More and more companies are becoming interested in Artificial Intelligence and Machine Learning and how they can implement them in their processes. In addition, Google's trend history shows that the focus on Big Data has remained the same over time, but the interest in the term Machine Learning has been increasing over the last three years. 

A decade ago, the emergence of mobile telephony was a paradigm shift.  Today it is the potential applications of Artificial Intelligence that are setting the pace for the technological future. Google's CEO, Sundar Pichai, explained that soon "we will move from a mobile-first conception to an AI-first world where we will see enormous opportunities for dramatic improvement"

However, this growing trend clashes with the reality of many companies. The reason is that their day-to-day operations are moving at a different pace to technological innovations. The main consequence of not implementing machine learning at an early stage is the loss of opportunity compared to the competition, especially in critical business areas such as:

• Process optimization
• Personalization of products and services to your target audience
• Detection of new business opportunities

When it comes to applying Machine Learning, many companies face great doubts. Mainly regarding the cost of adopting this technology and the benefits it can bring. But it is neither difficult nor expensive, there are many AI solutions. A good way to start is the pre-trained models of Google Cloud Platform that we will explain in this post.

What is Machine Learning?

Before addressing how to implement Machine Learning (ML) in a company, it is necessary to know what it is about. Machine Learning is one of the applications of Artificial Intelligence. While AI is a system that, through algorithms, manages to emulate human cognitive functions, ML focuses on how these systems learn through data analysis.
 

Data is the key to the successful implementation of these systems. At the same time, they are also the main difficulty faced by companies that want to apply Machine Learning in their processes because:

• The volume of data. For Machine Learning models to be effective, they need a large amount of data to evaluate - as this will be an essential part of their training.
• Data quality. The volume is essential for training. However, the model accuracy that is developed will depend on how valuable the data undergoing training is. In this sense, a parallel can be drawn with our learning as children: we may have many examples of adult behavior, but if these are bad habits, we will repeat them in the future.
• Data source. When implementing Machine Learning, we have to be sure how the data was obtained complies with privacy protection laws and regulations.  

The paradigm shift introduced by Machine Learning lies mainly in the way it is built. Before, there was a team of programmers writing lines of code anticipating all possible business rules. Now there is a team that trains an ML model from data labelled as good or bad. The result is a trained model that receives inputs and outputs based on what it has learned.

How to implement Machine Learning

Once we have secured the data, the next step is to consider what knowledge or practical applications we want to achieve with the adoption of Machine Learning. This aspect is crucial because depending on the specific need of the business, we will choose one of the different solutions that exist in the market.

In the implementation process that we carry out at Emergya, this first phase is defined. We accompany the client in brainstorming to detect their expectations and their real needs. And thus offer the most suitable model.

In effect, there are three ways to implement Machine Learning in a company: 

• Custom solution. This option is usually the most accurate, but it requires a large amount of data and defining the algorithm of our Machine Learning model, which increases the price. 
• Pre-trained models. Google has a library of proprietary resources open to any company. In Google's AI hub, you can find models that have been trained to identify people, for translations, to build assistants, etc.
• Mixed solution. This is where the pre-trained models are customized and adjusted to the specific needs of the company.

Pre-trained models are usually the fastest option for those companies that want to take the first step towards Machine Learning implementation because it does not require previous data to train the model.

Pre-trained models: the Google hub

Google hub is a compendium of AI resources ranging from tutorials and guides to processing flows, algorithms and other Machine Learning components. The key to this repository is that they do not need to be configured but are ready to be plugged in and start working. 

In this hub, Google makes available to companies all the public content it owns on Cloud AI. But at the same time, it establishes an accessible private space only to authorized individuals within a company.

One way to group the Machine Learning models in Google's hub is by the function they are focused on. Thus, we find APIs related to view function, language, conversation and structured data. 

View

There are two main APIs:
Vision. Among others, it classifies images by categories and discriminates between objects and people. It also extracts texts, entities and recognizes emotions.
Video Intelligence. It obtains metadata from clips, detects objects and people, separates signals from the noise...

Language

Beyond the model for translations, another of the available APIs is the Natural Language Processor (NLP), which allows the nuances and moods identification.

Conversation

These are models oriented towards human-machine interaction, among which we find:

Dialogflow. which is dedicated to building conversational interfaces, is the best known pre-trained model of this function. 
Speech-to-text. It can transcribe audio and, depending on its encoding, it can recognize the existence of two or more voices.
Text-to-speech. Encodes a text into speech by adding pauses, numbers, text formats and pronunciation instructions. Accessibility systems are based on this model.

Structured Data

Unlike other models that process unstructured data, there are other APIs designed to interpret structured data. 

These inputs are associated with labels, making it possible to search for patterns to make predictions. This function is the main of the Inference API, which can offer forecasts and probabilities based on time series analysis. 

Implementation support

Google's pre-trained models are a way to implement Machine Learning quickly, but it can be difficult for companies that do not have enough experience in these tools.

It is an excellent solution to introduce a business to Artificial Intelligence. But, it is advisable to go hand in hand with a company that can put it into production in the most efficient way possible to get good results and do it soon. Only then will the pre-trained models become real business accelerators.