Installment 1: AI Explained

Artificial Intelligence in the visual infrastructure-inspection industry

90% of the world’s data was created in the past two years (SINTEF, 2013), and more is generated each second. As a result of this exponential data growth, more and more AI-powered data management technologies are emerging, and expressions like Artificial Intelligence, Machine Learning, and Deep Learning are being used interchangeably.

Navigating this highly technical world is a challenge we all need to overcome, as Artificially Intelligent software-tools are making an entry in markets of all sorts.

This series of articles will focus on AI in the field of visual infrastructure-inspections and aim to providing an understanding of these technologies, their abilities, shortcomings, and possibilities.

Artificial Intelligence / AI 

Artificial Intelligence is a branch of computer science that seeks to replicate human intelligence in a machine, enabling it to perform tasks, which previously required human intelligence. 

There are two main types of artificial intelligence: narrow/weak AI and general/strong AI. As of 2020, we have yet to achieve anything more advanced than narrow AI.

Artificial Narrow Intelligence (ANI)

ANI has a narrow range of abilities and replicates human behavior rather than -intelligence.

Everything we know as AI today, falls under this category; voice assistants, self-driving cars, suggestive search engines, image recognition software or email SPAM filters.

ANI will be the topic of the second installment in this series.

Artificial General Intelligence (AGI)

AGI has abilities on par with human capabilities. It is capable of learning and applying those lessons to solve other problems. 

It is not yet clear when, and if, we will achieve this kind of AI. 

The likely capabilities and uses of an eventual AGI are discussed further in the third installment in this series.

A third type of AI-concept, which is more speculative, also exists:

Artificial Super Intelligence (ASI)

ASI is defined as a system more capable than a human and self-aware. 

Super-intelligent systems have long been a concern to forward-thinkers and a topic of science fiction movies. The likelihood of this type of AI coming into existence is debated.

To gain a better understanding of the Artificial Intelligence available to us today, this article will dive into some of the commonly used algorithms and how they work.

Machine Learning / ML 

Machine learning is one of many branches under the AI-umbrella, and one that has become increasingly popular lately, particularly in the visual inspection industry.  

The definition of Machine Learning is a model which makes decisions based on past data. Think of the SPAM-filter in your email inbox – determining which emails are SPAM and which are not, is the task of some established ML-algorithms. 

Machine Learning algorithms learn using a concept known as Transfer Learning, described below.

“So, say you want a computer to know how to cross a road, for example. With conventional programming you would give it a very precise set of rules, telling it how to look left and right, wait for cars, use pedestrian crossings, etc., and then let it go. With machine learning, you’d instead show it 10,000 videos of someone crossing the road safely (and 10,000 videos of someone getting hit by a car), and then let it do its thing.” - Ernest Davis, Professor at Computer Science at New York university

To put it into the context of the visual inspection industry, a Machine Learning algorithm would determine, from past inspection-data, weather forecasts and knowledge of encroachment-issues, which transmission-lines would have the highest risk of sparking fire within the next 5 months. 

ML, like AI, is also an umbrella-term covering many sub-sequent types of AI-models. Perhaps the most important one being Deep Learning. 

Deep Learning / DL

Looking at Deep Learning, we start understanding how AI-algorithms function. Deep Learning algorithms are the basis of most Machine Learning models that process visual data.

A DL-model looks at a huge amount of images, to learn how to correctly identify objects in those images (for a deeper understanding of DL in this specific context, this video is recommended). When training a Deep Learning model, you provide a large, diversified amount of labelled data, for the algorithm to process and specify markers. Then you let the model run through data on its own, allowing it to identify more markers of the object it is looking for. The more data you give the model, the more accurate it will become.

Deep Learning is built from layers of Neural Networks, which are at the core of many AI-algorithms. Neural Networks are structures made to imitate the brain, by creating a complex mesh of “neurons”. Each neuron can answer a simple question, and by creating several layers, the activations between the neurons, enable Neural Networks to turn many simple answers into one accurate, complex one. Fully explaining Neural Networks is not the purpose of this article, but for those interested in learning more, I recommend this video on the topic as a starting point. Let us put Deep Learning into the context of visual inspection data; Deep Learning algorithms could examine image-data and recognize trees, transmission-poles, individuals parts of those poles, and even discrepancies on those individual parts.

Now we move up a layer, to understand how Computer Vision uses Deep Learning models to detect- and act on discrepancies. 

Computer Vision / CV

AI-models that deal with visual data, are commonly grouped as Computer Vision. Computer Vision has been around for a long time, and used to rely on lots of manual data labelling, but today, CV is often based on a type of Deep Learning-algorithms called Convolutional Neural Networks or CNNs.

Computer Vision systems can recognize specific elements in visual data, and act based on the elements identified. This type of AI will be the topic of the following articles in this series.

Teaching computers to see

The role of visual data in AI-technology today is crucial, as Fei Fei Li, creator of ImageNet and pioneer in the industry said: “Understanding vision and building visual systems, is really understanding intelligence. And by see, I mean to understand, not just to record pixels.” (Macneal, 2015) 

Because the internet, which is our greatest source of data, is largely comprised of text and visual data, and because text is – for the most part - easy for a computer to understand, teaching computers to truly understand visual data, is the single biggest challenge at the moment. We have, in a sense, enabled computers to see, but not yet to understand. 

_

The next article in this series will dive deeper into the use-cases of Computer Vision in visual inspections, and take a look at where the technology is at today.

Article written by: Scopito