Grundlæggende om maskinindlæring for udviklere

I det nuværende teknologiske landskab forventes udviklere at have en række forskellige færdigheder. Og mange af dem gør det.

Der er også mange forskellige karriereveje tilgængelige for udviklere, der bruger mange af deres nuværende færdigheder med et lille twist.

Databaseadministratorer, advokater for udviklere og maskinlæringsingeniører har alle en ting til fælles med alle udviklere: de ved alle, hvordan de skal kode. Det betyder ikke noget, hvilke sprog der bruges, de forstår alle kernebegreberne bag at skrive god kode.

Det er en af ​​grundene til, at mange softwareudviklere overvejer at blive maskinlæringsingeniører. Med alle de tilgængelige værktøjer og pakker behøver du ikke have en dyb matematisk baggrund for at få nøjagtige resultater.

Hvis du er villig til at lære at bruge nogle biblioteker og få en høj forståelse af den underliggende matematik, kan du blive maskinlæringsingeniør.

I denne artikel vil jeg gå igennem nogle af de vigtigste begreber inden for maskinlæring, som du har brug for at forstå fra en softwareudviklerbaggrund.

Vi slutter med et eksempel på et helt maskinlæringsprojekt, fra at få data til at forudsige en værdi med en model. Til sidst skal du have tilstrækkelig viden til at gennemføre et lille maskinlæringsprojekt fra bunden.

Hvad er maskinindlæring?

Der er mange definitioner derude. Men maskinindlæring indebærer grundlæggende brug af matematik til at finde mønstre i store mængder data til at forudsige baseret på nye data.

Når den har fundet disse mønstre, kan du sige, at du har en maskinlæringsmodel.

Derfra kan du bruge modellen til at forudsige nye data, som modellen aldrig har set før.

Målet er at få computere til automatisk at forbedre sig med erfaring ved hjælp af de algoritmer, de får.

En algoritme er bare en matematisk ligning eller et sæt ligninger, der giver dig et resultat baseret på dine inputdata. Maskinindlæring bruger algoritmer til at finde de mønstre, vi leder efter.

Efterhånden som algoritmerne bliver udsat for flere og flere data, begynder de at foretage mere nøjagtige forudsigelser. Til sidst vil modellen, der er bygget af algoritmerne, være i stand til at finde ud af det rigtige resultat uden at være eksplicit programmeret til at gøre det.

Dette betyder, at computeren skal være i stand til at optage data og træffe beslutninger (forudsigelser) uden nogen menneskelig hjælp.

Maskinindlæring vs datavidenskab vs. kunstig intelligens

Mange mennesker bruger udtrykkene maskinlæring, datalogi og kunstig intelligens om hinanden. Men det er ikke de samme ting.

Machine learning er anvendt i data videnskab at gøre forudsigelser og opdage mønstre i dine data.

Datavidenskab fokuserer mere på statistik og algoritmer og fortolkning af resultater. Maskinindlæring er mere fokuseret på software og automatisering af ting.

Kunstig intelligens henviser til en computers evne til at forstå og lære af data, mens man træffer beslutninger baseret på skjulte mønstre, der næsten er umulige for mennesker at finde ud af.

Maskinindlæring er som en gren af ​​kunstig intelligens. Vi bruger maskinlæring til at opnå kunstig intelligens.

Kunstig intelligens er et bredt emne, og det dækker ting som computersyn, interaktion mellem menneske og computer og autonomi, hvor maskinlæring ville blive brugt i hver af disse applikationer.

Forskellige typer maskinindlæring

Der er tre typer maskinindlæring, du vil høre og læse om: overvåget læring, semi-overvåget læring og ikke-overvåget læring.

Overvåget læring

Dette er den kategori, hvor de fleste maskinindlæringsproblemer falder ind. Det er når du har input- og outputvariabler, og du prøver at lave en kortlægning mellem dem.

Det kaldes overvåget læring, fordi vi kan bruge dataene til at lære modellen det rigtige svar.

Algoritmen foretager forudsigelser baseret på dataene, og den korrigeres langsomt, indtil disse forudsigelser matcher det forventede output.

De fleste af de problemer, der overvåges læringsdækninger, kan løses med klassificering eller regression. Så længe du har mærket data, arbejder du med overvåget maskinindlæring.

Semi-overvåget læring

De fleste problemer i den virkelige verden falder inden for dette område på grund af vores datasæt.

