‘Passive’ visual experiences play a key part in our early learning experiences and should be replicated in AI vision systems, according to neuroscientists.
Italian researchers argue there are two types of learning – passive and active – and both are crucial in the development of our vision and understanding of the world.
Who we become as adults depends on the first years of life from these two types of stimulus – ‘passive’ observations of the world around us and ‘active’ learning of what we are taught explicitly.
In experiments, the scientists demonstrated the importance of the passive experience for the proper functioning of key nerve cells involved in our ability to see.
This could lead to direct improvements in new visual rehabilitation therapies or machine learning algorithms employed by artificial vision systems, they claim.
The passive visual experience has potential clinical implications, for the study of new visual rehabilitation therapies, and technological implications, where it could lead to an improvement of the learning algorithms employed by artificial vision systems
‘The development of artificial visual systems currently uses mainly “supervised” learning techniques, which require the use of millions of images,’ said lead researcher Davide Zoccolan, director of the visual neuroscience lab at the SISSA research institute in Trieste, northeastern Italy.
‘Our results suggest that these methods should be complemented by “unsupervised” learning algorithms that mimic the processes at work in the brain, to make training faster and more efficient.’
Humans’ very first visual experiences in the womb play a key role in teaching the brain to ‘see’ and are fundamental to vision development, the experts say.
From the early stages of gestation, our visual system is subject to continuous stimuli that become increasingly intense after birth.
These stimuli are at the centre of the learning mechanisms that, according to some theories, are fundamental to the development of vision.
The study shows the importance of passive visual experience for the maturation and the proper functioning of some key neurons involved in the process of vision
But there are two different types, the team argue, and one has been perhaps prioritised at the expense of the other.
‘Learning comes in two flavours – either “supervised” – guided by a teacher – or “unsupervised” – based on spontaneous, passive exposure to the environment,’ said Zoccolan.
‘The first is the one we can all associate with our parents or teachers, who direct us to the recognition of an object.
‘The second one happens spontaneously, passively, when we move around the world observing what happens around us.’
These two types of stimuli aid development of the visual system – in particular, the maturation and the proper functioning of some key neurons, or nerve cells, that are involved in the process of vision.
For their study, Zoccolan and PhD student Giulio Matteucci studied the role of spontaneous visual experience and, in particular, the role of the ‘temporal continuity of visual stimuli’ – a constant and immersive visual experience over time.
‘By temporal continuity we mean the typical, smooth transformation of the visual input that takes place during natural vision,’ said Zoccolan.
‘When we look at the world, we typically see things that change in the retinal image smoothly – for instance, because they move from one place to another place of the visual field or because they move closer to us, thus becoming gradually bigger in the retinal image.’
Temporal continuity is considered fundamental for the maturation of the visual system by some theoretical models that mathematically describe the biological learning processes.
In experiments at the lab, the researchers exposed two groups of young rodents to different visual environments in video clips on a daily basis.
Rats were shown a series of videos in either their original format or randomly shuffled single frames, which destroyed the temporal continuity of visual experience.
Much of what we will be as adults depends on the first years of life, on what we simply observe happening around us and not only on what we are taught explicitly
Rodents exposed to the discontinuous visual flow we showed impaired of the maturation of some cells of the visual cortex in the brain, called ‘complex’.
‘These neurons play a key role in visual processing – they allow recognising the orientation of the contour of an object regardless of its exact position in the visual field, a perceptual ability that only recently has been implemented in artificial vision systems,’ said Zoccolan.
‘Having shown that their maturation is highly sensitive to the degree of continuity of visual experience is the first direct experimental confirmation of the theoretical prediction.’
This finding reveals how important passive visual experience is for the development of the visual system.
‘The kind of learning process mediating such development is called “unsupervised learning” – a term often used also in the field of machine vision and artificial intelligence – given its spontaneous, passive nature that does not require a teacher guiding explicitly the learning,’ said Zoccolan.
Some forms of spontaneous or passive learning are the basis of elementary visual function developing, while other forms of learning, such as those experienced in school, only come into play later during the acquisition of more sophisticated skills, Zoccolan and his colleague argue.
As well as better AI and machine learning, awareness of the passive system could help develop new visual rehabilitation therapies for children with sight problems.
‘In some developing countries, there are children suffering from congenital cataract, who, after the surgery to remove it, have to develop substantially from scratch their visual recognition skills,’ said Zoccolan.
‘Already today, some rehabilitative approaches exploit the temporal continuity of specific visual stimuli – for example, geometric shapes in motion – to teach these patients to discriminate visual objects.
‘Our results confirm the validity of these approaches, revealing the neuronal mechanisms behind it and suggesting possible improvements and simplifications.’
The research has been published in Science Advances.