Chapter 2: The Basics Of AI. What Is AI? Why Now?

What Is AI? Why Now?

Last week in Chapter 1 we focused on the importance of data and how it enables the use AI. This week in Chapter 2 we will introduce AI from a business perspective, and why it is so important.

When to use AI?

Now, the level of automation depends on the current state of the art in AI/ML models. Technology is advancing fast and what wasn’t possible only two years ago now is or is close to being possible. There are a number of factors that determine the possibilities of using AI to help in a given process.

Number of rules to make the decision

If there are just a few rules, it’s enough to use a rule-based system (if this, then that). When the number of rules to make a decision start to increase, at some point these rule-based systems become unmanageable. Too many possible failures and maintaining those systems will take just too much effort. This is the case of many legacy systems used in conservative industries such as banking or healthcare, where those mammoth backend systems have become just too big to fail, and are kept in use albeit the monstrous amount of hours invoiced by IT companies just to keep them running. When too many rules are required, this is a good playground for AI. Another situation where AI becomes inevitable is when there are simply no known rules, for example, in the case of fraud detection. Where fraudsters and scammers keep coming up with new, never-seen before rules of fooling a given system. Obviously, a rule-based system won't work in a system where rules are just unknown.

Task difficulty

Closely related to the number of rules, as the more complex tasks have inherently more rules than simpler tasks, but different in the way that the person paid to perform the task will determine the return of investment (ROI) calculation baseline. In other words, simple or difficult, if the person paid to do the task has a high salary, then the ROI on tools to help that person will be shorter. It may be a better solution to hire more people with a low salary doing the task, if the salary is low and if the volume of tasks (see below) is manageable. But if a person is paid a high salary and spends his time doing time-consuming, low added value work, then there is definitely a space to consider if an AI tool can help, for example, discarding the high-volume obvious negative cases, and?showing the cases that need expert review.

Volume of cases

Obviously, if there is a very small amount of cases being processed, a more or less sophisticated AI model won’t be needed, and in the case of an extremely low amount of cases probably not even a simpler rule-based system. For AI to make sense, there needs to be a good amount of cases to be processed. So much so that stress starts to be a problem for the professional performing the task. AI systems don’t get tired, don’t have personal problems at home, will keep learning and improving their performance and will learn from mistakes as most of them are frequently retrained, and the quantity of data they are able to analyse is almost endless.

Technology?

Technology is advancing fast, AI models can do things now that weren’t possible only two years ago, especially in the fields of language and image processing. So when wondering if AI can do it, it’s not so straightforward to get an answer. As we will explain later in this ebook, uncertainty is an essential part of AI. There always needs to be a feasibility stage to answer the question: will it work? But many AI technologies have commoditized already and the feasibility is proven in many applications being used every day in millions of daily operations by tech companies.

In summary, in order to AI make sense to be implemented, these factors need to be in place:

1. Too many rules to handle, a rule-based system would become unmanageable.
2. The task is performed by a mid-to high qualified professional.
3. There is a high volume of cases.
4. The technology can handle the task.

Understandably, you may not know the answer to no. 4 (we are happy to help you with that), but do the remaining three ring some bell for you? Then you may be presented with the opportunity to unlock the value of AI. This is not a cheap line. This means you can truly transform the present and the future of your business with the help of AI, and aim for new levels of profitability (less costs), revenues (more sales), and as we saw before, an increase of the company value.

Automation Vs. Augmentation

When all of the factors above converge, we can use AI. But one thing is that AI makes sense, and a different thing is the level of automation. In some sectors where the margins are lower and the efficiency needs to be optimized to the maximum, hyper automation tools can fully automate given processes. In these cases, the algorithmic complexity is lower and it’s a numbers game, with a high amount of transactions performed. For example, classifying documents inside an robotic process automation workflow where a natural lounge processing (NLP) module classifies documents into three categories (invoices, claims, letter of intent).

