With more and more enterprises turning to AI for a myriad of tasks, companies quickly find out that training AI models is expensive, difficult and time-consuming.
Finding a new approach to deal with those cascading challenges is the aim of a new startup, MosaicML, that just came out of stealth – helmed by ex-Nervana, ex-Intelite Naveen Rao – and is now preparing to launch a cloud-based neural network training system that aims to attack the problems at the algorithmic and systems levels.
The idea is to make machine learning more efficient through a composition – a mosaic – of methods that together accelerate and improve training, the company announced in an Oct. 13 blog post from its founders.
MosaicML’s core idea is that since it is expensive to train machine learning models in the cloud, in datacenters or on-premises, that the answer lies in eliminating inefficiencies in the learning process.
The startup has built two components that will be part of its future product offering, Naveen Rao, the CEO and co-founder, told EnterpriseAI. Composer is an open-source library of methods for efficient ML training that can be brought together into “recipes,” starting with some 20 different methods curated and rigorously benchmarked for their performance benefits. Additional methods will be added as the product matures.
The other MosaicML component is Explorer, a visualization tool and interface that gives enterprise developers the ability to simulate, map out and choose the best routes for running models by comparing costs, quality and the time that will be needed to run the experiments. Explorer is designed to give users a visualization of the measured trade-offs of cost, time, and quality across thousands of training runs on standard benchmarks. Users can filter by method, cloud and hardware type to reach their optimal operating test protocols.
“The key here is that these techniques actually make the training process more compute efficient,” said Rao.
The idea and need for MosaicML came out of the rise of AI, machine learning and the steps that were initially established to create and test models, he said. The original technologies were established over time, and they worked, but it turns out that there are better ways to do things, he added.
“It is like anything else,” said Rao. “[Data scientists] came up with something that basically works but was pretty inefficient. The deep learning world has been about showing that things can work and not be efficient, it did not really matter because compute was relatively cheap.”
The problem was it was only true when the models were small, said Rao.
“Once models got very big, the compute side of it actually got very expensive,” he said. Now we are at this inflection point where the models got very big and data sets are very large, so the expenses are now quite big. GPT-3 [the natural language AI model] cost $5 million to train – that was one single experiment that cost $5 million.”
That is where MosaicML began seeing its opportunity in the world of AI and machine learning.
“We are focused on enterprise companies whose core competency is not AI or ML, but they need to be able to use these techniques in a cost-effective manner to extract value from their data,” said Rao. “If you are Facebook or Google, you have a huge team who can do this, and they can spare the expensive computing and manage it on their own. They will eventually probably use these tools as well … but they do not really need us upfront. The enterprise is where we go first.”
MosaicML was incorporated on Dec. 1, 2020, and has raised $37 million from investors so far, including Lux Capital, DCVC, Future Ventures, Playground Global, AME, Correlation, E14 and several angel investors.
Rao said the company is having conversations with customers but that it has not yet made any sales. MosaicML released its open source library so potential customers and developers can use it and get a sense of its capabilities and features.
The company’s product, which has not yet been officially named, is expected to be available in the beginning of 2022 in a free version and in a paid, supported version.
“When you are training a model, all you really care about is cost,” he said. But later you begin to think about other factors, including how long things will take and how it will perform.
“This Explorer visualizer allows you to see the difference,” he said. “If I want to not pay as much and just do a bunch of experiments for cheap, I can do that, and predict very rationally where I will be when they are done. The idea is to give users tools to allow them to understand how much things cost. If they don’t have any idea, they really cannot plan, and it becomes very difficult to run these experiments.”
Initially, MosaicML will work on models that are being done in the cloud, said Rao, since those variables are easier to measure based on rate costs from each vendor. He said he expects similar capabilities will be available for on-premises uses in the future. “But we are not there yet.” He added.
Rao has been involved with AI for some time. He founded AI chip company Nervana Systems, which was acquired by Intel in 2016. He then joined Intel and started and ran Intel’s AI division, he said. Intel shuttered Nervana in early 2020 and Rao left the company.
Karl Freund, founder and principal analyst at Cambrian AI Research, told EnterpriseAI that MosaicML has chosen a viable approach to helping AI users.
“MosaicML is going after the model optimization problem,” he said. “Some optimizations, such as Nvidia TensorRT, optimize for specific hardware, but MosaicML is going after algorithmic optimization.”
That makes it model-specific and not hardware-specific, he said.
“AI hardware for training is very expensive and exotic technology,” said Freund. “If you can reduce training time by 50 percent, that reduces costs accordingly. And the client does not have to hire the super-high-priced talent, either.”
Another analyst, Addison Snell, CEO of Intersect360 Research, said that the popularity of AI “is bringing more organizations to high-performance computing for the first time, whether they think of it that way or not. And for any organization moving into HPC, model creation and optimization is one of the biggest challenges, certainly more so than simply getting access to hardware.”