2022 Predictions: AI and Machine Learning

I work with HR Tech companies as an external consultant / PM and my role is to help them achieve process automation with machine learning, specifically NLP. My experience is mainly in text classification, entity recognition, text generation, and text summarization. My preferred stack includes HuggingFace and the Transformers architecture, Sagemaker and Google Vertext / AutoML for quick prototyping. I would still resort to Fast.ai for some projects, which is a great alternative when fast execution and lightweight prototyping are needed.

1. AutoML Platforms will drive massive productive adoption of ML and AI in businesses

Implementing machine learning-based automation in a production environment is always a nightmare. You have to define your problem well. You must handle data engineering, storage, model training, fine-tuning, deployment, scaling, retraining, and more around the MLOps cycle. If you are a small or a mid-sized business owner, your existing teams' cost and lack of competence to drive such innovations can cripple you. AutoML takes a large part of the machine learning and AI cycle and does it for you. Today, as long as you have some data, you can train a state-of-the-art deep-learning text classifier on Google Vertex and deploy it as a scalable API in just a few clicks. Both big players - Amazon and Google have released their AutoML platforms and I use them all the time as quick POCs or hypothesis-evaluation tools. Of course, AutoML will not work for everyone and if you need a more custom solution, you will have to go one level deeper. Thankfully, this is not so hard anymore because the Hugging Face (http://huggingface.co) platform has laid the foundation to train more customized models quickly and efficiently.

2. Vertical integration along the MLOps lifecycle reduces complexity and drives the adoption of AI and ML.

The toolset for AI and ML is now officially entering the mainstream. Companies like HuggingFace have streamlined the training and deployment of models in the cloud (Sagemaker) by vertically integrating steps from the MLOps lifecycle in the framework. With Sagemaker, AWS has scored a particular advantage over GCC regarding the fast operationalization of transformers-based models. The ease of training, deploying and scaling models into the cloud lowers barriers and accelerates the adoption of machine learning from small and mid-size businesses.

3. The use of AI will increase in software development

I have been waiting to make this prediction for a while now. With the advance of BERT and GPT-3, we are going to see more tools available to developers solving issues around:

• writing better and cleaner code faster (Github Copilot, Kite)
• finding security issues in existing code
• translating requirements to simple code blocks (Debuild)

4. NLP explosion

The Transformer architecture has pretty much changed the NLP landscape forever. Models like BERT, GPT-3, RoBERTa and XLM allow us to solve complex NLP problems with little training data thanks to large pre-trained language models. This development is a real game-changer for all major NLP use cases - classification, token extraction, summarization, and text generation. Gone are the days when you needed thousands of samples in your training data. Depending on your use and the complexity of the problem you are solving, you might be able to train an excellent text classifier with just 10-20 samples per class. Pack this together with the simplicity of AutoML or a fully vertically integrated framework like HuggingFace (if you need a more custom approach) and you have everything in place for an explosion of NLP use cases in pretty much every (digital) industry.

If you liked this text, subscribe to my blog.

Did you like my predictions? I wrote also:

2022 Predictions for Crypto, Web3 and Defi

2022 Predictions for HR Tech and Recruiting

Do you want to discuss this further - reach out to me on Linkedin!

Have a happy new year and a great start in 2022.