A beginner’s guide to demystifying the buzzword- AI

One of the words that’s being brandished like a sword often in the corporate world today is AI. Dare say Adobe Illustrator. Its Artificial intelligence. Why brandished like a sword and not just discussed about? Because for many, the lack of understanding of the potential and limitations of AI is often projected onto the digital experts who are expected to know all answers to AI.

AI is actually not a very new concept. Research on the subject began in the 1950s but it went through a phase of significant funding to negligible funding because of over estimation by researchers then. This happened again because of the absence of the necessary hardware. However, in the interest in AI increased manifolds in the 21st century along with breakthrough technological advances.

The reasons for it becoming a buzzword now are many actually. Here are a few that I know of.

Access to hardware platforms for AI- Tremendous progress in tech- based infrastructure like speed, cloud services etc. has enabled this today and so you have so many startups working on AI

Huge investments being made by some of the largest tech companies like Google and IBM in driverless cars, deep learning for strategic reasons.

There is big data that is available today across many industries that has resulted in demand for much more valuable and nuanced insights.

The advent of mobile personal assistants like Siri has also led to familiarity with the concept of AI among more people and not just corporations. So, this has obviously generated a lot of buzz around this word.

But it’s important to demystify this word- specially for many of us from a non- technical background and   understand it realistically so that we are clearer about what AI can and cannot do. Here are some myths and facts that I’ve been able to put together. Hope you find it useful.

Myth 1: AI is getting to a point where there are very intelligent robots coming over to take over our jobs

Fact 1: There are 2 separate ideas in AI- ANI (Artificial narrow intelligence) and AGI (artificial general intelligence). Today what we know of AIs functionalities are based on ANI – which enables a single function e.g. Self-driving cars, smart speakers, web search etc. AGI on the other hand enables a machine to do anything a human can do. That is the ultimate goal. However today all work is being done in the space of ANI. There is almost negligible work in AGI. So, humans being taken over by robots completely is still far away. At least those of us reading this can stop worrying about it.

Myth 2: Machine learning, data science, deep learning, and AI are the same…umm… similar

Fact 2: Machine learning is one of the tools in AI that is largely responsible for driving it. An example is when with a certain input A, we get the output B and the machine learns that and carries it out. For instance, online advertising uses this, where the AI system inputs some information about an ad, some information about the user and tries to figure out if this user will click the ad or not. Another example would be in manufacturing where AI is being used to identify defects during inspection. This type of AI is based on supervised learning. Supervised learning is not a new concept in AI but again it’s something that has grown by leaps and bounds, because over the years, the amount data that we have access to has grown manifold.  

Data science is nothing but generating insights from the data. – basically, the findings that are presented

Deep learning- It happens through something called artificial neural networks – basically a big mathematical equation that arrives at output B on the basis of certain inputs. This is a very effective technique for complex input-output systems -specially where there are many input variables.

Myth 3: Spend time getting data and only after we collect it perfectly, we can do AI

Fact 3: As data is getting collected, it should be shown to the AI team so that they can give feedback to the IT team on what types of data to collect. The earlier this feedback is taken the more robust is the infrastructure developed. So, if you are developing an AI based chatbot share with the AI team the kind of information you will be seeking and if it’s possible to get the desired insights areas from that.

Myth 4: We have so much data. We can definitely have an AI team generate value

Fact 4: While more data is usually better than less data for AI, it’s not a must that the AI team can make sense of it. So, if we are planning to invest in acquiring data its always better for the AI team also to have a look at what they will be getting and thrash out with them if it will be really valuable.

Myth 5: Ai can do almost everything.

Fact 5: Media tends to report only the success stories, rather than educating people by mentioning about failures also.

According to Prof. Andrew NG founder and CEO of landing AI and founder of deep learning AI, one rule of thumb that can be used to decide what AI may or may not be able to do is pretty much anything you could do with a second of thought.

So, for example when it comes to driverless cars it’s important to determine the position of other cars and that is what we as humans can do in less than a second. So, in AI this can be done through pictures and radar / sensor readings. However, if you want the car to understand the intention of a person waving at the car on the road, it’s quite difficult since people could wave differently with their hand and it can have different interpretations, for example to stop a car, to hitch hike or taking a turn etc. At this point of time AI is unlikely to be able to do because the number of ways people can communicate with gestures is huge. So, learning from a video or visual what a particular way of gesturing with the hand means is not a very simple concept for a machine to learn.

However, responding with an empathetic response to an irate consumer would not be really feasible at this point of time for AI

The second factor is the amount of data that is available. The more the data the greater the chances of feasibility. This would mean both input and output data.

So, we can conclude that machine learning tends to work well when you're trying to learn a simple concept, such as something that you could do with less than a second of mental thought, and when there's lots of data available. On the other hand, Machine learning tends to work poorly when you're trying to learn a complex concept from small amounts of data. 

So, it is always recommended to do some amount of technical diligence with the data to know if a project has potential to be an AI project.