Artificial Intelligence, Machine learning,Neural Networks…Let’s try to be simple!
Steven Spielberg is a visionary! He released, A.I. 15 years ago and unfortunately, he did not raise much attention at that time. Today it would have been probably very different as it’s hard to escape not only AI but some of the buzz words coming along–Machine Learning, Deep Learning.... AI, Artificial Intelligence, is even becoming a public debate mostly around the potential threat it could bring to humanity. On this topic, please read the excellent post Will Machines eliminate us?
In the high tech world, AI has recently gained lots of traction as it seems to be a crucial way to deliver business value out of the big data as well described in Is Big Data Still a Thing?
But recently, a day after Microsoft introduced an innocent AI chat robot to Twitter it had to be deleted after it transformed itself into a very evil one. It shows that AI needs to be well trained otherwise it can easily go out of track. When doing business, this is the last thing you want to happen: it’s better to know where you’re going.
Our customers or prospects often have questions such as "what’s the difference between Machine Learning and Deep Learning?”, so I thought it would be interesting to put down some answers to these questions. It is definitely not academics answers but hopefully it could be helpful for true beginners in the field.
Buzz word #1: AI
AI (Artificial Intelligence) is an automated decision making system, which can in an autonomous way continuously learn, adapt, suggest and even take actions. Google car is a good example of AI. Boston Dynamics made real cool stuff too. Robots and AI systems cannot rely on computer programs explicitly integrating every possible cases and appropriated actions. At the very core, they require algorithms able to learn from experience. This is where Machine Learning comes in.
Buzz word #2: Machine Learning
Machine Learning, or ML for short, is the science of designing and applying algorithms able to learn from past cases (kind of a child within a computer). It is worth saying that you cannot predict something that has never happened though: no past cases, no prediction. If some behaviour exists in the past, you may predict if or when it will happen again. Sounds weird? Let’s take an example: You want to know which customers of yours are going to buy product A in the next seven days. You start from extracting past cases of customers who bought product A and who didn’t. Then comes the “learning” phase. It’s the phase when the machine creates profiles of customers who are likely to buy. Now, by looking at the current Customer data, the machine can compute the likelihood of an individual customer to buy product A within the next 7 days.
Questions you might want to ask can be very diverse. You could detect which sensor is going to breakdown, which transaction is going to fail… There is virtually no limit - stating the event occurred in the past and there is a link between past data and the question (unfortunately there seems to be no chance of winning the national lottery with Machine Learning; if the game is not rigged).
When training an ML algorithm, one has to deal with complex tasks such as optimizing learning speed, number of samples analysed at each training step, evaluating multiples sets of parameters for each algorithms… And some ML algorithms are more complex than some others, so, please, beware!
Buzz word #3: Neural Networks
Neural Networks are a type of Machine Learning algorithms. Like a biological brain, neural networks consist of interconnected units called neurons; they mimic the way the brain can learn, changing the local patterns of neural connections. Artificial Neural networks are often organized in layers, exchanging, transforming and transmitting information (data).
The training phase is usually time-consuming due to the complex structure and the large number of hyper-parameters. That’s why it is worth training such networks mainly in cases like voice recognition, pattern recognition within pictures, facial recognition, etc. where one does not have to re-train models quite often.
Buzz word #4: Deep learning
Let’s go deeper in the family... Deep learning is part of the Neural Networks field, which is itself part of the ML. It is the subfield dedicated to the study of stacks of Neural Networks and often involves many (many, many) layers. Increasing the number of layers and combining different networks allows to dramatically increase the predictive performance of the overall architecture. The bad news is the resulting skyrocketing time it then takes to learn such complex structures. This is a show stopper for many marketing applications as they require fast learning and adaptation cycles: one cannot afford to spend the tremendous amount of time (and money) for learning complex, and performing enough, deep learning nets.
For other applications, like the ones we mentioned before, it is worth investing in the computing infrastructure and human experts required. And, maybe, deep learning will be the key to understand the way we dream.
Artificial Intelligence is a large and complex field, and getting a basic understanding might seem an overwhelming task. To summarize: it makes use of a family of techniques (ML) and sub-techniques (Neural Networks) giving powerful tools (Deep Learning) to extract value from Big Data, and as such brings deep changes for businesses using it.
What’s interesting nowadays is that AI is more and more used in the industry as every businesses, even the smaller ones, can now generate zillions of data record.
Now, you might be in such a position you’re sitting on top of a mount of data. And you want to ask your Big Data “what benefits could they bring to my business?” or “which techniques should I use?” That will be the next topic: insights to understand which problems benefit from Machine Learning!
Intrigued? If you want to get a sense of PredicSis.ai, we made it easy for you through our demo page