There are many paths into the field of machine learning and most start with theory.
If you are a programmer then you already have the skills to decompose problems into their constituent parts and to prototype small projects in order to learn new technologies, libraries and methods. These are important skills for any professional programmer and these skills can be used to get started in machine learning, today.
You must learn the theory to be effective in machine learning, but you can use your interests and thirst for knowledge motivate you from working examples into mathematical understandings of algorithms.
In this post you will learn four strategies a programmer can follow to get started in machine learning. This is the path of the technician, which is practical and empirical and will require you to perform research and complete experiments in order to build up your own intuitions.
The four strategies are:
- Study a Machine Learning Tool
- Study a Machine Learning Dataset
- Study a Machine Learning Algorithm
- Implement a Machine Learning Algorithm
Read through these strategies and select one that you feel suits you the best, then execute with abandon.
1. Study a Machine Learning Tool
Select a tool or library that you like and learn how to use it well.
I recommend you start with an environment that provides tools for data preparation, machine learning algorithms and the presentation of results. Learning an environment like this will allow you to get good at the process of machine learning end-to-end which is more valuable to you than learning a specific data preparation technique or machine learning algorithm.
Alternatively, perhaps you are interested in a specific technique of family of techniques. You could use this as an opportunity to deep dive into a library or tool that offers these methods and master the technique by mastering the library that supplies access to the technique.
Some tactics you could follow for this strategy are:
- Compare and contrast candidate tools from which you could choose.
- Summarize the capabilities of your chosen tool.
- Read and summarize the documentation for the tool.
- Complete text or video tutorials for the tool and summarize the key learning points for each tutorial you complete.
- Create tutorials for features or capabilities of the tool. Select things that you don’t know much about and create write a process for getting a result or record a 5 minute screencast on how to use the feature.
2. Study a Machine Learning Dataset
Select a dataset and understand it intimately and discover which algorithm class or type addresses it the best.
I recommend you select a modest sized dataset that fits into memory that may have been well studied before. There are excellent libraries of data sources available for you to browse and choose. Your objective is to understand the underlying problem that the data source represents, the structure in the dataset and the types of solutions that are most suited to the problem.
Use a machine learning or statistical environment to study the dataset. This will allow you to focus on the questions you are seeking to answer about the dataset rather than being distracted with learning about a given technique and learning how to implement it in code.
Some tactics that can help you with your study of an experimental machine learning dataset are:
- Clearly describe the problem that the dataset represents.
- Summarize the data using descriptive statistics.
- Describe the structures you observe in the data and hypothesize about the relationships in the data.
- Spot test a handful of popular machine learning algorithms on the dataset and discover which general class performs better than others
- Tune well performing algorithms and discover the algorithm and algorithm configuration that performs well on the problem
3. Study a Machine Learning Algorithm
Select an algorithm and understand it intimately and discover parameter configurations that are stable across different datasets.
I recommend that you start with an algorithm of modest complexity. Select an algorithm that is well understood, has many open source implementations from you to choose from and has few parameters for you to explore. Your objective is to build up intuitions for how the algorithm performs across a range of problems and parameter configurations.
Use a machine learning environment or library. This will allow you to focus on the behaviors of the algorithm as a “system” as opposed to concerning yourself with canonical mathematical descriptions and reference literature.
Some tactics you can use when studying your chosen machine learning algorithm are:
- Summarize the parameters of the system and the expected influences they have on the algorithm.
- Select a range of datasets suited to the algorithm that are likely to elicit varied behaviors.
- Select algorithm parameter configurations that you believe will elicit varied behaviors from the system and list the behaviors you may expect from the system.
- Consider the behaviors of an algorithm that could be monitored as the algorithm is run over iterations of the algorithms update process or other interval of time.
- Design small experiments using one or more combinations of datasets, algorithm configurations and behavior measures in order to answer a specific question and report results.
Your studies can be as simple or as complex as you like. At the higher end you can explore so-called heuristics or rules of thumb for applying algorithms and empirically demonstrate whether they have merit and if so under what circumstances they correlate with successful outcomes.