When making a decision, there is usually a common type of case complexity classification. There are the obvious cases that a novel professional fresh out of school could detect on their first day of work, and there are complex cases that require the intervention of the most senior staff. These are obvious black and white cases. However, there is a big gray zone in the middle where it is not so easy to make the call. As we will see later, AI/ML systems output a percentage, a degree of confidence the model is assessing the chances of a given output and companies need to learn to work not with absolute values of zeros and ones, but with a percentage from 0-100%.

This can get quite subjective to each company and business nature as some companies may feel comfortable with a 70% chance of the model being correct, and others may need much higher assurances.?

Augmentation is a very realistic, “meet me in the middle” compromise. AI can’t take the responsibility of making the calls, but can provide augmentation tools to professionals so they can make the call faster, and spending less resources whether those are company resources or mental resources so they are working in a less stressful environment.?

In Etnetera, we believe in developing augmentation tools that will make the lives of professionals easier so they can work in a less stressful environment, with a better quality of life, and making less costly mistakes. The companies embracing AI tools will have happier employees, will attract more talent, and will reach higher levels of overall quality and productivity.

Data keeps growing exponentially

Ok so we have seen how AI can help professionals and companies with a qualitative approach. Using AI-augmented tools increases the quality of their work. But let’s take a look now at a quantitative approach. It’s simple math. The data generated, collected and processed in the world grows exponentially every year, and will keep doing so.

To give you an idea, the data generated worldwide in 2021 was 79 zettabytes (source). And how much is 79 zettabytes? Well roughly, it is equal to 1,5 trillion blu rays, 800 million of the world’s biggest hard drives (100 terabytes), 80 million human brains (1 petabyte)....

And it is expected to double every two years or so. Events like the Covid-19 outbreak only sped up this trend, and we are all online much more time than before covid.?

How are companies going to deal with so much data? The companies who are able to aggregate, track, monitor, listen, watch, observe that data will have more business opportunities than those who don’t. The more advanced the data-processing systems? become, the better their output will be.?

AI and ML are the only ways to make sense of such a large amount of data. Rule-based programming systems cannot handle this amount of data. It’s just impossible to hard code for so many different rules. Those systems would become unmanageable, unstable. Essentially the use of AI is inevitable. And the earlier companies adopt AI, the earlier they will jump on the winning train.

What is AI??

One of the best definitions of AI is “the capacity of a computer to perform operations similar to learning and decision making in humans”.?

A (tiny) bit of history

Artificial Intelligence has been in the imagination of humans since the early 20th century in science fiction and movies. Robots and machines that became sentient and could think and make decisions like humans, like Tin Man in The Wizard of Oz, or the humanoid robot Maria in Metropolis. Actually, here in Czech republic where this book has been written, we have deep roots with robots and AI. The term robot was coined by Kamil ?apek who wrote in 1920 the science fiction nobel R.U.R. that stands for stands for Rossumovi Univerzální Roboti (Rossum's Universal Robots) (good moment to mention that Czech Republic has a vibrating AI ecosystem with companies, startups and universities).

In the 1950’s a new generation of scientists, mathematicians, philosophers and alike bloomed. Among them there was English mathematician Alan Touring who made himself the question “can machines think?”, and published a paper discussing if and how machines could learn, how they could be taught (actually coined the term training for machines, with the punishment and reward concepts which have been an important part of AI in later years), and who later became the unofficial start of artificial intelligence as an academic discipline, making Alan Touriing the father of AI.

However, even if the theory proposed by Mr. Touring was later proven correct, computers back then couldn’t store data, and were incredibly expensive to run (plus, they didn’t have the computing capacity needed). During the next 20 years, there were many “slowly but surely” advancements, in the field of language processing especially. Optimism was high and expectations were even higher. In 1970 Marvin Minsky told Life Magazine, “from three to eight years we will have a machine with the general intelligence of an average human being.” However, while the basic proof of principle was there, there was still a long way to go before the end goals of natural language processing, abstract thinking, and self-recognition could be achieved.

