ru ua en
c l o c k
Training AI with design
17 February 2020
blog

Training AI with design

In the previous article Artificial Intelligence Driven Design we have found out about the mechanism of influence of AI on the creativity process and appreciated the importance of the data in providing quality user experience. In this article we will dive into how we can train artificial intelligence (henceforth AI) through design.

Importance machine learning

With the development of AI-driven society, machine learning has also become a design decision. Machine learning is the most talked about topic in our industry right now and changes more and more in the process of creating and applying our products. We have a ton of new possibilities, but also some weighty issues.

The one of these issues is that machine learning is now UX problem. Explaining AI Learning Processes and designating your capabilities in this area as a designer, we hope to increase your confidence and motivation so that in the next project managed by AI you can purposefully apply this knowledge.

How does AI learn

There are several ways that artificial intelligence (AI) can learn and grow. Typically, you can say that AI is learning based on trial and error method. Unlike conventional programming, machine learning works the other way around so you don't programming a solution, the machine itself learns based on a set of rules. Imagine you want to make the best burger in the world. Instead of trying to come up with your own recipe, you buy some great burgers, and start with that.

You bought a bunch of hamburgers, told the machine what it is and which ingredients you like the most. The machine takes each part separately and comes up with the perfect burger. Magic effect of artificial intelligence is that you may not even know HOW it is done.

How does AI learn

Machine learning, first step to AI

Machine learning can be seen as a science that is designed to help. computers detect patterns and relationships in data. As for designers, templates are key elements for creating products, so having a good training model guarantees success. Basically, there are three ways of machine learning - supervised, unsupervised and reinforcement learning.

The types of machine learning

Supervised Learning
Supervised Learning. Teaches the algorithm to execute classification and regression of the specific dataset.
Unsupervised Learning
Unsupervised Learning. Teaches the algorithm to find a cluster and associations in unmarked dataset.
Reinforced Learning
Reinforced Learning. Training agent to perform certain interactions in an environment without data

Supervised Learning

Let's continue with the example of a hamburger. Do you want to create the best burger in the world. To do this, you first need find out and find out what a good burger is made of. This is where we will need machine learning. Machine in this case, informs the AI ​​useful information, for example, which bun is better, which meat is better or the best cheese to use (based on all ideal the burgers that you provided the car). Machine learning copes with individual specific tasks perfectly.

Each part of the burger (bun, meat, cheese) will be subjected to analysis to understand which type is best for each part. This is how the learning process takes place, you must provide examples of the algorithm and control its training to make sure the results are correct.

The data that educates the model can take two forms:

  • Labeled data (data that is already marked or training data), tagged data is just a set data that already matters, for example: This Apple
  • Unlabeled data (data used for training models)

In the training data, the model will know the features and how to label them correctly, while in test data, even if you know the label, you won’t reveal it to the models, that will be the test.

Unsupervised Learning

Unsupervised Learning

With this type of training you can find out which hamburgers the best ones, but you won’t know what they are made of - the machine will cluster the data set, which at first looks disorganized, and offer its own assumptions and associations.

Reinforced Learning

DeepMind was created to develop general AI. They act in accordance with the scientific mission to create systems that can learn to solve any complex problems without the supervision of a teacher. You need to achieve a level of general AI for this. Through reinforcement learning, DeepMind develops intelligent agents that perform learning through auxiliary tasks unattended.

Simply put, these agents carry out different strategies. (game methods in this case) to find most suitable. Every time they are “killed” or “game over”, the agent carries out an alternative strategy until it reaches a successful result.

Also, there are several other ways to learn, and all of them based on the principle of providing data and checking the accuracy of the results learning. This sequence continues until the data is grouped as needed.

Using Machine Learning in design

Now that we know HOW machines learn, let's focus on how design affects learning. Companies like Google and Netflix use machine learning in very efficient and extraordinary way. Google, for example, uses machine learning to make your email service more efficient with applying automatic responses. Netflix applies personalized movie covers to arouse the interest of a particular user.

Gmail effective quick reply

Google uses machine learning to make email experience more efficient, while reducing typical response time. Google pointed out the need to help users respond to email and suggested a quick response feature. Given that Google has many years of experience in linguistics thanks to the Google Translate platform, it was used to learn how to choose the most suitable answer.

Gmail effective quick reply

Personalized Netflix movie covers

Most of us watch TV shows on Netflix, and almost every day we come across with the result of machine learning, namely personalized system of recommendations. Netflix wants us to watch good movies and TV shows. But only the title of the content is not always enough. The process of selecting the perfect recommendations for the user is extremely complex.

There are so many different tastes and preferences that Netflix decided it was necessary to highlight aspects of the title that would have particular importance to the user.