Some algorithms you may consider to start with include: least squares linear regression, logistic regression, k-nearest neighbor classification, perceptron
4. Implement a Machine Learning Algorithm
Select an algorithm and implement or port an existing implementation to a language of your choice.
Select an algorithm of modest complexity to implement. I recommend performing some detailed research on the algorithm you which to implement, or select an implementation you like and port it to your chosen target programming language.
Implementing an algorithm by hand from scratch is a great way to learn about the myriad of micro-decisions that have to be made in transforming an algorithm description into a functioning system. By repeating this process with multiple algorithms you will quickly gain an intuition for how to read the mathematical descriptions of algorithms in research papers and books.
Five tactics that may help you when implementing machine learning algorithms from scratch are:
- Start by porting. Porting an open source algorithm implementation from one language to another will teach you how the algorithm is implemented and make it your own. It is the fastest way to get started and is highly recommended.
- Select one algorithm description to work from and collect other algorithm descriptions to support your disambiguation of the primary reference material
- Do not be afraid to reach out to algorithm authors, paper authors or even algorithm implementation authors to ask questions to help you disambiguate your understanding of the algorithm description.
- Read lots of implementations of your target algorithm. Learn how different programmers interpret the algorithm description and turned it into code.
- Do not get caught up on advanced methods. Many machine learning algorithms use advanced optimization methods in their core. Do not try to reimplement these methods unless that is the point of your project. Use a library that provides an optimization algorithm or use a simpler optimization algorithm that is easy to implement (like gradient descent) or is available to you in a library.
Small Projects Methodology
The four strategies being to a methodology I call “small projects”. It is an approach you can use to very quickly build up practical skills in technical fields of study, like machine learning. The general idea is that you design and execute on small projects that target a specific question you want to answer.
Small projects are small in a few dimensions to ensure that they completed and that you extract the learning benefits and move onto the next project. Below are constraints you should consider imposing on your projects:
- Small in time: A project should not take any longer than 5-15 hours from inception to presentation of results. This will allow you to complete a small project in a week of nights and weekend time away from your 9-5 job.
- Small in scope: A project should address the most narrow version of the question you are interested in and still be meaningful. For example, rather than addressing the problem “write a program that will tell me if tweet will be retweeted” in the general case, address the problem just for a specific twitter account for a given time period.
- Small in resources: A project should be able to be completed on your desktop or laptop with a connection to the internet. You should not need exotic software, web infrastructure, or third party data or service. Collect the data you need to file, load it into memory and attack your narrow question using open source tools.
Additional Project Tips
The principle of these strategies is to take action and make use of your programmer skill set. Below are three tips to help you adjust your mindset in order to take action:
- Write down what you learn. I recommend that you have a tangible work product for every step you take. This could be a note in a journal, a tweet, a blog post or an open source project. Each work product acts as an anchor and a milestone.
- Do not write code unless that is the purpose of the project. This tip is not obvious but may be the biggest in terms of accelerating your understanding of machine learning.
- The goal is for you to learn something not to create the a unique resource. Do not worry that no one will read your studies or tutorials or notes on an algorithm. They are your perspective and your work product to demonstrate that you now know something.
Here are the size strategies again with a clear one-linear for each to help you choose the one that is right for you.
- Study a Machine Learning Tool: Select a tool or library that you like and learn how to use it well.
- Study a Machine Learning Dataset: Select a dataset and understand it intimately and discover which algorithm class or type addresses it the best.
- Study a Machine Learning Algorithm: Select an algorithm and understand it intimately and discover parameter configurations that are stable across different datasets.
- Implement a Machine Learning Algorithm: Select an algorithm and implement or port an existing implementation to a language of your choice.
Which strategy would you choose and what will be your first step? Pick one and declare your intentions in a comment below.
If you like this self-study strategy, I have created a 32-page PDF guide you can learn and practice applied machine learning. Check it out:
I have also created a list of 90 project ideas (yeah, I went overboard) and provided it as a bonus with the guide.