What’s Trending in Machine Learning?

The first Terminator movie will soon turn forty, but the words “artificial intelligence” don’t scare us like they used to. AI methods are in use in many sectors, from healthcare to finance and architecture. Models train, algorithms comb through data, while humans reap the benefits. It seems like we’ve mastered the technology, and there’ll be no new machine learning trends in 2022 to surprise us.

This couldn’t be farther from the truth. The realm of AI, and machine learning (ML) in particular, is expanding in every direction. Here are a few rising trends to keep an eye on:

• MLOps
• Low-code machine learning
• Dimensionality reduction
• Self-supervised ML
• Reinforcement learning
• Specialized chips for AI

In this article, we’ll explain what each of these word combos means and how the technology behind them can give you a competitive advantage. Let’s begin with a term you may well have heard before.

MLOps

Setting up, running, and supporting ML systems is a multifaceted process, and code is only one of the components. You have to collect and verify data, devise models, train and update them, and manage and monitor the infrastructure. MLOps can help make these operations more sustainable and autonomous, especially as input datasets may change.

MLOps is essentially a combination of machine learning and DevOps. On top of ensuring the functioning of the ML model, it adds continuous testing and delivery. The approach aims to build a complete ML pipeline that will include steps like:

• Obtaining and preparing the necessary datasets
• Setting up the ML architecture 
• Training models
• Running tests and choosing the most suitable model
• Evaluating and deploying the model
• Monitoring performance and managing infrastructure

There’s another set of practices with a similar name — ModelOps. Let’s quickly define it and look at the purpose of following MLOps vs ModelOps.

Model operations for AI, or ModelOps, is an approach for creating operational ML models and automating the development process.

The two concepts overlap, but ModelOps concerns itself more with the development cycle of an ML solution, while MLOps focuses on communication and collaboration. Thanks to ModelOps, we can create a self-correcting system capable of adapting to changes in datasets. This way, the same ML models can be reused and scaled without rewriting code. It’s especially relevant in the context of today’s ever-growing data pools and changing business objectives.

But it’s not the only solution.

Low-code machine learning

Machine learning is a complex discipline. Implementing its algorithms properly requires special training and can be time-consuming unless you have pre-built blocks of code and a visual interface to move them around instead of starting from scratch. 

That’s what low-code machine learning is all about. Its main purpose is to bridge the gap between business demand for ML features and the shortage of ML professionals on the market. Low-code ML solutions allow a simplified, streamlined workflow and bring business users into the process. 

• Software developers can use tools like Google’s Auto ML, Microsoft’s Azure Machine Learning designer, or Outsystems ML Builder to leverage the might of ML in their projects.
• Data scientists will benefit from using platforms like DataRobot, Dataiku, and RapidMiner, and learning libraries such as Pycaret.
• Some platforms, like Apple’s Create ML, even offer no-code options for users with little to no background in software development.

Even though such platforms constantly add features and integrations, it may be difficult to find a package for your needs. In this case, opting for a custom-tailored ML solution is recommended.

Dimensionality reduction

In machine learning, dimensionality refers to the number of features in the input data. Imagine rows and columns of data as coordinates in a geometric system and you’ll understand how the overall area can quickly grow. This can negatively impact the performance of the algorithms, so data scientists constantly need new solutions to the problem.

Currently, several techniques exist that help reduce dimensionality. Let’s briefly describe them:

• Manifold learning is a method of creating a low-dimensional projection of high-dimensional data. It effectively reduces the volume of data, preserving the original relationships between fields.
• Feature selection is a set of techniques that use statistical methods or scoring to keep certain features in favor of others. 
• Matrix factorization methods use linear algebra to reduce datasets to their components. 
• Autoencoder methods involve building a deep learning neural network that will compress the input data into a bottleneck layer. The resulting layer will have fewer dimensions than the original.

There’s no ideal fix, but a specialist well-trained in these techniques can quickly select the right method for a particular project.

Self-supervised learning

In essence, self-supervised learning (SSL) is an approach for devising ML systems that don’t require the input data to be labeled by humans. However, to achieve that, small labeled datasets are often used for pre-training. That’s very similar to how we learn new concepts — by comparing them to what we already know, singling out defining features, and then making inferences.