Source

However, computing power and lack of data storing capabilities. Actually, it was only in the 1960’s when Moore’s Law was defined and has been right all along. In summary, Moore’s Law says the capacity of computers in terms of computing power and memory storage capabilities double every year. So eventually, computers catch up with AI and AI is able to make breakthroughs like when in 1997 IBM’s Deep Blue was able to defeat Gary Kasparov in 1997, and when Google’s AlphaGo was able to defeat Chinese Go champion, Ke Jie, only a few months ago. This has happened all along AI’s history, theorists predict AI’s capabilities but computing power and storage capacity are not there so the industry goes into a so-called “AI winter”. Since 2012 we are definitely living an AI summer, and it’s getting hotter and hotter (admittedly, this is not a fun example due to the current climate warming).

There are three different types of AI:

1. Artificial narrow intelligence (ANI): AI with a narrow range of abilities. All the AI applications you see currently in the market are narrow AI applications. They are trained to perform specific tasks and they can’t transfer the knowledge to learn new tasks on their own. For example, an anomaly detection model for fraud detection in the banking industry can’t be used to analyzing text, or recommending content. They are carefully crafted pieces of code engineering developed for specific purposes and each year there are significant advancements in their developments, making them more and more powerful every year. All this book is about ANI applications.
2. Artificial general intelligence (AGI): AI on par with human capabilities.?Artificial general intelligence (AGI) is the ability of machines to think, comprehend, learn, and apply their intelligence to solve complex problems, much like humans. AGI is able to replicate human level intelligence by being able to remember and transfer the knowledge previously acquired to solve new problems never seen before by the AGI. It hasn’t been developed yet, in other words it doesn’t exist yet, although a few organizations have set their minds to develop an AGI and are spending a tremendous amount of effort in research and development. Companies like Open AI (acquired by Microsoft) and like GoodAI here in Czech Republic, are paving the path to who knows what amazing discoveries they’ll unveil. In the meantime, the whole industry benefits from their R&D efforts and we are thankful for that.
3. Artificial super-intelligence (ASI): AI that surpasses human intelligence. This overlaps with the science fiction realm, as it’s a theoretical concept for now. However, this is where it gets a bit scary, since according to the Moore’s Law already described above, the machines could learn exponentially, which would turn them into something far beyond human comprehension, as we wouldn’t even be able to understand how they evolved. More like ants and humans (we would be the ants in this case). There are two main variants that researchers are proposing. Either ASI would eliminate the human race, or we would use it to transcend to other types of beings and different worlds, becoming half human half superhuman machines. Before you start panicking, please note that there are groups denying the feasibility of Moore’s Law, since the minimal distance of 2 nanometers cannot be overcomed at this point. Buy Quantum computing opens a totally new dimension to this, and as the normal physics laws don’t apply the same way in Quantum computing, there are fascinating developments on this regard to be unveiled in the following few decades.

How AI works

Think of an AI/ML system as a human brain. We receive stimuli (inputs), we have the cognition part where our brain processes the information (data), we generate an action (output) and we get better over time (learning).

Inputs

As for the inputs, it can be images or video, it can be sounds, it can be speech or text, or it can be just numbers. Each of these have become disciplines by themselves and here is a high level description of the most frequent applications.

Image and video processing

Image and video processing have advanced tremendously in recent years mainly due to the advance of the internet and social networks. Companies like Google, Facebook and Instagram have collected a huge amount of images that allow them to train big AI models with millions of parameters. The processing of an image is also called semantic processing, meaning the image processing models output some degree of understanding of the image. Based on difficulty, there are three main types:

• Classification: it’s a cat. The simplest task is classification, what am I seeing in an image?
• Detection. There is a cat here. Inside the image, the model detects the object and its localization. It can also be multiple detection.
• Segmentation: There is a cat, a dog, and a duck, sitting in the grass on a sunny day. This is the hardest task and it’s done at a pixel level. These models are for example, the ones in charge of self-driving vehicles. They can understand the whole image and make deeper interpretations.

