Why do simple human skills challenge AI? Exploring Moravec's Paradox, computational creativity, efficiency of AI vs the brain, and more.

Transcript

Hello! I’m Leo Isikdogan, and welcome to the Cognitive Creations podcast. Today’s topic is simple things we can do that artificial intelligence cannot master, at least not yet.

A decade ago, the narrative around AI was all about how it would revolutionize manual labor, with autonomous driving leading the charge. The expectation was that truck drivers would be the first to feel the impact. Yet, here we are, with truck driving still in high demand.

Surprisingly though, it's the jobs once considered safe – like those of software engineers, lawyers, and artists – that are experiencing significant shifts because of AI. Tasks in these fields are increasingly being automated.

Moravec's Paradox

This shift aligns with Moravec's Paradox, which suggests that tasks that are easy for humans are hard for robots, and vice versa. It's simpler for AI to write code or create art than to master basic sensory motor skills.

It's relatively easy for a computer to beat a human in chess. However, it's very challenging for a robot to match a human in skills like picking up a pencil or cleaning up a cluttered room. Yes, robot vacuums today, with all the advanced AI powered smart features they have, still suck, both literally and figuratively.

These tasks are so natural to humans that a child can do them effortlessly, but they're a lot more difficult for robots. We're seeing AI make strides on this front but still, the progress is not nearly as fast as what we have seen in computer vision and language models.

This paradox highlights a key point: the brain's billion-year evolution has made complex tasks seem effortless. Replicating these in AI is a monumental challenge.

Skills like recognizing a face, navigating space, or understanding social cues are deeply ingrained in us, developed over millions of years. They seem simple because they are automatic, but in reality, they are incredibly complex.

Contrast this with skills like proving theorems and engaging in complex scientific reasoning. These don't come as naturally to us because they're relatively new in the timeline of human evolution. They require conscious effort because they have had at most a few thousand years to be refined, and that's mostly by cultural evolution.

This brings us to another question: Where does creativity fit into all this? How does AI measure up when it comes to creativity?

Today's generative models can produce text and images that are remarkably convincing. They are still not so great on the musical side, but they are getting there.

Why is music hard to get right?

But why is that? Why is music harder to get right? Maybe let's first talk about that before we get back to the question of whether machines can truly be creative.

One challenge in music, unlike in images and text, is that everything is transparent in a sense. Objects have boundaries, and words and phrases have beginnings and ends, but sounds are all mixed together throughout.

In music, every element is intertwined, where boundaries are less distinct than it is in visual objects or textual phrases. This overlapping nature makes it difficult to separate individual sound elements, whether it’s in a waveform in the time domain or a spectrogram in the frequency domain.

Another challenge in music is that sounds are not as localized as images or text. In images, neighboring pixels are highly correlated. Pixels that make up an object likely cluster together. In text, relevant parts are usually not far from each other. Music, on the other hand, often has longer-term dependencies. It requires understanding and generating a long context.

Lastly, there's the issue of data diversity and availability. Unlike text and images, high-quality, diverse datasets for music are less abundant, which can limit the training and performance of generative models in this domain.

These are some of the reasons why research from the language and vision domains doesn't translate well to music. But none of these are fundamental limitations that prevent AI from ever generating good music.

The Nature of Creativity in AI Models

Back to the question: whether machines can be creative.

I think whether machines can create is no longer a philosophical question, we know they can, at least in some domains with some limitations.

One of those limitations is the extent to which they can be creative. Creations of generative models are often interpolations of existing data. So maybe there is still some room for philosophy here to what it means to be truly creative.

Generative AI today is great at generating new samples that are a mix of what it has already seen but struggles to create something truly novel. This doesn’t necessarily mean they copy what’s in their dataset, but their creations remain within the boundaries of their training data distributions.

Here’s what I mean by that:

Take a neural network trained to do division, as a toy example. We can teach an AI model to divide one number by another by showing it some examples. Division operation is actually not so easy for neural networks to learn. Still, the model would be able to approximate it as long as the inputs are within the range of the examples in the dataset. But it would most likely fall apart if we ask it to divide numbers that are out of range of the samples in the dataset unless we employ some special tricks, like operating in the logarithmic space where division becomes subtraction.

This limitation is evident in many AI applications. So, what generative AI models do today doesn't go beyond imitation and interpolation.

That's why I see AI models as tools, rather than being artists in their own right, but I keep an open mind. One day perhaps we can call AI systems non-human artists. But today, they are far from being fully autonomous when it comes to creative tasks.

With machine learning, we don't have to fully understand the underlying mechanics of a problem to solve it. AI may figure it out for us, given a set of examples and a function to optimize for.

That doesn’t mean we need no understanding of the problems we are attempting to solve though. We still need to have a substantial amount of understanding so that we can properly formulate the problem. So, there's still an ample amount of human involvement in today's AI systems.

AI as a Tool in the Creative Process

To build truly creative models, we need them to go beyond what they see. Maybe if we can mathematically quantify creativity, then we can train AI to optimize for that. Maybe then we can manage to build models that can extrapolate beyond what's in the dataset.

But as of today, we're still a ways off from AI being truly creative. As of now, AI remains a tool, an incredibly powerful one, but not yet a creator. So, artists shouldn't worry, at least not for now. New tools bring new forms of art.

As someone involved in both research, engineering, and art, I see this as a positive thing. AI taking over the grunt work means that we can focus on the things that matter.

When photography was invented, many believed it would be the death of painting. But instead, it led to new art movements like Impressionism, where artists focused less on capturing reality and more on expressing their perception of it. Similarly, AI in art is leading us to unexplored territories.

As a side note, photography itself also has evolved a lot since then. Computational photography today is pushing the boundaries of what can still be considered a real photo. There is no such thing as no-filter in digital photography. Even at a very low level, before photons hit a sensor and create a current, they go through a filter, a physical color filter, called a color filter array.

But anyway, I digress.

Generative AI hasn't changed the fundamental role of an artist. An artist's role is still to create something new, and artists have always found ways to innovate with the tools they have. Without innovation and storytelling, all art, AI or otherwise, would start looking the same and wouldn’t be called art at that point. Speaking of which, check out my artwork, just search my name on the web and you’ll find it — and here is my shameless plug.

Efficiency of Brain

Finally, another area where AI cannot match the human brain yet is in its efficiency. The human brain is incredibly energy efficient, despite its high energy demands by biological standards. It runs on a mere 20 watts. A single high-end GPU runs at about 400 watts, and we need a cluster of those GPUs to train and run large AI models.

Moreover, it's not only about energy efficiency but also data efficiency. Humans can learn with far fewer examples than machine learning models can. We don’t need to see all kinds and breeds of cats from all different angles to understand what a cat is. More on that in the next episode.

Alright, that was pretty much it. Thanks for listening and I’ll see you next time.