📈Machine Learning
How machines learn from data · 73 terms
ML Basics
24The words every machine learning project uses. They name the data, what goes in, what comes out and the settings.
A way to build AI. The computer learns the rules from data. People do not write them.
The examples a model learns from. The model is only as good as this data.
One fact about a case that a model reads. In a table, it is one column.
The right answer that comes with an example. The model learns to guess it.
A number inside a model that training learns. Together they hold what it knows.
A setting you pick before training. It shapes how the model learns.
Train a model many times, each with new settings. Keep the best.
Data Splits
4How the data is cut into parts for learning, tuning and the final test. Also how that goes wrong.
The part of the data a model learns from. It is usually the biggest part.
Data kept out of training. You use it to pick the best settings.
Data kept locked away. You use it once, at the end, for a fair final score.
When the answers sneak into training. Test scores look great and real use fails.
Data Preparation
6Getting raw data ready to learn from: tagging it, turning it into numbers and filling the gaps.
Each example gets a tag with the right answer. The model learns from it.
You make new, better inputs from raw data. Then the model learns with ease.
Turns a word into a row of 0s with one 1. Now a model can read it.
Puts number columns on one range. Then big numbers do not drown out small ones.
Make more examples by changing the ones you have. Flip them, crop them, tint them.
When one answer is much more rare than the rest in your data.
Learning Types
7The ways a model can learn, sorted by where the right answers come from.
The model learns from data. Each row has the right answer on it.
Learning from data with no answers on it. The model finds the patterns by itself.
The model learns from a few rows with labels and lots of rows with none.
The data makes its own answers. The model guesses a hidden word, then checks.
Learning by trial and error. Good actions earn rewards, so the agent repeats them.
Reusing a model trained on one job as the head start for a new job.
One model learns on many phones. Only the changes travel. The data stays put.
ML Tasks
6The jobs machine learning is used for, named by the kind of answer they give.
Sorting each case into one of a few known groups, like spam or not spam.
Guessing a number from the facts you have. The number might be a price or a time.
Putting things that are alike into groups. Nobody names the groups first.
Finding the rare cases that do not fit, like a fake charge on a bank card.
It squeezes many columns down to a few. It keeps most of what counts.
A system that guesses what you will like next. It learns from what you and others chose.
Classic ML Algorithms
12Well-known methods that learn from tables of data. They range from straight lines to forests of trees.
Guessing a number with a straight line. The line is the one that best fits the data.
It gives the chance that something is true. Then it turns that into yes or no.
A fast model that sorts things by adding up the odds from each clue.
A model that asks yes or no questions. Each answer leads to the next one, until it is done.
Lots of decision trees, each a bit different. They all vote on the answer.
Many models each make a guess. Joined up, they beat most single models.
Many small trees built in a row. Each one fixes what the last got wrong.
A fast, free tool for boosted trees. Often the best pick for table data.
It finds the closest cases it knows. Then it copies what most of them say.
It splits data into k groups. It moves k centre points until they stop.
Squeezes many columns into a few new ones. It keeps as much of the spread as it can.
It splits two groups with a line. The line leaves the widest gap it can.
Neural Networks
21Models built from layers of small math units, and the parts they are made of.
Small math units in layers. Each one passes numbers on to the next.
One tiny unit of a network. It weighs its inputs, adds them up and passes it on.
One row of neurons. Data goes through each row in turn.
A number on each link in a network. It sets how much an input counts.
A rule inside each neuron. It decides what number gets passed on.
Turns raw scores into chances that add up to 100%.
Machine learning with lots of layers. Most AI you use today runs on it.
Network Training
7How a neural network learns. It measures its error, then steps its weights toward a better answer.
A score of how wrong a model is. Training tries to make it small.
How training improves a model. Each small step nudges the weights downhill.
The rule for how a network moves its weights at each step as it learns.
It traces the error back through the layers to see which weights to fix.
How big each training step is. Too big jumps around, too small crawls.
One full pass through all the training data. Training runs many of them.
How many examples the model reads before each update. Often 32 or 64.
Network Types
7The main designs of neural network. Each one is built for a kind of data or a kind of job.
The most plain kind of neural network. Each layer links to all of the next.
A network made for photos. It slides small filters over the picture.
A network that reads in order. It keeps a short memory of what came before.
An RNN with gates that choose what to keep and what to forget.
Two networks in a contest. One makes fakes, the other tries to catch them.
It starts with random dots. Step by step, it cleans them up into a picture.
A network that packs data into a small code. Then it builds it back.
Model Fit
7Whether a model has learned the real pattern, too little of it, or the noise as well.
The model learns its old data by heart. Then it fails on new data.
The model is too simple. It does badly on old data and new data alike.
How well a model does on new data. It is the whole point of training.
Too simple misses the same way each time. Too complex jumps around.
Ways to stop a model from memorising. They keep it simpler.
Random neurons switch off in training. The rest learn to cope.
Stop the training when the score on new data stops going up.
ML Frameworks
3The free code libraries people use to build, train and run machine learning models.
A free tool in Python. You use it to build and train neural networks.
A free tool to build and train neural networks, then ship them to apps.
A free Python tool full of classic models. You use them all the same way.