Oct 26 — 2020

Artificial Intelligence Machine Learning and Deep Learning: A fruitful chat on buzzwords

Intro

In this article, we are sharing our vision of Artificial intelligence (AI), Machine Learning (ML) and Deep Learning (DL), including an overview of what lies behind these technologies and practices, and providing some real examples of their application in business. They are interrelated fields: ML is a subfield of AI, and DL is a subfield of ML:

machine learning

At Bedrock we have consistently seen that C-level leaders often ask and talk about AI, and it’s common for them to hold high, often way too high, expectations, and therefore not surprisingly, misunderstandings. Our objective with this article is to go through these well-worn concepts again, but this time in the form of an easy-to-digest tale.

Recently, I had a chat with my good friend, Alice Smith, who has been leading a renowned steel manufacturing company for over 20 years, and she asked:

 

  • I have been hearing a lot about AI, ML, and other fancy words lately, but I’m not sure I really understand what any of that is. And more importantly; should me and our board pay attention to these at all?
  • Well, Alice, we should first talk about the differences between how a machine and how a human think. Actually, machines don’t really think at all,although it may look like it when very complex models are applied and that’s exactly what AI is: advanced models that were programmed trying to mimic the human way of thinking. Sometimes these programs are highly successful, even capable of passing the Turing test, which is a method for determining whether a machine is thinking like a human being or not.
  • Ok, I get it; so machines think in a way that might look human, but it’s actually different. Still sounds weird though, how is this possible?
  • AI is currently able to make predictions, mainly through the use of statistics and algebra, combined with large amounts of historical data. All the processes followed are deterministic, although they have reached such a high degree of complexity that they make us feel that machines are showing some sort of creative-thinking behaviour.
  • OK, now I get it, it’s just mathematical models. And now we’re able to apply them because computers are more powerful and cheaper than they used to be. Another question: can AI help a traditional manufacturing company like ours?
  • For instance, let’s assume that you´re planning to set-up a new factory for your company. You have to decide where to place the entrance, every machine and electrical outlet, and much more. You even have to consider if that investment is going to be worth it based on the potential return.
    You can do this using classical industrial and econometric models, like we´ve always done. But you could also go one step further, and do it using Artificial Intelligence models. If fed with the right data, AI models can make much better predictions, allowing you to optimise your investment.
    This is just an example, but there are many more, like finding the right composition of materials for a packaging, or the best temperature for a process.
  • But are all kinds of AI the same, or are there many types of AI?
  • Not at all. For starters, there are two types of AI: narrow and general. Right now we can only use narrow AI, which only performs the exact specific tasks it was designed for, and in the way that it was designed to perform them. General AI would work in a similar way to the human mind; flexibly and independently, adapting and learning to carry out completely new tasks with experience. The latter is not possible at the moment, and doesn’t look like it will be in the near future so you don’t have to worry about advanced robots taking over civilisation!
    There are also many different AI algorithms, tailored and specific for each kind of problem and there are subcategories of these, like Machine Learning or Deep Learning
  • Ok, since we’re already on the subject, those are words I’ve heard, a lot, but I don’t really know what they mean. What does ML entail?
  • I would start by asking you a question, do you know what an algorithm is?
  • Yes, it´s like my wonderful grandma’s recipes, but for computers, right?
  • Exactly! You have a clear set of instructions that you apply to your ingredients, but on computers your recipe is your algorithm, and the ingredients are, most of the time, the data you feed your algorithm with.
  • Now it will be easy to tell you in simple terms what ML is by comparing it to a traditional algorithm. Think of it as algorithms being able to evolve over time, learning from the past whilst adapting to changes. It is as if your recipe evolves, and adapts by itself; just imagine your recipe automatically detecting the flavours and foods that your family likes best (learning), and changing itself (adapting) to modify some ingredients accordingly.
  • Would ML technology be useful for us?
  • Sure, there are countless applications that will work incredibly well already. Imagine you have different items on your conveyor belts, that need to be classified and routed accordingly. In a traditional approach you have to specify very precise rules to identify a route for all possible items, which would require an exhaustive list of possibilities and ranges, and once an unlisted item crosses the identification mark, it would have to be treated as an “unknown item” and therefore discarded. By implementing ML, you can make your classification model ‘learn’ about the kind of items, avoiding hard coding all possibilities. The ML model will then figure out the appropriate route even for unforeseen items with great accuracy.
  • That would help us save a lot of manual programming, effort and money!
  • Indeed 🙂
  • What about Deep Learning then?
  • Well, Deep Learning is a specialised subset of ML. They may look like they’re the same because the differentiator is completely invisible to an external viewer: it uses Artificial Neural Networks as the enabling technology. As the ‘neural’ in the name suggests, they basically try to mimic brain biology functions to perform calculations.
  • Can you please explain, how could you mimic a brain?
  • Well, I guess you’ve heard something about a biological neuron, right?
  • It’s the brain cell, right?
  • That’s right. In extremely simplistic terms, a human brain possesses a network of interconnected neurons. A neuron receives signals from other neurons, and the way they respond to the received signals is based on the strength of signals incoming from other neurons. In the brain these communications are chemical and electrical. Some people have come up with a simplified abstract model of this, which results in sets of algebraic formulas called neural networks.
  • Ouch, this sounds quite complex.
  • Well, it is… and maybe it’s best if we meet some other time to discuss it further.
  • Actually, I’m glad you proposed that. But even if we don’t go over this today, I’d like to know why we go over the trouble of complicating data models so much.
  • It may sound surprising to you, and in fact it was surprising to the teams who did research on this. For some applications, like recognising handwritten numbers, they provide much better results than traditional machine learning models. They don´t solve different problems to other ML algorithms, but they solve the same ones better.
  • Ok, I understand, so I could just use this to solve the classification and routing problem in conveyor belts that you told me before.
  • You could, yes.
  • And before we finish, my main concern has always been: won’t it be a threat to our employees´ jobs?
  • Alice, I see why you might think that, and many people ask the exact same question. It’s true that some jobs will be replaced by automation, but there are also many new jobs that will be created, and some experts argue that these could even outnumber the traditional jobs lost. There need to be people developing and monitoring the algorithms, selling them, and even explaining them to people, like I’ve just done with you. And not only technical people, to develop every algorithm there needs to be someone with high business domain knowledge to assist the team with their implementation
  • Ok, I guess there is no need to worry as long as we create a data science culture at our company. Out of curiosity, where do you see Artificial Intelligence going in the next few years?
  • It will surely continue growing at a fast pace based on what you can see these days. More and more problems, across all industries, are being solved using AI, taking advantage of affordable computational power, easily scalable cloud solutions, and huge amounts of accessible data. If I am honest with you, at Bedrock every month we kick-off more projects than the month before, and this is a self-feeding circle; as more data initiatives are finished with validated results, the news about the transformative potential of data and AI keeps spreading to more executives, like you, who become interested in how we can assist their businesses.
  • You’ve got me. I’ll swing by your office next Monday so that I can meet your team to start discussing how you could apply DS for some of our needs.
  • To be honest, I wasn’t even trying to get you to work with us, but you’re more than welcome!