07/19/2021 at 08:00 CEST
Einstein said in 1929 that imagination is more important than knowledge, but he never imagined that Artificial Intelligence could also achieve this feat of the human mind.
Imagination is a superior creative process that allows the mind to represent objects, sensations and ideas that do not always have a correlate in reality: for example, an orange cat.
It is a complex cognitive process that allows us to think in possibilities: for example, we can imagine possible consequences of a decision that we must make.
According to Einstein, while knowledge is limited, imagination is universal, it can encompass anything, and science recognizes that it is the field of scientific hypotheses.
Technological development has managed to replicate in a machine some of the attributes of human intelligence, such as learning and solving problems. We call it Artificial Intelligence (AI) and we have been developing it since 1956.
We have designed intelligent systems that think like humans (artificial neural networks), that act like us (robotics), that reason like people (expert systems), and that even behave rationally (intelligent agents).
Related Topic: Chaos Helps Thinking, Even Artificial Neural Networks
Artificial imaginationNow, new research adds one more power to AI: we can also make it imagine things, like humans do.
When we imagine something, for example, the orange cat, the neurons in our brain are activated to generate chromatic variations of a well-known object, such as the cat.
A team of researchers from the University of Southern California has made an AI use human-like capabilities to imagine an object with different attributes. Just as we would a cat that changes color.
The idea is totally disruptive, because until now advances in deep neural networks and the most advanced AI systems are based on algorithms that cannot handle subjective information, like the one we use when we imagine colored cats.
One of the architects of this technological feat, Yunhao Ge, explains in a statement: “Human beings can separate their learned knowledge by attributes – for example, shape, posture, position, color – and then recombine them to imagine a new object. Our development tries to simulate this process using neural networks ”.
AI developmentThe artificial imagination achieved is based on a development of current AI systems: artificial neural networks can now generate multiple images of cars from photos of different makes of vehicles embedded in their architecture.
The AI extracts rules from some examples and applies them to a wide range of previously unknown car models. And you can represent the new models in any color and from multiple angles. But it lacks common sense.
The new development goes a little further by taking advantage of a capability already developed by AI, known as Deepfake: it allows, for example, to generate people’s faces using machine learning algorithms and videos and stock images. The end result is a very realistic, albeit fake, video of someone who doesn’t exist.
Similarly, the researchers explain, the new development takes a group of sample images, rather than one sample at a time, as traditional algorithms have done, and extracts the similarity between them to achieve something called ‘untangled representation learning. controllable”.
Then, he recombines this knowledge to achieve a “controllable novel image synthesis”, or what we might call imagination, stand out.
In other words, artificial imagination is achieved by offering AI the ability to construct fiction by combining several simultaneous sources of archived information and then synthesizing it into a real-looking image.
The researchers consider that this system does the same thing that we do when we imagine cats of different colors: use known patterns (cats, colors) to combine them in non-existent ways (for the neural system) and derive it to a combined image of an orange cat.
Using this technique, the researchers have generated a database of more than 1.5 million images that could aid future imaginative developments in AI.
Potential applicationsAlthough artificial imagination is based on pre-existing ideas, researchers consider that its development can be compatible with almost any type of situation, with multiple potential applications.
The ability to imagine the unknown can empower computers to make AI systems fairer, by completely removing cultural biases related to race or gender.
It could also help physicians and biologists discover more useful drugs, synthesizing new drugs derived from imagining possibilities from the properties of various drugs.
Infusing machines with imagination could also help create safer AI, for example by allowing autonomous vehicles to imagine and avoid dangerous scenarios never seen during rehearsals.
“Deep learning has already shown unsurpassed performance, but this has been accomplished many times without understanding the attributes that make each object unique. For the first time, we have a new sense of imagination in artificial intelligence systems, ”concludes Yunhao Ge.
ReferenceZero-shot Synthesis with Group-Supervised Learning. Yunhao Ge et al. The International Conference on Learning Representations (ICLR), Vienna, May 2021.
Top image: The new AI system is inspired by humans: when a human sees a color of an object, we can easily apply it to any other object by replacing the original color with the new one. Illustration / Chris Kim (USC).