CloudLess Data Sharing: How Machine Learning Finally Reached Micro and Small Enterprises

Original article was published by Miguel Tomás on Artificial Intelligence on Medium

CloudLess Data Sharing: How Machine Learning Finally Reached Micro and Small Enterprises

Share on Twiter | Facebook| email

In the past decade, “big data” has become the biggest buzzword everywhere. When trained on an incredibly large data set, machine learning (ML) models can deepen their understanding of a given field, thereby bringing breakthroughs to top high-tech companies. For example, Google adjusts its ranking algorithm by tracking and analysing more than one trillion search queries every year. Facts have displaced the daunting task of past times when answering all questions and is now reduced to small algorithms that can run from your pocket everyday.

But there is a catch, most companies are limited to “small” data. In many cases, those companies and enterprises only have dozens of examples of processes to be automated using ML. And this is one of the limitations of ML and AI technologies. The need of big amounts of data in order to perform well.

If your company is looking to build a robust ML system, it needs to develop new technologies to overcome the problem of insufficient data.

In transforming small data into big data, two technologies, especially “transfer learning” and “collective learning”, have proven to be crucial, and enables medium-sized companies to learn from machine learning that was once reserved only for big tech companies. Another important fact about this is that only 15% of companies have deployed AI or ML, technologies that have proven huge opportunities to change the business world.

From the figure above can be seen that using only one company’s data, even modern machine learning models only have an accuracy rate of about 30%. But thanks to collective learning and migration learning, is possible to determine with an accuracy of more than 90%.

the road to Open Source

Of course, data is not the only prerequisite for building a world-class machine learning model-it doesn’t matter if the model is initially built. Given the shortage of machine learning engineers, for most organisations, there is no option to hire a team of experts to build an ML system from scratch. This gap helps explain why resource-rich technology companies like Google get disproportionate benefits from ML.

But in the past few years, some open source ML models released by Google in 2018 (including the famous BERT model for understanding languages) have begun to change the rules of the game. The complexity of creating a BERT model, named “large” version of the model, has approximately 340 million parameters, which means that few organisations can even consider using this kind of strategy to apply directly on their businesses. However, because it is open source, the micro and small companies world can now adjust this publicly available playbook to address its specific use cases.

To understand what these use cases look like, consider the point of view of a small company. This company itself does not have enough data to build and train effective ML systems for internal use cases. However, its small data does contain a lot of insights, waiting for ML to unlock it. By using new technologies to gather these insights, micro and small companies become more efficient, by identifying which internal workflows need to be paid attention to on their day to day business activities.

Great rewards with small data

So this is a trillion-dollar question: how do you adopt an open source ML model designed to solve a specific problem and apply that model to different problems in the enterprise? The answer starts with transfer learning, which undoubtedly requires the transfer of acquired knowledge from one field to another with less data.

For example, by adopting an open source ML model like BERT that aims to understand a common language and perfecting it at the margin, ML can now understand the unique language employees use to describe IT problems. Language is only the beginning, because we are just beginning to realise the huge potential of small data.

From the figure above it can be seen how transfer learning uses knowledge from related fields-usually providing large amounts of training data-to increase the small amount of data for a given ML use case.

More generally, this approach of providing very small and very specific training data choices to ML models, called “few-shot learning”, and this term has quickly become a new buzzword in the ML community. Some of the most powerful ML models ever created, such as the landmark GPT-3 model and its 175 billion parameters, are orders of magnitude higher than BERT. .

GPT-3 essentially takes the entire Internet as its “tangential domain”. By establishing a strong knowledge base, it will soon be able to master these novel tasks, just like Albert Einstein can become a checker without a lot of practice. Like a master. Moreover, although GPT-3 is not open source, applying similar one-time learning techniques will enable new ML use cases in enterprises with little training data.

The power of collective and collaboration

With the help of transfer learning and fast learning functions on the powerful open source model, ordinary enterprises can finally start to play in the field of machine learning. However, although the amount of data required to train ML through migration learning is several orders of magnitude less, further efforts are needed to achieve powerful performance.

This step is collective learning, which works when many companies want to automate the same use case. Each company is limited to small data, and third-party AI solutions can use collective learning to merge these small data sets to create a large enough corpus for complex ML. In the context of understanding language, this means abstracting out a company-specific sentence to discover the underlying structure:

Collective learning involves abstracting data (in this case, using ML for sentences) to reveal common patterns and structures.

The combination of migration learning and collective learning and other technologies are rapidly reshaping the limitations of enterprise ML. For example, aggregating data from multiple customers can significantly improve the accuracy of models designed to understand how their employees communicate. Of course, in addition to understanding language, we have also witnessed the rise of a new type of workplace-a workplace supported by small data machine learning.