Imagine you are driving down a mountainous road when suddenly you hit thick fog. Your response is instinctive. You narrow your vision and sharpen your focus to see any approaching cars.
The human brain is able to handle such rapid changes, but with an artificial intelligence system (AI), things can easily go wrong.
AI systems today are very accurate in good visibility. A self-driving vehicle can recognize pedestrians, signs, and other cars with accuracy on a sunny, clear day. They are, however, extremely sensitive to changes in the environment.
When it’s raining, dark, or foggy the standard AI system becomes blind and incapable of detecting any obstacles.
We decided to mimic biology in our research conducted at the University of Valencia. Instead of subjecting AI models to images of millions of road conditions, we chose to emulate biological processes.
Why, biologically speaking can humans see well in such diverse conditions?
Click here to read more
What we see and what we don’t, tells us much about our consciousness
Volume control in the brain
Neurons do not operate alone in our brains. Neuroscientists refer to this form of adaption as divisive normative normalisation.
For a better understanding (without going into math), we can imagine it as a “volume control system” with neuronal teams working together. Imagine that one neuron is focusing on a dark part of your field of view, like a car in the night. This weak signal is amplified by neighbouring neurons, which increases the volume.
The same happens if we turn the brightness down. To prevent being dazzled, the brain lowers volume.
It is this mechanism that allows us to see and adapt in many different conditions. Modern AI systems, in their quest for accuracy and speed, have forgotten this biological source of inspiration.
Click here to read more
AI and human systems’see’ things differently, and this is why AI images are so garish.
Driving simulator with AI
We used AI to process images in our study. Then we added layers that simulated the “volume control mechanism” of the brain. We forced the neurons of these animals to adapt and communicate, much like our brains.
To find out if cars would be safer if they imitated biology, we ran a series of tests. We tested both the standard AI model and our modified brain-inspired version.
We compared responses using databases of real-life driving data from European cities and night driving photos from Switzerland.
Imitating the brain’s functions was successful. Both AI models were able to drive well after being taught, but when fog and darkness entered the equation, they began failing. The AI model lost its ability to differentiate cars and buildings from roads.
On the other hand was the AI system equipped with our brain inspired mechanism. It performed better in complete darkness or fog than the original AI system.
From the inside we analysed how the new system perceives the world, and it did exactly what was expected.
The system was enhancing and capturing the fine details of cars hidden by fog, which would have otherwise been invisible. Its performance improved as a result of the weather changes.
Click here to read more
Next-generation driverless vehicles will need to not only see the road but also think about it.
Nature teaches us a lot
The safety of pedestrians and passengers in autonomous cars are major factors to consider when trying to convince society at large of the benefits of AI. Smart systems must work in ideal conditions. They must be safe and reliable in real life, in any weather condition.
Our research suggests that artificial intelligence may be more accessible than we think. It is not necessary to have more powerful computers, or much larger amounts of data. We can learn a lot by looking at how our brains have evolved over millions of years.
Nature has often already found solutions to many of the issues that artificial intelligence is currently facing. It’s up to us to take the lessons from nature.
Weekly e-mail featuring the expertise of scholars and researchers. The newsletter introduces the diverse research that is being done on the European continent, and also considers the most important issues for European countries.
Subscribe to the newsletter.
