What Can AIs Learn From Humans? Artificial Intelligence Simplified

Article originally posted on www.ocularpoint.com

Artificial intelligence (AI) has been at the forefront of major technological innovation for the past decade, with its usage growing every day. In fact, it is being utilised by around 37% of businesses and organisations today, with the AI industry estimated to be earning shy of £100 billion by 2025.

AI is described as a machine agent’s ability to perceive its environment, whilst reacting in a way that successfully maximises or completes its operational outcome.

This endeavour has long been of interest to philosophers and scientists alike, not only as a way to introspectively understand our own brains, but as a way to accelerate technological growth and assist humans in data-heavy tasks in which they typically underperform.

In this article I would like to go over where AI currently stands at the moment; what AIs can learn from humans; and how we can do so with a practical example.

Where is AI heading?

Looking into the future, the goal of AI is to develop systems that display flexibility (neuroplasticity) similar to that observed in humans; the ability to take knowledge learned from one task and apply it to unseen tasks. This is known as transfer learning or task generalisation.

As technology moves towards the integrated application of AI in major fields such as engineering, design, business and finance, a focus on addressing this discipline’s key shortcomings becomes crucial.

Unfortunately, transfer learning has notoriously been a weak spot for AI since its inception. Whilst efficient at optimising specific problems, little progress has been made towards advancing transferable skills. A key aspect of intelligence is the capacity to transfer and employ previously learned information into new environments.

Benchmarking tools such as the OpenAI Retro Contest have helped inspire new approaches towards achieving this goal, however, there is still room for improvement, as these approaches perform considerably worse than humans in task generalisation.

As such, it is natural to ask: “Can our AIs learn anything from human performance?”

How can AIs learn from humans?

There is a machine learning approach, called supervised learning, that has been successfully implemented in the past, in cases such as AlphaGo, however, it has never been carried out and tested properly in a transfer learning environment.

It is important to streamline our methodological approaches as much as possible, so as to utilise the currently limiting available hardware to its maximum capacity. Can our AIs be taught how to quickly learn from humans?

The aim of a supervised learning approach on an AI machine is to have a positive impact on learning rates and times (ie the time taken to learn a specific task), which ultimately leads to a reduction in computational expense. As mentioned above, with our conventional hardware complexity plateauing, it is important to streamline our AI learning techniques to best utilise these computational resources.

Of course, with quantum computers around the corner, we are bound to see a profound impact on the speed and evolution of machine learning — you can read more about the current state of quantum technology here.

Putting this to practice

During my university career, I put this very idea into practice with none other than Sonic the Hedgehog (inspired, of course, by the OpenAI Retro Contest).

What you are about to see below is a comparison between an AI that has been ‘taught’ how to play the game by a human, versus an AI which has been let loose and made to figure out the game for itself (both of these are trained for the same amount of time). These are then placed into a level of Sonic that they have never seen before and using only prior knowledge of the game mechanics, they must figure out how to adapt to complete the level.

Below is their progress after 5 minutes:

Unsupervised AI

As you can see, an AI that has been trained by a human is at a more advanced skill level than an AI that has learned how to play on its own from scratch — as expected. Of course, were we to leave our unsupervised machine to play for a while, it would eventually reach the skill level of our supervised player. However, we must remember that we are trying to conserve computational power!

More importantly, what’s most impressive, is that these skills are transferred into a never before seen environment and used to adapt and learn on the spot.

Now, there are, of course, much deeper nuances as to what is going on here, such as our algorithm architecture (for anyone interested, the above are a variation of a Proximal Policy Optimisation [PPO]) — as algorithms are not all created equal.

However, in a scenario where we are using the same algorithms, we can see drastically improved results just through our training method!

Brief Summary

So, AI affects most areas of our lives nowadays and it is important that we try to tackle its shortcomings if we are to progress in this field. AI is still a long way away from sci-fi levels of intelligence, however, progress is still being made.

The above brief overview provides just a glimpse into the future of transfer learning and how it could help streamline our AI learning approaches. It has been shown that AIs can effectively gain a headstart in the learning process by observing and mimicking human patterns. This ultimately saves on our computational expense.

Having said that, this performance, when compared to a human, is still considerably sub-par; so there is a long way to go before achieving an AI that displays similar neuroplasticity to that observed in humans.

This article was originally posted on www.ocularpoint.com