I mange tilfælde har du et stort datasæt, hvor nogle af dataene er mærket, men det meste er det ikke. Nogle gange kan det være for dyrt at lade en ekspert gennemgå og mærke alle disse data, så du bruger en blanding af overvåget og ikke-overvåget læring.

En strategi er at bruge de mærkede data til at gætte om de umærkede data og derefter bruge disse forudsigelser som deres etiketter. Derefter kan du bruge alle dataene i en slags overvåget læringsmodel.

Da det også er muligt at lære uden opsyn på disse datasæt, skal du overveje, om det ville være en mere effektiv vej at gå.

Uovervåget læring

Når du kun har inputdata og ingen tilknyttede outputdata, og du vil have en model til at fremstille det mønster, du leder efter, det er når du indtaster læring uden opsyn.

Din algoritme vil gøre noget, der giver mening for det baseret på de parametre, du giver det.

Dette er nyttigt, når du har mange tilsyneladende tilfældige data, og du vil se, om der er interessante mønstre i dem. Disse problemer er normalt gode til klyngealgoritmer og giver dig nogle uventede resultater.

Praktiske anvendelser af maskinlæring for udviklere

Klassifikation

Når du vil forudsige en etiket til nogle inputdata, er dette et klassificeringsproblem.

Machine learning handles classification by building a model that takes data that's already been labeled and uses it to make predictions on new data. Basically you give it a new input and it gives you the label it thinks is correct.

Predicting customer churn, face classification, and medical diagnostic tests all use different kinds of classification.

While all of these fall under different domains of classification, they all assign values based on the data their models used for training. All of the predicted values are exact. So you'll predict values like a name or a Boolean.

Regression

Regression is interesting because it crosses over machine learning and statistics. It's similar to classification because it's used to predict values, except it predicts continuous values instead of discrete values.

So if you want to predict a salary range based on years of experience and languages known, or you want to predict a house price based on location and square footage, you would be handling a regression problem.

There are different regression techniques to handle all kinds of data sets, even non-linear data.

There's support vector regression, simple linear regression, and polynomial regression among many others. There are enough regression techniques out there to fit just about any data set you have.

Clustering

This moves into a different type of machine learning. Clustering handles unsupervised learning tasks. It's like classification, but none of the data is labeled. It's up to the algorithm to find and label data points.

This is great when you have a massive data set and you don't know of any patterns between them, or you're looking for uncommon connections.

Clustering helps when you want to find anomalies and outliers in your data without spending hundreds of hours manually labeling data points.

In this case, there's often not a best algorithm and the best way to find what works for your data is through testing different algorithms.

A few clustering algorithms include: K-Means, DBSCAN, Agglomerative Clustering, and Affinity Propagation. Some trial and error will help you quickly find what algorithm is the most efficient for you.

Deep learning

This is a field of machine learning that uses algorithms inspired by how the brain works. It involves neural networks using large unclassified data sets.

Typically performance improves with the amount of data you feed a deep learning algorithm. These types of problems deal with unlabeled data which covers the majority of data available.

There are a number of algorithms you can use with this technique, like Convolutional Neural Networks, Long Short-Term Memory Networks, or the Deep Q-Network.

Each of these are used in projects like computer vision, autonomous vehicles, or analyzing EEG signals.

Tools you might use

There are a number of tools available that you can use for just about any machine learning problem you have.

Here's a short list of some of the common packages you'll find in many machine learning applications.

Pandas: This is a general data analysis tool in Python. It helps when you need to work with raw data. It handles textual data, tabular data, time series data, and more.

This package is used to format data before training a machine learning model in many cases.

Tensorflow: You can build any number of machine learning applications with this library. You can run it on GPUs, use it to solve IoT problems, and it's great for deep learning.

This is the library that can handle just about anything, but it takes some time to get up to speed with it.

SciKit: This is similar to TensorFlow in the scope of machine learning applications it can be used for. A big difference is the simplicity you get with this package.

If you're familiar with NumPy, matplotlib, and SciPy, you'll have no problems getting started with this. You can create models to handle vehicle sensor data, logistics data, banking data, and other contexts.

Keras: When you want to work on a deep learning project, like a complex robotics project, this is a specific library that will help.

It's built on top of TensorFlow and makes it easy for people to create deep learning models and ship them to production. Y

ou'll see this used a lot on natural language processing applications and computer vision applications.

