Using machine learning to process data has increased revenue and efficiency for enterprises. Big tech companies have hundreds of millions of data, which is humanly impossible to process. Machine learning is, therefore, being used to build complex models.
AutoML is no longer a new term. Since Google released its first AutoML product, discussions around this technology have been quite prominent. Some regard it as a weapon to achieve general artificial intelligence, while others deem it to be exaggerated. But what everyone agrees on is that AutoML does have extraordinary significance in realising AI advancements.
Over the past few years, automatic machine learning algorithms have been used in areas including image recognition, natural language processing, speech recognition, interactive AutoML optimisation, semi-supervised learning, reinforcement learning and more. But it comes with its own set of challenges.
Sign up for your weekly dose of what's up in emerging technology.
Business Challenges For AutoML
AutoML faces problems such as data and model application. For example, high-quality labelled data is far from enough, and data inconsistencies during offline data analysis will cause bad effects. In addition, teams need to do automatic machine learning processing of unstructured and semi-structured data, which is technically difficult.
Furthermore, the current AutoML system optimization objectives are fixed. Often realistic problems are a combination of multiple objectives, such as the need to make subtle differences between decision-making and cost. With this kind of multi-objective exploration, people have limited ways to effectively judge before the results are obtained.
Such a situation is difficult to support in the current AutoML. The actual business may have customised requirements for the actual machine learning process, for example, for which only a certain type of data processing tool can be used. Such requirements cannot be met in the current black box AutoML solution. Whether it is effective or efficient, AutoML has a lot of room for improvement.
Another point worth paying attention is that it is more difficult to perform AutoML in a dynamic environment than in a static environment, as the environments keep dynamically changing. Dealing effectively with dynamic environments is an open issue in the academic community, and researchers are continually exploring the field. For dynamic feature learning, companies will need to adapt to changes in data faster, detect distribution changes and automatically adapt to different types of models, etc.
The current mainstream computing frameworks (such as Tensorflow, PyTorch, etc.) are only optimised for single machine learning model training. Companies need to work on redesigning the underlying computing architecture for automatic machine learning, providing configuration evaluation and optimisation for multiple model learning. For dynamic environment learning, it needs to be able to automatically perform model adaptation based on changes in the data distribution.
While automated machines can find solutions, it may not necessarily be what the user wants. The user may want an explainable model. In fact, experts say explainability itself has great uncertainty, because everyone’s understanding is different, and it has a great relationship with personal judgment. It is even more challenging to make the model explainable.
Organisations have to work on advancing the development of standards related to explainable machine learning. AutoML can give the results, and experts judge whether they meet their own interpretable and explainable standards and consistency.
Security and Privacy
Security is the other hot research topic of AutoML. In terms of the security of AutoML, organisations are exploring different technical solutions for different scenarios, such as automatic machine learning for privacy protection, automatic multi-party machine learning, automatic federation, etc.
But the current implementation lacks the support of laws, regulations and industry standards. Organisations need to promote the establishment and improvement of standards for federated learning and secure multi-party computing.
The AutoML technology has now been implemented in many scenarios, but the challenge is to implement it on a large scale and in more industries. The obstacle is that technological breakthroughs of AutoML require deeper research on the theoretical and algorithmic levels.
Companies are experiencing many iterations of automated machine learning. It has evolved from the earliest two-category expansion to multi-category and regression, from structured data to unstructured data such as images and videos, is used in automatic supervised learning that covers low-quality data, and in automatic multi-party machine learning that protects privacy. In theory, researchers are exploring the boundaries of the AutoML algorithm, because there is no general algorithm that can solve all problems.