SSL holds a lot of promise when it comes to working with unlabeled data. These systems help save costs (labeling takes time, and someone has to be paid for it) and can digest data of lower quality. It’s usually based on artificial neural networks and is often used for image and speech recognition.

Here is an example that will help explain how SSL works (in general terms). 

Let’s say our goal is to create an image de-noising tool (as used by professional and amateur photographers to rescue images taken in low light). For this purpose, we’ll turn to autoencoders (AE) — a set of components that can be used in many ML paradigms, including SSL. 

First, we need to pre-train the algorithm on a small labeled dataset:

• We take the original noise-free image and corrupt it with a noise pattern
• We then feed the original and altered images into the neural network, with the original serving as our supervision signal (telling the algorithm “this is the result we want”)
• The SSL algorithm compares the two sets of data and attempts to define patterns
• After multiple iterations, the system will be able to understand what noise is
• Then we can feed the algorithm large sets of unlabeled data so it can fine-tune the process
• As a result, we can effectively use the newly developed algorithm to reverse the process and remove noise from images

With enough data, a similar succession of steps can be applied to text. For instance, an SSL system can teach itself some basic concepts and their interrelations by filling in blanks in sentences. This way, AI can learn the human language in a similar way babies do — without direct explanation. Meta, the successor to Facebook, uses SSL to enhance its speech recognition module.

Reinforcement learning

Here’s another analogy with human behavior: ML models can be trained with the carrot and stick method.

Take AI-based bots in video games, for instance. By receiving rewards and penalties for certain movements and actions, AI can quickly learn the map and its objectives. And while we’re talking about bots, robots are another great use case for reinforcement learning. Guiding a robot with reinforcement and punishment can teach it to perform actions that humans can’t formulate or demonstrate directly.

The key here is to allow the model to find its own way, only specifying the end goal.

We can take it a step further, letting two neural networks compete against each other in a zero-sum game. Coincidentally, that is as close to the generative adversarial networks (GAN) definition as it gets. In this approach, the competing networks are called the generator and the discriminator. 

The former generates content based on real data from training sets, while the latter tries to evaluate its authenticity. With sufficient repetitions, the GANs can produce original content that is indiscernible from real-world samples. That’s why they can be successfully used for creating fake social media accounts. 
Other applications of generative adversarial networks are found in science, design, advertising, and art. GANs are capable of producing photorealistic renders of objects and photos of people, as well as de-noising astronomy images.

Specialized chips for AI applications

As it turns out, the development of ML isn’t limited to software. Neural networks require exceptional computing power to process data, and specially optimized chips can help speed things up quite a lot. 

Google Tensor

Google designed its new chip for use in its latest Pixel 6 series phones. A premium system on chip (SoC), Tensor consists of several modules that enable ML and security features. It can run new versions of ML models for computational photography and speech enhancement. The new chip also has improved power efficiency, often using half the power previously needed for the same job.

Tesla’s Dojo D1

During the recent Tesla AI Day, the company shared some info concerning its much-awaited Dojo supercomputer. Replacing the current Nvidia-based clusters, the new computer will rely on a total of 3,000 of Tesla’s own D1 chips. The company will use the system to train neural networks, mainly to improve the performance of its electric cars’ autopilot.

Tesla’s AI development team expects that the new chip’s architecture will allow the supercomputer to complete tasks that used to take days in mere hours. All this power will come in handy when processing video feeds from over a million Tesla vehicles sold worldwide.

Summing up

The global AI market is currently experiencing 40% year-on-year growth, and the demand for reliable and diverse solutions is higher than ever. We hope you’ve enjoyed our overview of the most promising trends in machine learning that are currently gaining momentum. By now, one thing is clear: AI will continue to make headway, penetrating every industry and solving multiple puzzles.

If you’re ready to augment your workflow with artificial intelligence, Postindustria will gladly help. We’ve been building software for over 15 years and can offer you a custom package of AI development services. Our experts navigate the sea of AI-based options with confidence, so you’ll get the best solution for the needs of your business.

Let’s get in touch and start the dialogue.