NLTK: Natural language processing is a huge area of machine learning and this package is focused on it.

This is one of the packages you can use to streamline your NLP projects. It's still being actively developed and there's a good community around it.

BERT: BERT is an open-source library created in 2018 at Google. It's a new technique for NLP and it takes a completely different approach to training models than any other technique. B

ERT is an acronym for Bidirectional Encoder Representations from Transformers. That means unlike most techniques that analyze sentences from left-to-right or right-to-left, BERT goes both directions using the Transformer encoder. Its goal is to generate a language model.

Brain.js: This is one of the better JavaScript machine learning libraries. You can convert your model to JSON or use it directly in the browser as a function and you still have the flexibility to handle most common machine learning projects.

It's super quick to get started with and it has some great docs and tutorials.

Full machine learning example

Just so you have an idea of what a machine learning project might look like, here's an example of the entire process.

Getting data

Arguably the hardest part of a machine learning project is getting the data. There are a lot of online resources you can use to get data sets for machine learning, and here's a list of some of them.

  • Critical care data set
  • Human heights and weights
  • Credit card fraud
  • IMDB reviews
  • Twitter airline sentiment
  • Song data set
  • Wine quality data set
  • Boston housing data set
  • MNIST handwritten digits
  • Joke ratings
  • Amazon reviews
  • Text message spam collection
  • Enron emails
  • Recommender system data sets
  • COVID data set

We'll use the white wine quality data set for this example and try to predict wine density.

Most of the time data won't be this clean when it comes to you and you'll have to work with it to get it in the format you want.

But even with data like this, we're still going to have to do some cleaning.

Choosing features

We're going to pick out a few features to predict the wine density. The features we'll work with are: quality, pH, alcohol, fixed acidity, and total sulfur dioxide.

This could have been any combination of the available features and I chose these arbitrarily. Feel free to use any of the other features instead of these, or feel free to use all of them!

Choosing algorithms

Now that you know the problem you're trying to solve and the data that you have to work with, you can start looking at algorithms.

Since we're trying to predict a continuous value based on several features, this is mostly likely a regression problem. If we were trying to predict a discrete value, like the type of wine, then that would likely be a classification problem.

This is why you have to know your data before you jump into the machine learning tools.

It helps you narrow down the number of algorithms you can choose for your problem. The multivariate regression algorithm is where we'll start. This is commonly used when you are dealing with multiple parameters that will impact the final result.

The multivariate regression algorithm is like the regular regression algorithm except you can have multiple inputs. The equation behind it is:

y = theta_0 + sum(theta_n * X_n)

We initialize both the theta_0 (the bias term) and theta_n terms to some value, typically 1 or 0 unless you have some other information to base these values on.

After the initial values have been set, we try to optimize them to fit the problem. We do that by solving the gradient descent equations:

theta_0 = theta_0 - alpha * (1 / m) * sum(y_n - y_i) theta_n = theta_n - alpha * (1 / m) * sum(y_n - y_i) * X_n

where y_n is the predicted value based on the algorithm's calculations and y_i is the value we have from our data or the expected value.

We want the margin of error between the predicted value and the real value to be as small as possible. That's the reason we're trying to optimize theta values. This is so we can minimize the cost function for predicting output values.

Here's the cost function equation:

J(theta_n) = (1 / 2m) * sum(y_n - y_i)^2

That's all of the math we need to build and train the model, so let's get started.

Pre-processing data

The first thing you want to do is check and see what our data looks like. I've done some modifications to that wine quality data set so that it will work with our algorithm.

You can download it here: //github.com/flippedcoder/probable-waddle/blob/master/wine-quality-data.csv.

All I've done is take the original file, removed the unneeded features, moved the density to the end, and cleaned up the format.

Now we can get to the real pre-processing part! Make a new file called multivariate-wine.py. This file should be in the same folder as the data set.

The first thing we'll do in this file is import some packages and see what the data set looks like.

import pandas as pd import numpy as np import matplotlib.pyplot as plt df = pd.read_csv('./wine-quality-data.csv', header=None) print(df.head())

You should see something like this in your terminal.

C: multivariate-regression-wine.py 7.e 6.3 9.5 97 .ø 7.2 7.2 8. 1 17ø.ø 132.ø 186.ø 186.ø 3.øø 3 30 3.26 3. 19 3. 19 9 6 6 6 6 6 1. øele ø.  994 € ø.  9951

The data looks good to go for the multivariate regression algorithm, so we can start building the model. I do encourage you to try and start with the raw white wine data set to see if you can find a way to get it to the correct format.

