Source: Deep Learning on Medium
The Deep Learning Series: Part I
When the thought of computers and artificial intelligence was first on our mind, it was scary how computers will replace humans in everyday tasks and how we’ll all be out of jobs and as I’ve seen a part of media is looking to do exactly that. I remember reading an article (I guess Forbes) that said will robots replace you in the future? Well, that is an interesting question and I would like to quote Peter Thiel on that note “Computers are compliments for humans, not substitutes”.
In this series I will begin my journey to a very small part of Artificial Intelligence i.e. Deep Learning covering through the maths, basic machine learning moving on to the current state of Deep Learning and then finally to future research in the field.
I will also try to provide a philosophical view on it rather than just providing a core scientific view. This might be entirely based on the book that I’ll be reading at the time because I generally tend to forget what I have read in the past. (I’m trying to improve my habit on that). So, at the end of every article, I’ll be mentioning the sources that might influence my thinking on the subject. All throughout the journey, I will try to write these articles as if I’m teaching somebody and not just narrating it.
If you were to ask What would you choose? Computer or Human, if you’re not given any context of what the task is, think to yourself what would you choose?
My guess is most probably a human because humans are intelligent and have a way of figuring out complex tasks on their own given enough time but computers don’t. Now surprise surprise, the task is to calculate multiplications of 2 numbers quickly. Now, in this case, if you don’t pick a human calculator, the person who picked computers will win. That’s a task of computation. Let’s talk about another task, say, you have to see an image and identify some objects in it. Well, you’ll think that’s a very easy task for a human and it is. A computer, however, will lose this task.
Computers have speed, significant speed, like faster than Lightning McQueen from Cars but only in a task they know and they are limited to tasks they know. Now, what if we have a billion images to label if you want to think of a real-life example, imagine you’re trying to solve a robbery and you have the face footage of the person who robbed but now you have to look through thousands of faces in your database to see if one matches. I don’t know about you but after seeing 100, I would actually forget what the robber looked like. In this case, even though humans are better at extracting information from images, it’s not feasible for humans to do so.
So what if we could teach computers who are extremely fast to recognize and compare faces and find the robber in minutes?
How would we begin to do that?
Let’s begin with a simple task first and then move on to more complicated ones. Let’s just start by finding a car maybe in an image. Now, do note that computers perceive an image as just numbers in the form of RGB (many other formats exist but let’s assume for simplicity)
If you think about it, how do we recognize if there is a car in an image? Think about it for a minute before moving on.
You might think that we see a wheel, we know what a car’s shape looks like, we know a car has doors, windows, etc. These are all correct but if you really think about it hard, a cart also has wheels, we have different shapes of cars, a house has windows and doors, so how do we teach computers to recognize a car.
What really distinct’s a car from other objects in the universe? What precise concrete things we can say about a car that a computer can understand?
To be honest, I don’t think we can really, to make a 4-year-old understand what a car is you just show “look that’s a car” and the kid understands by observing carefully.
Let’s do that with computers, let us let them figure it out on their own and maybe, if we show them enough pictures of cars, they might start recognizing it. Hence, the field of artificial intelligence steps in. In reality, we don’t need to teach a computer rather just provide the tools for it to learn itself.
Now, deep learning is the field that has been making the most advancement here. It’s quite simple actually but then all complex tasks are quite simple at the end. We just have to find a combination of linear algebra functions that work together in layers to form a representation of the image and to generalize the representation to all cars and not just a single car image but lots of car images.
Now, that’s all I want to say for the first article, next up I’ll be writing about linear algebra and things will start to get mathematical. So, if this is something you feel like following, hit that clap and follow me for more.
Zero to One by Peter Thiel