Helping machines ‘learn’ and be more useful: An introduction to Machine Learning

Today, technology powers some of the most delightful experiences in your favorite apps and products. It helps organize your photos, understand your voice commands, show you the most relevant reviews for products and businesses, navigate from one place to another, and much more. If you are wondering what goes on behind the scenes or want to understand how technology can perform these wonderful tasks, you are in for a treat! Read on about what has become an integral part of your day-to-day life.

Why is ML popular now?

The key algorithms powering machine learning were created decades ago. They come from disciplines like statistics, linear algebra, biology, physics. Today, ML has become an important tool to solve difficult computational problems, and its popularity has surged in the recent past due to Big Data, the increased computing power of technology, availability of Cloud infrastructure, and advanced algorithms that can analyze the data to solve problems.

Because of this availability of data, algorithms, and technology, ML-based approaches can solve complex problems that rule-based approaches can’t.

In the above example, we see how the ML approach will ‘learn’:“what is a carrot” by trial and error, and this approach can be quickly developed, deployed, and can show an impact on a global scale with a fraction of the resources required otherwise.

That’s great. But how does Machine Learning work?

1. Focus on the user: As with any good product development philosophy, the ML process also starts with understanding the user's needs or wants. Not every problem needs ML to be solved, and conversely, ML can’t solve every problem. Only user problems that are too complex to solve by traditional programming / rule-based systems, and deal with a specific user need, that can be solved by computers ‘learning’ themselves about the problem, should be identified and analyzed for the next stage.

2. Define the objective: Clearly define the problem statement. i.e Problem-framing. For example, your objective could be to predict which friends a user is likely to share a photo with or to suggest where a user should eat in a new city based on the history of the restaurants they have visited in the past. Your goals should quantify success metrics as well.

3. Collect data: Identify the input data that your ML model needs, to make successful predictions. Since ML models learn from examples to recognize patterns, it is essential to find a large existing source of data that is relevant to your problem. Experts say, collecting, cleaning, exploring, and other data processes tend to be the longest but most critical part of the process. The better you know your data, the more useful hypothesis you can make.

4. Train and test the model: Once the data has been collected, the model is trained. Part of the data (called training data) is fed into the algorithm to predict outcomes, and the outcomes are compared to ‘ground truth’ (i.e. data that is known to be correctly identified and labeled). This repeated process runs thousands, sometimes millions of times and when the error reaches the minimal threshold set by the team, the model is ready to be tested. It is then tested on the rest of the data (called test data) to validate the model and make it ‘ready for implementation’. In short, a test set is a data set used to evaluate the model developed from a training set.

5. Predict and evaluate: Once the ML model is rolled out, it will continue to be refined and improved with every user interaction. This is an ongoing step in the ML development process and the iteration takes us back to step 3 (data collection) or step 1, depending on the outcome and progress.

Key concerns and where Crowdsource plays an important role

The ML model is only as good as the examples. Machine learning models face two important limitations: 1) bias and 2) variety.

Machine learning models are not inherently objective. Engineers train models by feeding them a data set of training examples. Human involvement in the provision and curation of this data can make a model's predictions susceptible to bias.

When building models, it's important to be aware of common human biases that can appear in your data, so you can take proactive steps to mitigate their effects.

A biased data sample will teach the algorithm to look for similar patterns and hold them ‘true’. These biases limit the usability of the algorithm, especially in a global context. For example, a company’s AI recruitment tool heavily favored men for the job openings because the 10-year dataset they used to measure whether a candidate would be a good fit for the job mostly came from men, a reflection of the male dominance across the tech industry. Similarly, the lack of variety will cause the model to have a very narrow view of what it is supposed to predict. For example if all of the examples of cats have triangle-shaped ears, it might fail to recognize a cat which has flat / no prominent ears.

The Crowdsource community plays an important role as a driver of diversity of data for the ML models in Google, to help remove biases and add variety, to make our products more inclusive and help us serve everyone, everywhere.

If you’re new here, you can get started at the main Crowdsource website or download our Android app. We would love to hear your feedback and thoughts. Please feel free to write to us on Facebook or find us on Twitter using #GoogleCrowdsource.

This blog post is a recap of our ‘Introduction to Machine Learning’ on On Air with Crowdsource - a series of virtual learning sessions with Googlers about a variety of topics. These sessions help our influencers and top contributors from around the world to build their knowledge and skill sets through practical learning. Due to the overwhelmingly positive response that these sessions have received, we’re super excited to now make this content available to all of our readers.