Our explainer video series tackles complex machine learning concepts in five levels of escalating difficulty, starting from a kindergartener and moving up to an ML expert.
Our first episode? Machine learning itself!
00:00 – Intro
00:38 – Level 1: Kindergartener
02:10 – Level 2: Teenager
03:38 – Level 3: Non-expert adult
05:55 – Level 4: Computer science major
07:41 – Level 5: Machine learning expert
A 5 year old
Have you ever seen a robot before? I’m pretty sure you must have they’re pretty cool, right?
In fact, I’m sure you must have one lying around somewhere in your bedroom or your toy chest. But let’s talk about the reason why robots even exist in the first place. Do you know how the very first robot was created? Robots were invented by people like you and me who at some point in the past got really really tired and very frustrated of doing the exact same thing over and over again, so at some point someone really smart said: “Hey! Let’s make a machine to do this one thing for us instead.”
And now some of those machines have gotten even better and even faster than we are at doing this one specific task. Now i’m also sure you must have heard that people sometimes say that robots are smarter than people, but a robot is not smart the way that you are or I am. They can look really smart though by repeating things that they’ve seen over and over again, so it’s a little bit as if you could train a robot or teach a robot how to get really good at this one thing that you want them to be good at.
That’s what scientists call machine learning.
I’m sure that you’re aware that technology is all around you, but i would say that the applications that you enjoy using the most, the ones that your parents are always trying to pull you away from, are the ones that use machine learning. So you might ask: What is machine learning and how does it make these applications so engaging? It’s actually quite simple! Machine Learning is simply the process of taking large amounts of data, finding patterns in these data and then actually predicting something useful based on this data.
So let’s take an example: look at Netflix. I’m sure you’ve watched a movie or tv-show on Netflix. What you might not have noticed is that, while you’re watching this, Netflix is recording different data about what you watched and what other people similar to you are also watching. Based on all this data Netflix predicts what movie you’re more likely to watch and enjoy watching in the future
Other applications also do something similar: for example, Instagram will recommend to you content pictures that you’re going to find most interesting based on what other people in your social groups have found interesting. Sometimes these predictions are perfect sometimes you love the content Instagram provides you or you love the movies that Netflix recommends you. But other times it just doesn’t make sense. You’re like: “Why was this recommended to me? I have no interest in whatever this is offering!”. However, researchers are working on this problem and they’re trying their best to improve the accuracies of the Machine Learning models, so that we can actually provide users with information that’s engaging and interesting
So they stay interested in using their applications!
A non-expert adult
A lot of people misunderstand Machine Learning and Artificial Intelligence today and frankly I blame the movies. After all, most movies written about AI are focused on AI death robots like the terminator, but Artificial Intelligence is a lot more commonplace and mundane than you might expect and many of us interact with AI systems on a daily basis without even seeming to know about it.
So what sort of successes have we had so far? Today we’re really good at something called narrow or specific intelligence. It’s the idea of doing one particular task really well. A good example of this is IBM’s deep blue, an AI chess robot, that was able to beat Gary Kasparov, the world’s best chess player at the time. But this is something that was done back in the 90s and even three decades later on, we’re still not much different from that. We’re really far away from something called Artificial General Intelligence, the idea of a robot being able to do many different tasks all at once.
So if we aren’t doing Artificial General Intelligence, then what are we doing today basically is we can consider a set of three criteria that companies typically evaluate before they approach a Machine Learning problem that first criteria is that they need to make sure that the problem they’re trying to automate is able to save them time and money. Companies aren’t going to try and automate something that they could better do manually or with human beings. Second, companies need to make sure that they have a sufficient amount of data available to them before they approach that problem. In that chess example the only reason why deeply was able to succeed was because they were able to create virtually unlimited amounts of chess games for that robot to train on. But without something like that, it would have been impossible for them to even make it close to making a good chess robot, Finally, they’re able to understand or companies are able to understand that Machine Learning algorithms approach complexity in a different way than human beings: for example, Machine Learning algorithms can do something like predictive maintenance or medical diagnoses and learn that in a much shorter time period than a human being might take. Whereas human beings can do something like walking around or talking which is a really difficult task for Machine Learning algorithms.
With these criteria in mind, you can better understand the type of applications you’re going to see in 2020s!
A CS Student
As someone who recently graduated with a master’s degree in computer science and who started a job in the industry, I faced many challenges which I would like to talk about today: the first one unsurprisingly was data collection and labeling, so finding out for a particular problem what data should be collected how much should be labeled who should label it and what’s the best way to solve that problem is a science on its own which they don’t really teach you about in the classroom.
Another thing that I feel in the academia is focused upon is arbitrary improvement of benchmarks and irrelevance of applications like question answering system or 6D pose estimation. I was so proud when I trained my first ImageNet model, but here we deal with much more complex problems.
And then another thing is that with open source data sets right you the data set is fixed you cannot have external data sources but here in the industry we can have multi-modal data meaning that, let’s say, for autonomous driving systems I can have image data and Lidar data and also sound data if I choose to use it.
And lastly, data and models are ever evolving, so things keep changing. You may find seasonal patterns in many of the datasets you use this may include something like the stock market and so we need to keep retraining our models.
So what’s the best way to do that and what models are particularly suited for that task is something to keep in mind.
A ML Expert
Let’s talk about the advancements in Machine Learning in the past decade. It’s undeniable that we have made significant amounts of progress in the past 10 years. Imagine 10 years back we were talking about classifying cats from dogs or classifying numbers and today we’re talking about understanding a complex scene in an autonomous driving setting.
A lot of these advancements could be attributed to the advancements in hardware. Now the bigger question is if Moore’s law were to hold true for the next decade (which i personally think is not going to be the case), is it going to push our models to achieve this superhuman level intelligence? Is it going to help our models pass the turing test?
I personally think ‘no’, unless we restrict ourselves from following some of the heuristics that we follow when building our Machine Learning models. You know how people say that we treat Deep Learning as sort of this black box. This specifically is not only a problem for the users, but for you as well if you don’t know how your models work. How are you going to build a better model?
Still don’t believe me? Think about hyper parameter optimization. We still do this in sort of a brute force fashion most of the times.
But on the brighter side of things we have seen significant amounts of progress in fields like unsupervised learning, meta learning, reinforcement learning or even like multi-modal learning which I’m a big fan of. But unless we see a significant amount of progress in the theoretical space of these fields, I’m afraid AI would sort of be this augmented intelligence rather than artificial intelligence in the next decade.