Souce

Another very interesting field in image processing is image generation. For example, OpenAI’s Dall-e 2, or Google’s Imagen are text-to-speech tools that generate images based on text. They are creating huge opportunities in the fields of art and creative work.

OpenAI’s Dall-e 2

Google’s Imagen

For video processing, the same categories apply. The principles are exactly the same, only the required computing power is more in most cases. And because of this, the edge technology is also experiencing big advances in recent months/years. Edge technology runs the inference of the models (inference means the output of a model, that is, the answer the model gives to a given call) on devices. To process video, usually bigger AI models are needed, and those models need to run either offline in a local computer with a somewhat powerful GPU, or run in the cloud which needs an internet connection. In many cases both options are problematic and the most efficient is to run the models on the device. Microchip brands like Nvidia or Intel and device device manufacturers of video smart cameras offer built-in AI capabilities. They offer limited computing power but they can be quite effective in performing simpler tasks, for example real-time people tracking or detection of certain events.

For audio processing in AI, the techniques are usually the same as for image processing, as audio is usually converted to an image (using spectrograms) and once the audio signal is in an image format, it’s processed with the same AI models as images do.

Natural Language Processing (NLP)

How computers understand human natural language in written and verbal forms. The term natural language refers to the way we, humans, communicate with each other. Language is full of complex small variations and making sense of those would be just impossible with rule-based programming. For example, we humans can match the dots if someone makes a simple typographic mistake, or a misspelling but a rule-based programmed system would have a real hard time to overcome it.

Natural Language Understanding (NLU)

Whereas NLP could be understood as merely processing the text data as in just doing something with it, NLU purpose is to understand the data. What is the intent of the text? What is the intention of the person who wrote the text? What does it mean? Is it well written? For example, in a customer care center, the calls are first transcription from audio to text, and when is in text format the customer request is classified (it’s an urgent breakdown, it’s non urgent administrative claim, they want to talk with a specific person or department) and the NLU engine channels the request inside the system. Natural Language Generation

Natural Language Generation (NLG)

NLG generates text.It’s a complex field but the recent technology developments are allowing to develop production ready applications. For example NLG can write a summary of an article, or can write ghost stories, essentially it creates meaningful sentences that look like written by a human.

In real life, the most common applications of NLP are chatbots, text summarization, text categorization, parts of speech tagging, text mining, machine translation, ontology and others. It is a rapidly advancing field since there have been tremendous algorithmic advances lately. Tech giants like Google, Facebook, Microsoft and Amazon are leading the research. They have access to huge datasets and that allows them to experiment with large models. Some examples are OpenAI’s (largely owned by Microsoft) GPT-3, a language generation model trained on 180 billion parameters, or Google’s PaLM, or Facebook’s recently released Blender Bot 2.0.?

There are many subcategories or tasks covered by NLP but it is not the purpose to cover in such detail in this ebook. Maybe in later releases!

Numerical values

Many AI/ML models who work with numerical data as inputs. Numbers, databases, historical records feed models that can make predictions, recommendations, anomaly detection over large amounts of data. We will go through it in the next chapter.

Learning

Just like a human brain, AI models receive stimuli from the input data and learn from the data, detecting patterns in the data, and performing different actions. Most AI and ML models are made using neural networks.?

Difference between AI and Machine learning

Machine learning is a subcategory inside artificial intelligence, but for the sake of simplicity, in this book AI and ML refer to the same concepts, although as you can see in the figure below they are not exactly the same.

Figure XXX: DAI vs. ML vs. DL source

Machine Learning is a subset of Artificial Intelligence that uses statistical learning algorithms to build systems that have the ability to automatically learn and improve from experiences without being explicitly programmed.

