Advances in machine learning are powering the current boom in AI, transforming industry and many everyday tasks. But what does the future of machine learning hold?
Table of Contents
- AI and the different types of machine learning
- Where machine learning fits in
- Companies using machine learning
- Machine learning vs human learning
- Should we all be worried?
- The future of machine learning
Today, artificial intelligence is everywhere. It’s in our phones, powering voice assistants such as Siri and Alexa. It’s in our GPS systems, giving us the most efficient route to get home. And it’s in our social media networks, giving us personalised news feeds and targeted advertising.
All these applications of artificial intelligence rely on machine learning (ML), where algorithms are taught to act with similar cognition to the human brain. This self-education has already transformed many everyday tasks and continues to transform industries. Its benefits are clear: instead of relying on humans to process information and make decisions (which can be time consuming and costly), machines can make the same judgements almost seamlessly.
But what happens when machine learning goes down the other path – the one that dystopian sci-fi thrillers love to explore? Is an Ex Machina-style future possible, where machine learning surpasses human learning? And even if not, what are the other consequences we need to consider as we build this technology?
AI and the different types of machine learning
To understand machine learning, it’s important to define it. In 1943, scientists discovered that ‘artificial neurons’ could carry out logical functions such as human-style learning. However, the term ‘artificial intelligence’ wasn’t coined until 1956, when John McCarthy used the term to sum up a range of advanced research topics during a science conference held at Dartmouth University.
Today, AI refers to the general capability of a machine being able to imitate human behaviour. This includes tasks such as understanding language, perceiving objects and surroundings, as well as continued learning.
There are two main types of AI:
The AI can exhibit some aspects of human intelligence, but is lacking in others. For the traits that the AI can demonstrate, it has the ability to do that task extremely well. For example, consider DeepMind’s AlphaGo AI. It can defeat the world champion of the game Go, but it can’t do much more than that.
The AI has all the characteristics of human intelligence. This include tasks such as reasoning, planning, abstract thinking and learning from experience. This is a huge step-up from narrow AI and is yet to be fully achieved.
Researchers believe there is a 50% chance of AI outperforming humans in all tasks in 45 yearsGrace et al, 2017, When Will AI Exceed Human Performance? Evidence from AI Experts
Where machine learning fits in
Machine learning supports the goals of AI. It’s not the only way to achieve AI – but it is the most successful so far.
This means it’s possible to build AI without machine learning. Just think of your own intelligence. Sure, pattern recognition is partly how you’ve learned to speak, calculate and read – but this learning must be paired with experimentation, decision-making and emotion for it to be complete. As such, AI – especially general AI – requires more than just looking at algorithms, decision-trees and data.
In that sense, machine learning can be seen as a clever processing technique. The term ‘machine learning’ was first coined in 1959 by Arthur Samuel who defined it as “the ability to learn without being explicitly programmed”. Instead of being fed instructions, the machine’s algorithm is ‘trained’ to self-adjust and improve to compute data better. Machines find a solution based on the data they have.
It allows engineers to make the most of data without having to explicitly program machines to follow set paths.
There are several variations of machine learning. These include:
This is where systems are designed to automatically discover the representations needed for data classification or detection. Different forms of representation learning include:
- Supervised machine learning
- Unsupervised machine learning
- Reinforcement learning
Deep learning refers to when there are a number of hidden layers in an artificial neural network (ANN). These layers are designed to mimic the biological way the human brain processes information – be it turning sound into speech recognition, or images into information. Within the ANN, there are layers and connections to other ‘neurons’. Each layer is dedicated to learning a particular aspect of a task, and multiple layers are used to complete the whole task.
Companies using machine learning
Today, most big tech companies are investing in at least some form of machine learning.
Facebook is using machine learning to predict user actions. This includes pre-empting things such as what a user will ‘Like’ and what posts they’ll click on. Machine learning is also being used to order the News Feed and make recommendations.
Google is often regarded as the most advanced company when it comes to AI and machine learning. As well as its continual acquisitions of AI startups, it offers developers a range of cloud-based tools to encourage further development in the field. One such example is its Google Cloud AI machine learning tool.
In 2017, Microsoft acquired the deep learning startup Maluuba, which it described as “one of the world’s most impressive deep learning research labs for natural language understanding”. It hopes this acquisition will advance its progress in creating fully literate machines.
Machine learning vs human learning
On an extremely basic level, it could simply be said that human knowledge resides in the brain, whereas machine knowledge exists on servers. But the differences and similarities between the two go much further than that.
While artificial neural networks may mimic human brain functions, they still haven’t achieved the same level of human intelligence. This is because artificial neurons cannot self-organise and adapt in the same way human neurons can. As well as that, machine learning cannot be programmed to include intrinsic human learning characteristics.
One such characteristic is motivation. Humans learn because we enjoy doing it and find it personally rewarding. Machines, on the other hand, can only be motivated to do things for external rewards or to avoid negative consequences.
Combining human and machine learning
Today, computers can memorise vast amounts of information and undertake complex supervised machine learning tasks. Individually, these tasks far exceed the capacity of the human brain.
But as it stands, machines can’t apply knowledge to think in more abstract ways.
However, as our propensity to rely on automated tasks grows (after all, many answers to our questions are a mere Google away), human learning may evolve differently. We can expect less of a focus on information retention in tomorrow’s classrooms, and more of a focus on problem-solving and creativity.
Should we all be worried?
The current hype around AI means there has been a lot of fear around the potential negative consequences of machine learning. But this attitude neglects to consider the benefits of joint human and machine learning.
Singularity: a potential tomorrow
Much fear around machine learning comes from singularity, a hypothetical point where AI and robots surpass human intelligence. The term comes from the gravitational singularity that occurs at the centre of a black hole, where gravitational fields are infinitely strong and the laws of physics collapse. In 1993, Vernor Vinge wrote an essay in which he applied the term to a moment in the future when technology’s intelligence exceeds our own, and life as we know it will be forever changed. It’s a theme that has formed the basis of many science-fiction stories (and usually for the worse).
Multiplicity: where we are today
This is when people and machines work together to solve problems. It doesn’t exist in the realm of science fiction – rather, it already exists in many smart systems we have today. Humans are essential to multiplicity. Diverse groups of people interact with diverse groups of machines to translate languages, make recommendations on books, and suggest tags for images and videos. This is the approach many researchers are taking with machine learning.
The future of machine learning
Humans and machines need to evolve with each other, and not in silos apart. With smarter machines, our human abilities become augmented. Just as the printing press democratised knowledge and information in the 15th century, computers with AI are doing the same now.
But as machine learning evolves, we’ll have to carefully consider a number of complex and far-reaching questions surrounding its applications. How can we make sure machine learning doesn’t promote systemic bias that comes from existing data sets? And how can we make sure humans remain purposeful in a world without work?
Because while we’re still coming to grips with narrow AI, general AI may eventually come to fruition. And once that happens, the next hurdle will be superintelligent AI. Fingers crossed an intelligent machine can help us tackle that momentous step.