Building the model

We need to add a bias term to the data because, as you saw in the explanation of the algorithm, we need it because it's the theta_0 term.

df = pd.concat([pd.Series(1, index=df.index, name="00"), df], axis=1)

Since the data is ready, we can define the independent and dependent variables for the algorithm.

X = df.drop(columns=5) y = df.iloc[:, 6]

Now let's normalize the data by dividing each column by the max value in that column.

You don't really have to do this step, but it will help speed up the training time for the algorithm. It also helps to prevent one feature from being more dominate than other features.

for i in range(1, len(X.columns)): X[i-1] = X[i-1]/np.max(X[i-1])

Let's take a look at the data since the normalization.

print(X.head())

You should see something similar to this in the terminal.

The data is ready now and we can initialize the theta parameter. That just means we're going to make an array of ones that has the same number of columns as the input variable, X.

theta = np.array([1]*len(X.columns))

It should look like this if you print it in your terminal, although you don't need to print it if you don't want to.

[1 1 1 1 1 1]

Then we're going to set the number training points we'll take from the data. We will leave 500 data points out so we can use them for testing later. This is going to be the value for m from the gradient descent equation we went over earlier.

m = len(df) - 500

Now we get to start writing the functions we'll need to train the model after it's built. We'll start with the hypothesis function which is just the input variable multiplied by the theta_n parameter.

def hypothesis(theta, X): return theta * X

Next we'll define the cost model which will give us the error margin between the real and predicted values.

def calculateCost(X, y, theta): y1 = hypothesis(theta, X) y1 = np.sum(y1, axis=1) return (1 / 2 * m) * sum(np.sqrt((y1 - y) ** 2))

The last function we need before our model is ready to run is a function to calculate gradient descent values.

def gradientDescent(X, y, theta, alpha, i): J = [] # cost function for each iteration k = 0 while k < i: y1 = hypothesis(theta, X) y1 = np.sum(y1, axis=1) for c in range(1, len(X.columns)): theta[c] = theta[c] - alpha * (1 / m) * (sum((y1 - y) * X.iloc[:, c])) j = calculateCost(X, y, theta) J.append(j) k += 1 return J, j, theta

With these three functions in place and our data clean, we can finally get to training the model.

Training the model

The training part is the fun part and also the easiest. If you've set your algorithm up correctly, then all you'll have to do is take the optimized parameters it gives you and make predictions.

We're returning a list of costs at each iteration, the final cost, and the optimized theta values from the gradient descent function. So we'll get the optimized theta values and use them for testing.

J, j, theta = gradientDescent(X, y, theta, 0.1, 10000)

After all of the work of setting up the functions and data correctly, this single line of code trains the model and gives us the theta values we need to start predicting values and testing the accuracy of the model.

Testing the model

Now we can test the model by making a prediction using the data.

y_hat = hypothesis(theta, X) y_hat = np.sum(y_hat, axis=1)

After you’ve checked a few values, you'll know if your model is accurate enough or if you need to do more tuning on the theta values.

If you feel comfortable with your testing results, you can go ahead and start using this model in your projects.

Using the model

The optimized theta values are really all you need to start using the model. You'll continue to use the same equations, even in production, but with the best theta values to give you the most accurate predictions.

You can even continue training the model and try and find better theta values.

Final thoughts

Machine learning has a lot of layers to it, but none of them are too complex. They just start to stack which makes it seem more difficult than it is.

If you're willing to spend some time reading about machine learning libraries and tools, it's really easy to get started. You don't need to know any of the advanced math and statistics to learn the concepts or even to solve real world problems.

The tools are more advanced than they used to be so you can be a machine learning engineer without understanding most of the math behind it.

The main thing you need to understand is how to handle your data. That's the part no one likes to talk about. The algorithms are fun and exciting, but there may be times you need to write SQL procedures to get the raw data you need before you even start processing it.

Der er så mange applikationer til maskinindlæring fra videospil til medicin til produktionsautomatisering.

Bare tag lidt tid og lav en lille model, hvis du er interesseret i maskinlæring. Når du begynder at blive mere behagelig, skal du tilføje den model og fortsætte med at lære mere.