The learning part happens when feeding the training data, the ML is able to induct (different that deduct), it means the model learns patterns within the data. Induction happens in supervised learning and deduction happens in unsupervised learning. The goal of a good machine learning model is to generalize well from the training data to any data from the problem domain. This allows us to make predictions in the future on data the model has never seen. There is a terminology used in machine learning when we talk about how well a machine learning model learns and generalizes to new data, namely overfitting and under-fitting. Overfitting and underfitting are the two biggest causes for poor performance of machine learning algorithms.

Neural networks and Deep Learning

Neural nets are a type of machine learning. They are computer programs inspired in the human brain and the neuronal nervous system (hence the name), in which a computer learns to perform some task by analyzing training examples. A neural net consists of thousands or even millions of simple processing nodes that are densely interconnected.

Figure XXX: Neural network architecture, source

Most nets are organized in layers of nodes, and each node is connected to different nodes in different layers beneath and above to which it sends data. To each of the incoming connections, the node will assign a number called “weight” and it will multiply the input value times its node weight to the output nodes. If the result of that multiplication is above a certain threshold, the node will not send data to the next layer. If the number is below the threshold value, the node “fires” which in modern neural nets design usually means sending the number (the sum of the weighted inputs) down to the output connections.

To train a neural network, a training data set is needed. It’s the data that will be used to train the neural network, hence the term training. A portion of data (a general rule of thumb in best practices is ? of the data) should be kept for validation. Now, the neural network will be trained to learn patterns from the data. When a neural network is trained, initially all weights and thresholds are set to random values, data is fed to the first input layer, which propagates forward (forward-propagation) to the next layers (there is also back propagation but we will leave out for now), and values get multiplied and added together in complex ways until finally arrives radically transformed at the output layer. During training, the weights and thresholds are continually adjusted until training data with the same labels consistently yield similar outputs.?

Deep learning is a term you may have heard or read about on many occasions. Now that you understand the basics of AI, ML and a neural network, it’s easier to understand that deep learning is learning by using deep neural networks, that is, neural networks with many (hundreds, thousands, millions) of layers, and the learning part happens through those deep nets in million, billions and even trillions of repetitions during the training process.

Supervised learning

In supervised learning we have input variables (x) and an output variable (Y) and we use an algorithm (or model) to learn to predict which will be the out based on many examples of inputs and outputs. When training the model, the training data used is labeled with the correct answers, so the model has a reference to learn from, like a supervision from a teacher, hence the name “supervised”. Supervised learning has two types of models, classification (what is it), or regression (based on the previous history, this is the most likely outcome).

Classification

A classification problem is when the output variable is a category, such as “Red” or “blue” , “disease” or “no disease”. Is this an apple or an orange? Is this email spam or not?

Regression

A regression problem is when the output variable is a real value, such as “dollars'' or “weight”. Based on his study time, is Tom going to pass the exam? Based on the purchasing history, is Tom going to buy this product? Based on the leakage history, is this pipeline going to break?

Unsupervised learning

Unsupervised Learning is used when the data is not labeled. Think of it as a big data set where we don’t know much about it, and the mission of these unsupervised learning models is to find patterns and trends in that data. It is called unsupervised because the models are left?

Without any supervision, they need to find similarities in that unsorted data. It is often used as part of exploratory data analysis, and it is most commonly used to find clusters and reduce dimensionality in the data. Customer segmentation, recommendations and anomaly detection are the majority of the use cases.

Reinforced learning

The model learns by experimenting, if the model does it well it gets a reward, and if it doesn’t do well it gets a “punishment”. By maximizing the rewards the model learns to find the best course of actions. Reinforcement learning is used in self driving cars, inventory planning, or other optimization problems.

Ok, now we can mark the theory part as checked. Next week we will cover real-life business applications and how AI is transforming businesses.