Your browsing history plays a big role in deciding which one from the thousands of movies to offer you. The image below shows how your story affects proposed artwork based on the genres you prefer. If you watch a lot of romantic movies you may be interested in Good Will Hunting, if Netflix shows Matt Damon and Minnie Driver. Whereas if you like comedies featuring Robin Williams, Netflix will get more of your attention. The same logic can apply to actors and actresses. If you watched a lot of films with John Travolta, then you would rather choose a film with John Travolta than the same film, but with Uma Thurman on the cover.

Personalized Netflix movie covers

Та ж логіка може застосовуватися до акторів і актрис. Якщо ви дивився багато фільмів з Джоном Траволтою, то ви швидше виберете фільм з Джоном Траволтою, ніж той же фільм, але з Умою Турман на обкладинці.

Netflix Personalized movie covers

How does Netflix conduct machine learning?

Most Netflix recommendations work on machine learning. Netflix applies contextual approach in which they test two algorithms against each other.

Netflix about their method: “Traditionally, we collect a data package on how Our customers use the service. Then we launch a new machine learning algorithm on this batch of data. Then we test this new algorithm against the current one. A / B test helps us see if a new algorithm is better than our current production system, having tested it on random number of users "

This means that members of group A + receive a new algorithm. If the group B has a higher level of engagement, a new algorithm will be launched. But unfortunately this approach did not present the expected results, so Netflix adopted the online machine learning approach, where the algorithm is trained in a constant mode.

Users loop

As AI learns from retrospective data to predict result, it can become outdated very quickly, for example: Let's say you build a model to establish which email is most important in user’s mailbox, and to do this, you take into account several things: subject, content, time required for the user to interact with this letter, response time, and other things.

You are training intensively with retrospective data and now your output is 90% accurate and the model gives you pretty good results, so whenever a user receives a new email, he knows its priority. But unfortunately, this accuracy cannot be constant because the model cannot constantly operate with retrospective data, without updating it becomes useless. Adding your user reviews to the above data may help update the result.

How to create a feedback loop in ML mode?

Let's take an example of email and see how Google fixed it. Gmail priority folder ranks emails by the likelihood of how important a message is to the user. But it is necessary to consider that priorities change quickly enough, therefore, having learned what is important for each user, it is necessary to update as often as possible.

How exactly it was done ?

Google provides new users with a “test model” and keep training it every day, actively watching how people interact with their emails. Thus, this model will improve every 24 hours.

How to create a feedback loop in ML mode

Design stimulates AI

Communicating with users and getting their feedback is an important part of designers work. While creating intelligent systems, we can constantly develop, by making sure that it is the interests and goals of our user are what drives artificial intelligence forward.

Built-in feedback

Developing interfaces with AI, we must keep in mind that AI can make mistakes. Therefore, it’s worth starting with creating interfaces that have built-in structured feedback - in the event of an error, structured feedback is much more effective than just yes or no answers.

In the Netflix example mentioned earlier, they created a complete system just to try and show you high quality content. But what if you really didn't like this improvement? If there is no structured feedback, you will be only rely on sure-fine machine work, not including human factor.

Do cover up important information

In my opinion, the main danger of AI is that we can delegate important life decisions under the black box responsibility, pretending to be a person at this moment we can only aggravate the situation. Do not forget to tell users what is exactly happening, show them why you recommend something.

Youtube recommends you new videos based on the other people views: “Kevin Kenson’s audience is watching this,” so you can give your users more transparency and gain their trust, because now they know why you recommend them this or that content.

Prejudice

When we hear the word prejudice, we automatically think about something bad. But do not rush to judgment. Prejudice is a gift of evolution, the ability to quickly accept solutions, instant reaction that can save your life - this is just one of thinking types. In the book Think Fast and Slow Daniel Kahneman explores the idea of ​​a fast and slow brain, fast brain that makes quick decisions and slow, which is our mind. Fast brain scans fast everything that you have experienced in life and makes really fast decision.

But what do we mean when we talk about prejudice against AI? There are many examples of how AI becomes racist, sexist and so on - how does this happen

Example

Amazon has been working on AI to sort hundreds of CVˆs and find the best candidate for the post, but during the model testing they noticed that the algorithm choses only men, but also giving a negative assessment to every female candidate. When you feed the model with retrospective data, you will get not quite relevant result. Naturally machines do not have prejudices, but we do them based on data with our prejudices, and there is no easy way to fix this. Most of the time we don’t even know if we have prejudices, because that’s what we have been taught.

The first step is to recognize that we have an unconscious prejudices and after that, only testing and communication with your users will help you control yourself and your product.

Join the ML!

In this article you read a lot about the importance of data, machine learning process, teaching methods and the prejudices which affect on AI. Many information to read, but not much interaction and fun, therefore Pedro Marquez prepared the site, which allows you to experience for yourself what is machine learning. Enjoy it!

Are you ready for leadership on the internet?

Let's discuss the prospects for promoting your site!