Conceptos básicos de aprendizaje automático para desarrolladores

En el panorama tecnológico actual, se espera que los desarrolladores tengan varias habilidades diferentes. Y muchos de ellos lo hacen.

También hay muchas trayectorias profesionales diferentes disponibles para los desarrolladores que utilizan muchas de sus habilidades actuales con un ligero giro.

Los administradores de bases de datos, los defensores de los desarrolladores y los ingenieros de aprendizaje automático tienen una cosa en común con todos los desarrolladores: todos saben cómo codificar. No importa qué lenguajes se estén utilizando, todos comprenden los conceptos básicos detrás de la escritura de un buen código.

Esa es una de las razones por las que muchos desarrolladores de software consideran convertirse en ingenieros de aprendizaje automático. Con todas las herramientas y paquetes disponibles, no es necesario tener una base matemática profunda para obtener resultados precisos.

Si está dispuesto a aprender a usar algunas bibliotecas y obtener una comprensión de alto nivel de las matemáticas subyacentes, puede convertirse en un ingeniero de aprendizaje automático.

En este artículo, lo guiaré a través de algunos de los conceptos principales en el aprendizaje automático que debe comprender desde su experiencia como desarrollador de software.

Terminaremos con un ejemplo de un proyecto completo de aprendizaje automático, desde obtener datos hasta predecir un valor con un modelo. Al final, deberías tener el conocimiento suficiente para completar tu propio proyecto de aprendizaje automático desde cero.

¿Qué es el aprendizaje automático?

Hay muchas definiciones por ahí. Pero el aprendizaje automático básicamente implica el uso de matemáticas para encontrar patrones en cantidades masivas de datos para hacer predicciones basadas en datos nuevos.

Una vez que haya encontrado esos patrones, puede decir que tiene un modelo de aprendizaje automático.

Desde allí, puede usar el modelo para hacer predicciones sobre nuevos datos que el modelo nunca ha visto antes.

El objetivo es lograr que las computadoras mejoren automáticamente con la experiencia utilizando los algoritmos que se proporcionan.

Un algoritmo es solo una ecuación matemática o un conjunto de ecuaciones que le dan un resultado basado en sus datos de entrada. El aprendizaje automático utiliza algoritmos para encontrar los patrones que estamos buscando.

A medida que los algoritmos se exponen a más y más datos, comienzan a hacer predicciones más precisas. Con el tiempo, el modelo creado por los algoritmos podrá determinar el resultado correcto sin estar programado explícitamente para hacerlo.

Esto significa que la computadora debería poder tomar datos y tomar decisiones (predicciones) sin ayuda humana.

Aprendizaje automático frente a ciencia de datos frente a inteligencia artificial

Mucha gente usa los términos aprendizaje automático, ciencia de datos e inteligencia artificial de manera intercambiable. Pero no son las mismas cosas.

El aprendizaje automático se utiliza en la ciencia de datos para hacer predicciones y descubrir patrones en sus datos.

La ciencia de datos se centra más en estadísticas y algoritmos y en la interpretación de resultados. El aprendizaje automático se centra más en el software y la automatización de las cosas.

La inteligencia artificial se refiere a la capacidad de una computadora para comprender y aprender de los datos, mientras toma decisiones basadas en patrones ocultos que serían casi imposibles de descifrar para los humanos.

El aprendizaje automático es como una rama de la inteligencia artificial. Usaremos el aprendizaje automático para lograr inteligencia artificial.

La inteligencia artificial es un tema amplio y cubre aspectos como la visión por computadora, las interacciones entre humanos y computadoras y la autonomía, donde el aprendizaje automático se usaría en cada una de esas aplicaciones.

Diferentes tipos de aprendizaje automático

Hay tres tipos de aprendizaje automático sobre los que escuchará y leerá: aprendizaje supervisado, aprendizaje semi-supervisado y aprendizaje no supervisado.

Aprendizaje supervisado

Esta es la categoría a la que pertenecen la mayoría de los problemas de aprendizaje automático. Es cuando tiene variables de entrada y salida y está tratando de hacer un mapeo entre ellas.

Se llama aprendizaje supervisado porque podemos usar los datos para enseñarle al modelo la respuesta correcta.

El algoritmo hará predicciones basadas en los datos y se irá corrigiendo lentamente hasta que esas predicciones coincidan con el resultado esperado.

La mayoría de los problemas que cubre el aprendizaje supervisado se pueden resolver con clasificación o regresión. Siempre que haya etiquetado los datos, estará trabajando con el aprendizaje automático supervisado.

Aprendizaje semi-supervisado

La mayoría de los problemas del mundo real caen en esta área debido a nuestros conjuntos de datos.

En muchos casos, tendrá un gran conjunto de datos donde algunos de los datos están etiquetados, pero la mayoría no lo está. A veces puede ser demasiado caro que un experto revise y etiquete todos estos datos, por lo que utiliza una combinación de aprendizaje supervisado y no supervisado.

Una estrategia es usar los datos etiquetados para hacer conjeturas sobre los datos no etiquetados y luego usar esas predicciones como sus etiquetas. Luego, puede usar todos los datos en algún tipo de modelo de aprendizaje supervisado.

Dado que también es posible realizar un aprendizaje no supervisado en estos conjuntos de datos, considere si esa sería una forma más eficiente de hacerlo.

Aprendizaje sin supervisión

Cuando solo tiene datos de entrada y no hay datos de salida asociados y desea que un modelo haga el patrón que está buscando, entonces ingresa al aprendizaje no supervisado.

Su algoritmo inventará algo que tenga sentido en función de los parámetros que le proporcione.

Esto es útil cuando tiene muchos datos aparentemente aleatorios y desea ver si hay patrones interesantes en ellos. Estos problemas suelen ser excelentes para los algoritmos de agrupación en clústeres y le dan algunos resultados inesperados.

Usos prácticos del aprendizaje automático para desarrolladores

Clasificación

Cuando desea predecir una etiqueta para algunos datos de entrada, este es un problema de clasificación.

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.

There are so many applications for machine learning from video games to medicine to manufacturing automation.

Just take some time and make a small model if you're interested in machine learning. As you start to get more comfortable, add on to that model and keep learning more.