs, Fugatto 1 Program There is a “Foundation Model” for audio synthesis, an AI model that can handle any combination of text instructions and sound clips and change the clip based on instructions. “I am really excited about Fugatto,” Catanzaro told ZDNET.
Too: I have tried a lot of AI image generators, and one is one to defeat Nvidia and MIT speed
For example, Fugatto can produce a sound at request, such as the cat’s meow. This can separate a song sample to reproduce each separate singer. It can merge the sound of ripping water with a classical guitar sound to create a hybrid sound that is an entrance of both.
Fugatto has a nerve net Developed in Google in 2022 It can work on “spectrogram”, it seems as a wave pattern. The original contribution of Rafael Valle and his team of Nvidia is a new dataset and a training diet that teaches the model to handle complex text orders.
Too: Rise of AI PC: How Business is rebuilding its technology
Nvidia projects such as Fugatto manufacture on many pre -innovations, as any research lab does. One of the important aspects of NVidia’s research papers is that they provide more technical details about the hardware implementation used in research, such as, for example, the number of GPU chips used, while other laboratories often leave that data.
AI informs chip development
Research projects such as LLAMAFLEX and Fugatto do many functions. They highlight many ways that can be used for Nvidia chips, which is always a great way to promote the capabilities of those parts. They also include Nvidia in the art of art for AI, which can inform the company’s chip development. They help the company to attract talent by displaying projects that can win the award and get colleagues recognized.
Too: The latest chip of Google is about reducing a huge hidden cost in all AIs
And they show how the raw power of Nvidia chips plays a big role as a region in AI. The “acceleration” of AI is a story that is not known as it should be.
“It is my belief that a lot of progress in AI in the last 15 years has actually come from acceleration,” said Catanzaro.
Check all NVidia Research publications Main research website,
Get top stories of morning with us in your inbox every day Tech Today Newsletter.
