1. Start by focusing on the business problem, you want to solve, and how you want to implement the solution.
2. Don't be scared by the buzzwords (i.e. "If you don't use AI, you will be out of business in five years). Sometimes a standard report is fine.
3. Although you should start small by solving existing pains, be curious, and understand the potential for large scale solutions. Understand enough of the technology, that you understand the scale and type of problems that can be solved. But don't do this until you are done with 1. and 2.
Good question and good luck with it! If you are interested in further dialogue, feel free to schedule a call.
Best regards
Kenneth Wolstrup
AI is a broad category. But the first distinction to be familiar with is Elon Musk's AI vs. Google's AI. In other words: the AI of the movies is very different from the AI being developed in leading companies today.
In theory, it may be someday possible to make an actually "artificial intelligence" - a mind that can reproduce itself and has all the creative, linguistic and generative functions of the human mind. But no one is anywhere close to building this today.
This is very important to understand, because AI is not as unapproachable or intimidating as it initially seems. Machine Learning - which is what any company does that claims to do "AI" - is all about automated pattern recognition. AI/ML tools are ones that have been trained to recognize specific patterns in massive amounts of data.
Take, for example, face recognition. The first step - even before recognizing WHO is in a picture - is to recognize that there is a face in the picture. If there is no face, no recognition. And if there is some frog's head that looks like a face, the algorithm may get tripped up trying to match it to a person's face. If it indeed selects Donald Trump or Hillary Clinton as the most probably match for this frog face, you have a PR crisis on your hands.
This is how it works: offer thousands, millions of images tagged as either 'no-face' or 'face' - a '0' or a '1' - a 'null' or a 'match'. The null images have no human face present. The match images have at least one face present. Then run it through an image segmentation or object identification training algorithm to create a face detection algorithm. Then test it on images that were not part of the training set (these not-used-before images are called the 'holdout set') and evaluate the performance of the algorithm. If needed, train again and again.
You'll notice a couple things here:
1. You can make an algorithm for almost anything. Yes! You could make an algorithm to recognize the Iron Throne in Game of Thrones. There are many obscure algorithms out there for very specific purposes. Some brands I've talked to have requested algorithms for identifying a certain shoe in a massive image database, for instance.
2. There are many versions of the same algorithm. Yes - many person, face, car, horse, chair, airplane, etc. etc. etc. algorithms exist, all with varying degrees of accuracy. See this link (if Clarity will let you click it) http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php?challengeid=11&compid=6
3. It's all about probability. Yes. Ultimately the output of every ML algorithm is a probability of how likely a thing is to be the thing you want it to be. In other words, how close to 1.0 (certainty) a prediction is for whether a face is present.
Now, empowered with an accurate face detection algorithm, you can take the next step: face recognition. This will follow the same steps, except now you are creating training sets where 'null' means there is a human face that is NOT the target face (the person you want to recognize) and 'match' means there human face present is the target face.
If you want to find Donald Trump's face in a massive library of images, you'd first train an algorithm on a large array of images, some of which have Donald Trump's face. You'd want the 'null' images to be as diverse and vast in their coverage of non-Donald Trump faces as possible. Preferably, there'd be some very Donald-like faces in there, marked as not-Donald.
Notice here we have a layering of one algorithm on top of another: first, run the face detection. Then, run the face recognition. My current gig involves over 30 sequenced and parallelized (they run at the same time, separately) algorithms fo image and audio classification.
Not all AI is about images though. Some AI is for chatbots, which just means it's making a lot of predictions about the likely meaning of a text and outputting highly probably-appropriate responses. Some AI is audio analysis, for speech recognition and speech generation (like Amazon Alexa). Some AI is to drive cars - which is a many-layer-algo environment for image analysis and math about mass/velocity/trajectory of moving objects. Very hard to do.
It's important to know that the more you train AI, the better it will be. That's also why it's a long way to go before we have to be afraid of it. We're still training baby AI algorithms, in the grand scheme of things. Consider that a human baby can recognize a face - and even which one is its mother's - immediately upon exiting the womb. That is with ZERO face detection/recognition training. There's a lot we don't know about truly artificial intelligence.
Armed with this basic understanding of AI, you can collaborate with data scientists to uncover opportunities for the application of machine learning to large datasets in your possession.
If you have specific questions about applying AI to your business, don't waste time googling stuff. Let's chat.
