After hearing about creating and implementing a model… what is a machine learning model really and how does it actually work?
In: Technology
Machine learning is just a mathematical equation with a lot of inputs.
The best Eli5 way I can illustrate is to imagine a radio. You can tune the radio to some station. As you get closer you can tell you’re almost there by how clearly the sound comes in. As you get closer you slow down until you reach the maximum clarity.
Machine learning is much the same. Except instead of 1 dial you have thousands, if not millions. But the premise is the same. We have some idea of what the output should look like. So what we do is put in the inputs and see what the static sounds like, the output. We then run an analysis based on what we want the output to be. This is averaged over the thousands of training data sets we give it. This tells us by how much we need to tune our radio.
The averaging works so that the further away you are, the greater the change will be. But then as you get closer you tune the radio slower, so you get more precise.
A machine learning model is set up like this.
Input -> x (weight) -> (sum of all inputs times their weights) x (another weight) -> output
Each connection is a weight, and each node sums the weighted inputs and then multiplies it by a weight. Then it uses this new number as the new input.
Every weight is a dial we can use to tune the radio. So it gets pretty complex pretty fast. As far as what it is actually doing, we don’t know. You can guess that maybe it’s looking at specific features of a picture or maybe it’s analyzing some pattern on a chess board, but then at the same time you can feed static into a handwritting ML and it will confidently tell you that it’s a 5. So we ultimately aren’t sure why it works.
The point is that it is just a mathematical function. It’s like a radio, except it has a million dials. If you can find the radio station you’re looking for, whether thats playing chess or telling you if that picture is a cat or a dog, then you’ve created an effective machine learning algorithm.
Saying a machine learning model is pretty broad.
It’s really any type of mathematical equation or process to group things.
The idea is that you have a bunch of data, you then come up with a model. Then it’s the computer’s task to figure out what to do to best fit the model using initially random numbers.
Once there is a “best” way to fit the model, this is referred to as a trained model.
You can then take a brand new piece of data and see what the trained model thinks about it. Is it accurately placed in a group of like things? Is a number of a field calculated which is close to accurate?
This is a huge field and there are many types of models from k-means to deep learning architectures. It really depends exactly what you want to know.
K-means is a very simple concept if you look into that. It’ll give you a concrete machine learning algorithm that you can either implement yourself with any programming language or you can simply call it from a library.
Without going super deep into the math, you basically take two sets of data…known inputs with known outputs, and you apply a general mathematical filter (“the model”) to the inputs to calculate the model output. Then compare that to the known output. Initially, the model output will be wildly wrong. Machine learning algorithms are clever ways of adjusting the filter based on comparing the model output to the known output to try and adjust the filter so the model output more closely corresponds to the known output. You iterate that lots of times until the model stops getting any better or you’re as accurate as you want to be. Now your model is “trained”…you know that, for the known inputs, it will produce outputs that match the known outputs (“truth”) to some degree. You assume that now, if you feed it new input, it will give you correct output.
An example…I want to train a machine learning model to pass those annoying captcha things that say “Click on all the pictures of traffic lights”. I collect a whole ton of pictures, some with traffic lights and some without, and I go through them myself and label them “yes” or “no” based on whether it has a traffic light or not. The set of pictures is the known input, my yes/no list is the known output. Then I run a machine learning model on that training data and the model “learns” how to recognize traffic lights by continuously adjusting a bunch of internal math values according to the algorithm until it’s pretty good at recognizing traffic lights. Now I feed it a new captcha that it’s never seen before…if my model is good, it will correctly detect traffic lights in the new photos too.
This makes machine learning *really* dependent on having good training data. Getting this wrong can cause your model to have all kinds of weird biases and errors that can be very difficult to detect.
I’m not super well versed in machine learning, but I have a good enough explanation that I feel like I can try and make it understandable.
A machine learning model is kinda like a really big mathematical problem that has a lot of variables. There are input variables, middle variables, and output variables.
The purpose of machine learning is to try to make that model “learn” what inputs lead to the specific outputs.
So, let’s say that you want to train a model to distinguish between oranges and apples. So, what you’d be doing would be giving the model an image of, let’s say, an orange, and telling it that it’s an orange. The model then tries to fine tune the variables in a way that it can be pretty sure that the image is of an orange.
You repeat this with a large number of images of apples and oranges and quiz the model every so often. The goal is to minimize the amount of incorrect guesses so that you can use it to, say, sort fruit into different bins in a factory.
Latest Answers