AI has become a modern goldrush, with every company rushing to sprinkle AI's magical fairy dust on every corner of the enterprise, regardless of how far afield deep learning might be from the application in question and whether the resulting AI model performs dramatically worse than its classical predecessor. Even the most simplistic rules-based systems composed of a few lines of code are being switched out for deep learning replacements that perform far less accurately, have a fraction of the reliability and require orders of magnitude more computing power to execute. This goldrush is being driven by the hauntingly accurate results AI has delivered in fields like image, audio and video recognition. Yet, at the end of the day, these algorithms are merely correlation machines, sifting through vast piles of numbers to record subtle correlations among inputs without any high order understanding that would allow them to divine causative relationships. In the end, we are building our AI revolution on a correlation house of cards.
Even for those with deep technical backgrounds, the siren song of deep learning can be seductive. Rather than spend days, weeks or even months writing complex pieces of unmaintainable code, any task can be magically automated merely by feeding training examples to a deep learning algorithm that will, with a few hours of training time, magically produce an algorithm more accurate than the world’s foremost human expert.
Of course, the reality is that writing production-quality deep learning code can be vastly more complex than traditional code, the resulting models can be difficult to execute, let alone maintain and with the rate at which the underlying deep learning toolkits are evolving, today’s production code can become next month’s legacy code relying on a depreciated workflow that is no longer supported by the toolkit.
Deep learning systems are black boxes with unknown edge cases whose human-like accuracy on some cases can lead us to temporarily forget their spectacular failures in others and our inability to know where those edge cases are.
Without an understanding of how the deep learning systems we are betting our societal future on work, we are making society vastly more fragile, unpredictable and vulnerable to discriminating against entire demographics whose existence is not captured in the free-but-poor-quality datasets driving the AI revolution.
Lost in the hype and hyperbole of the deep learning revolution is the underlying great leap of faith into the unknown: the shift from causation-based business rules we deeply understand into a world of massive ledgers of weak and obscure correlations recorded inside black boxes.
We are literally betting the future of our planet on statistical correlations.
Today’s deep learning systems don’t actually “understand” the world.
They do not take their reams of inputs and abstract upwards to high-order entities described by properties and connected to other entities through relationships and transitive causation.
Instead, our most powerful AI systems merely discover obscure patterns in vast reams of numbers that may have absolutely nothing to do with the phenomena they are supposed to be measuring.
Most importantly, we have little way of testing whether those correlations are wrong until our algorithms fail in the most spectacular, and unfortunately sometimes fatal, ways.
The early era of AI was built on the idea of machines understanding the world in the same way we do: as semantic abstractions and relationships whose interactions are guided through causation.
Building such AI systems is really hard, so we’ve gone back to what’s easy: building correlation engines.
It is a lot easier to build a machine that can spot patterns in a pile of numbers than it is to build a machine that can take those numbers and use them to build a mental model of the world they describe.
In the same fashion, we have built our AI revolution upon free data that we acknowledge is extraordinarily biased in ways that severely harm entire demographic groups. Yet, we refuse to pay for good data that would mitigate these biases because that requires extra work and money.
In short, whether it is the training data that represents our world or the algorithmic models built upon that data, we have chosen the short-term path that is easy and cheap, rather than the long-term path that is far more expensive but is the only path that takes us towards truly robust and less biased AI.
Our current flirtation with correlation-based AI cannot last. The resulting systems are too brittle, unpredictable and unable to reach beyond the narrow confines of their training data to allow us to build truly robust and generalizable AI.
Yet, we continue to build up this house of cards.
We compensate for lack of real training data by building simulators to flood our algorithms with artificial edge cases to narrow their instability.
We augment our deep learning algorithms with hand-built code to handle the really important parts that need understandability and repeatability.
We come up with ever-more-convoluted ways of wrapping the learning pipeline with algorithmic ducttape that lets us take a few more steps down the correlative path before the ground gives way beneath us.
We search for creative abstractions, exotic sensors and new forms of representation that allow us to shoehorn new applications into a correlation framework.
Most importantly, we project a reality distortion field around the world of deep learning, touting AI as the ultimate solution to all the world’s problems, even while acknowledging internally its immense limitations.
Putting this all together, we have built our modern AI world upon a correlation house of cards that is beginning to buckle under the strain of all our hopes and aspirations colliding with statistical reality.
In the end, this house of cards we have built upon correlations will come crashing down. Building a world based exclusively upon correlations with no connection back to the causative reality connecting those patterns can stretch only so far. While there is considerable room to achieve quite a few more success stories with this correlation-only approach, there is a greater awareness of the immense fragility of the resulting systems and their very real limitations.
Eventually we will have no choice but to let this brief experiment in correlation fail and watch our AI house of cards collapse.
After all, for machines to truly understand the world we must evolve from correlation to causation.
