Artificial Intelligence: Let’s be serious

By integrating several elements, software and hardware, in its latest iPhone and iOS 11 available since September 2017 for hundreds of millions of users, Apple allows companies to integrate Artificial Intelligence in their applications if they know how to choose and train the predictive models. We share some of our insight with you here.

Artificial Intelligence: Let’s be serious …

We do not believe its Artificial intelligence as most of you would imagine. We think it is still going to be several years before “real”Artificial Intelligence (strong AI) comes into play. Although many people are fantasising about it and the commercial esteem of the companies that promise are firing from all cylinders.

It is more serious to speak of algorithms created by humans.

“The evolution will be to think machines and that I do not know how to do it,”

– Luc Julia (the creator of Siri)

 

Artificial Intelligence is a true scientific subject, vast and already old. Strongly correlated to mathematics, the subject quickly becomes technically arduous.

However, some disciplines of Artificial Intelligence become more accessible, with today’s Machine Learning.

Apple has just made available in the iPhone2 and iPad, solutions to democratize the use. But what for and how?

Can we predict the future with a smartphone?

Machine Learning is neither clairvoyance nor divination. Under this name are scientific methods based on mathematical models, sometimes borrowed from biology. These algorithms analyze data to make identifications and rankings.

This subject is the buzz, but these algorithms have been known for several decades and you already experience them in your daily life, below are some examples;

Estimation of the price of a house according to its characteristics on a real estate site;

Calculation of the accident risk and the rates of your insurance;

Filter spam e-mails from your software or vendor.

And so on and so forth.

What changes very quickly, however, is the extension of these treatments to new data (text, audio, photos, location, video, IoT sensors) and the availability of huge storage and computing powers to propose new data. less vertical, less “industrial” uses.

This computing power available (in the cloud and with decreasing costs) and the abundance of data reinforce the capabilities and precision of a model for:

Define the content of an image on the web or taken with your smartphone: distinction (dog or cat), identification (coffee or restaurant), ranking (number, order, plans), for digital marketing it is the possibility on an e-commerce site to mount any photo to buy clothes for example;

Determine the risk of occurrence of a failure on a machine according to exogenous parameters (weather, events, users, …); very useful now with the rise of connected objects to embody your services to your customers;

Predict the next word on a keyboard according to the context (interlocutor and history of the relationship, measured emotion, …) (Google has been doing this for some time, and recently made significant changes)

What is new is that Apple (and Google with another approach) provides companies with a machine learning solution embedded on smartphones.

And in this area, several axes of use of Machine Learning seem particularly interesting to us because “accessible” for you.

Beyond visual recognition and voice, the added value of machine learning within an application could well be found in Context Aware Computing (any message – whether audio, video, photo, textual – is adapted in real time to the interlocutor, the place, the temporality, the context, … according to the “learned” rules).

It is about adapting a service to a whole series of environmental data: the user and his known historical data, but also what he wears as clothing, his face, his emotions, the place, the temperature, the people around him, his previous actions, … All of these data will have been previously modeled to provide the algorithm with a basis for recognition and selection of the best answer.

Product catalog navigation: assisted search, trend selection, promotion targeting, mobile predictive download, image scheduling, cross product recommendation and up-selling are no longer based on business rules but on the recognition of use of each customer.

 

CoreML, from iOS 11

Behind this name, CoreML, we find an Apple software package to optimally run algorithms based on machine learning models.

The execution of the calculations and the storage of data are done on the smartphone. This allows a fast processing time without connection to a server and will certainly be an element of valuation of your services in the face of concerns about the protection of personal data.

The field of possibilities has served, in the cloud to give birth to a universal matrix that allows local fast analysis and ranking without the need for a connection to mass transit data.

Depending on the intended applications, Apple proposes to use different types of Machine Learning models. Some are supported by default, others can be imported.

Les us select the most appropriate and efficient algorithm for the requirement.

The use of existing models reduces development times. Apple offers tools to convert models for CoreML, models are reusable elsewhere in the company. Some ready-to-use templates are also available, especially for the detection of objects on images for example.

Another novelty, these developments are not reserved for mobile app developers, they are designed in languages ‚Äč‚Äčaccessible to as many people as possible (Python for example or C ++): thus developers who work on websites, business applications, statistical models, … can work with these models.

But it remains to mobilize two other pillars of Machine Learning that Apple or Google can not offer you: the human (to set goals & set) and data (to train models).

Think and identify the improvements that Machine Learning can bring to your business? as an audit of your existing and a recommendation of Strategy Data integrating the training of the model with your data: “Train Your Model”.

 

Machine Learning = Learning

In order to be able to decide, the machine must learn …

From “known” data, we will train the model by iterations. At each iteration, the quality of the forecasts obtained is tested (on a “control” data set). If the forecast error is important, some parameters of the algorithm are varied. Iterations stop when the forecast error no longer drops and the model is implemented in production; the machine has learned “patterns” and is able to detect similar cases very effectively.

So few methods allow the algorithms to re-configure themselves independently (except the Reinforcement ML: which automatically determines the ideal behavior in a specific context). It is therefore necessary to regularly enrich the training of the model; the analysis of the logs that we do allows to do it.

Independent learning is not yet a reality. Most of today’s ready-to-use bots are based on pre-programmed questions.

Insign accompanies you in learning your algorithms because if your brand participates in conversations (business conversational), it is up to you to manage this production of contents.

“Train your model”: the steps

The following process allows you to develop with you a model of Machine Learning to embed in a mobile application or a website.

1 / Collection of data

Study structured data from your transactions, digital devices, stores, sensors, mobile and web analytics. If unstructured data exists, Data Scientists will make it usable by the model.

2 / Optimization of the data

Eliminate noise (errors, inconsistencies) and normalize data (to make them compatible, comparable to each other) to produce a reliable initial dataset.

3 / Dimensional reduction

Data Scientists then proceed to identify and select the explanatory features of your data.

For example the navigation of a mobile application can be made dynamic to reduce the number of images and thus the download time. By analyzing the data and modeling the relationship between navigation and image loading, a model is produced to limit unnecessarily loaded images based on the current navigation, potentially for each user.

4 / Model selection

Depending on the data and the processing to be performed, it is possible to immediately identify the most relevant model or, on the contrary, to test different models.

5 / Learning the model

Finally, study the performance of the model by comparing the results obtained with raw data with those optimized and adjust its settings.

Finally, the model can be put into production in an iOS or Android application.

1 The real danger of Artificial Intelligence today is still the man who is at the helm of the algorithms and who can reinforce the bones of their use.

2 The possibilities are similar in Android “O”, the update of the operating system of Google. Facebook, Amazon, IBM and others have also undertaken to make Machine Learning capabilities, or even other Artificial Intelligence disciplines, accessible to companies.

3 The famous neural networks attempt to reproduce the architecture and the functioning of the circuits of the brain.

4 Alphabet (parent company of Google) created a Machine Learning solution through a TensorFlowTM open source software library, coupled with a hardware solution; its own electronic cards optimized for the processing of data submitted to the Machine Learning (TPU: TensorFlowTM Processing Units) accessible in Cloud.

Any comments you have please feel free to leave them below.

Leave a Reply

Your email address will not be published. Required fields are marked *