Artificial Intuition

The limitations (and ridiculous power) of ANN creativity

Despite the marketing talk about ‘Cognitive Computing’, ANNs are in many ways more Artificial Intuition than intelligence. They fill in gaps and make intuitive leaps to produce an appropriate response to a given situation — System 1 thinking, in Kahneman’s terms.

They are powerful enough to take over any human activity that takes no longer than about one second, or a series of such moments in a loop — driving a car, recognising faces, reading handwriting, labelling objects in a scene.

1. So, we have traditional computers which are great for calculation.

2. ANNs, particularly Deep Learning, give us Artificial Intuition and potentially super-human pattern spotting.

3. That leaves us with the question of true intelligence: True Reasoning about things.

Reinforcement learning can create AI that learns about situations but cannot conceptualise them. True conceptualisation is something we do not have, and will not until another revolution in AI — one that could take a few years or a few decades. 

Some of the latest developments from MIT are able to combine multiple discrete elements of something, to syncretise something new, which seems promising. OpenCog and Wolfram Alpha are built in a top-down model whereby a system is explicitly taught things, rather than inferring properties from data (bottom-up). In theory this can lead to a more reasoning-type process. 

Both require enormous human grunt work to build, and neither is optimised. Marcus Hutter’s AIXI design would be a near-ideal AI system if it could be implemented, but it is considered computationally non-viable.

Recent work on generalising learning between ANNs is promising — though it risks smuggling biases and misconceptions in through the back door — as is learning from a single example without needing a large curated dataset. 

My hope is that some of the recent leap in bottom-up approaches, and the GPU/FPGAs/ASICs now powering AI systems, can translate to making this top-down reasoning process faster and easier.

To sum up, machine intelligence can do a lot of creative things; it can mash up existing content, reframe it to fit a new context, fill in gaps in an appropriate fashion, or generate potential solutions given a range of parameters.

Outside of a few potential hints at something deeper, ANNs do not appear to be generating purely original concepts or ideas, or performing abstract reasoning. Surprisingly few human tasks or roles actually require this kind of mental function. Most people are Cooks, and not Chefs, businesspeople rather than entrepreneurs, and they have not been taught how to reason from First Principles either. 

One area where some kind of reasoning is generally required is in ethics. This is why a project which I have co-founded, OpenEth.org, is working to create ethical constraint solutions for narrow AI, in this niche but crucial area.

Bias in how a machine perceives something can come from the algorithm, but it can also come from the data. An algorithm will generally be tweaked over time to extract better sense from whatever data is available. 

But an incorrectly weighted algorithm can reinforce existing biases in the data — stereotypes get cemented, or implicit discrimination occurs without warrant, as when certain individuals never see a job ad because they do not fit the standard pattern of hires. The worst abuses may occur within the justice system, as decision support and probabilistic engines are increasingly being used to calculate things like bail. 

Much of the engineering behind today’s most powerful machine learning happens within a small geographic zone, by demographically similar individuals, and that is unlikely to change soon.

I’m therefore proud to serve as an advisor to Diversity.AI, an organisation fighting for better, more open, and more accountable use of machine learning. Machines are meant to help liberate us. Let’s make sure they do.