Contributed by: Matei Mancas, Faculté Polytechnique de Mons

Attentive Computers: What For?
Attention is so natural and so simple: every human, every animal and even every tiny insect is perfectly able to pay attention. In reality as William James, the father of psychology said: “Everybody knows what attention is”. Why in this case should one study something that everybody knows? And why adapt attention to machines? That is precisely because everybody “knows” what attention is that few people tried to analyse it before the 19th century.

Attention is sometimes conscious (for complex living forms only) but most of the time unconscious and it is the key to survival. Attention is also a sign of limited computation capabilities. Vision, audition, touch, smell or taste, they all provide the brain with a huge amount of information. Gigabits of rough sensorial data flow every second into the brain which cannot physically handle such an information rate. Attention provides the brain with the capacity of selecting the main information and building priority tasks.

Attention of live beings largely contributes to the brain computation optimization and its importance is crucial. However, this key process of attention is currently rarely used within computers. As with the brain, a computer is a processing unit. As with the brain it has limited computation capabilities and memory. As with the brain, computers should analyse more and more data. But unlike the brain they do not, or rarely, pay attention.

That is why a new transversal research field appeared for a few years gathering engineers and computer scientists, psychologists and neuroscientists: computational attention. Beyond the theory of attention, there is a wide application field and this new research domain may have an important impact on future. The aim of computational attention is not to replace classical signal processing techniques but to complement them in various situations.

Attention: a step towards intelligence
Few references to attention were made through history. Such a natural process, partly unconscious, did not capture philosophers’ attention. Some thoughts about the attention concepts may be found in Descartes, but no rigorous and intensive scientific study was done until the beginning of psychology.

There is no widely accepted definition of attention because of the diversity of the disciplines which focused on it. Even if, at the beginning, only psychologists studied attention, its huge importance led other specialists like philosophical, clinical, anatomical, physiological and even computational scientists to provide their own definitions of attention. Nevertheless, John. K. Tsotsos suggested that the one core issue which justify attention regardless the discipline, methodology or intuition is “information reduction”.

Similarity may be used by the brain to obtain several steps of information reduction and comparison may represent the process of information understanding. A meaningful structure comes out from the rough initial information by comparisons at several scales:

• Comparison within the acquired basic signal components
• Comparison between groups of basic signal components
• Comparison between groups of basic signal components and the working memory. More complex groups may rise here from the acquired signal
• Comparison between complex groups of signal components and the long term memory which leads to a semantic classification

Through these comparison steps, attention reduces information by transforming it into a meaningful structure. The concept of attention may be defined as the transformation of a huge acquired unstructured data set into a smaller structured one while preserving the main information: the attentional mechanism turns rough data into intelligence and, for sure, there is no intelligence without attention.

Bottom-up and Top-down competition
Attention is a continuous competition between a bottom-up approach which uses the features of the acquired signal and a top-down approach which uses observer’s a priori knowledge about the observed signal. Comparisons between basic or groups of basics signal components may lead to bottom-up attention, while comparisons with the long-term memory may mostly lead to a top-down influence.

The influence of bottom-up or top-down approaches depends on how familiar the acquired signal is for a given observer.

Computational attention: a step towards artificial intelligence
The purpose of computational attention is to automatically predict human attention. An example may be seen for images (visual attention) in the following figure:

Similarly to the fact that attention is the beginning of intelligence in biology, computational attention may be the starting point of artificial intelligence in engineering applications. Computational attention provides machines with human-like reactions and behaviours and let them free to make decisions even in unexpected situations:

• A computer which pays attention is able to be surprised and interested in novel data
• A computer which pays attention is able to understand novel situations and to choose the important data it will learn

Applications of computational attention are numerous and among them we may cite:

• Automatic pathology detection and localisation in medical imaging or defect detection in machine vision applications
• Automatic feature selection for supervised learning
• Automatic detection of auditory events
• Smart signal filtering, coding and transmission through networks with unpredictable quality of service
• Image ergonomics as automatic zooms on smart phones, advertisement efficiency quantification, street lightning efficiency quantification
• Video tracking
• Image registration
• Object recognition
• Detect interest points and match already seen objects in robotics
• Automatic focus on digital cameras
• Natural HCI (Human computer Interfaces) as avatars

More generally, when a set of data is acquired and it should be processed, it may be interesting to reduce it by keeping only data which may attract attention.

Want to know more?
Attention web page