AI for everyone (week 1)

Source: Deep Learning on Medium


Artificial Intelligence (AI) is transforming all sectors. What are its possibilities, limits and impacts? Andrew Ng’s latest course (IA for everyone) gives you the keys to understanding this digital transformation of our society and the tools to apply it in your activities.

This article is part of the series “Deep Learning in Practice” (read as well the versions in French and Portuguese).

Introduction

Are you a programmer or at least interested in coding? You can access dozens of MOOCs on the Internet (sometimes free) to learn how to create AI models (such as deep neural network models or Deep Learning) such as those of Jeremy Howard (fastai) and Andrew Ng (deeplearning.ai).

With these MOOCs — and with perseverance — anyone interested in coding and AI can become an AI practitioner.

However, the digital transformation of public organizations and companies by AI also requires an understanding of the possibilities, limitations and impacts of AI by all employees, not just the technical teams. Indeed, incorporating the AI ​​into existing processes obviously requires a validation of the executive (which must therefore understand the issues) and may also result in partially or completely modify the existing activities.

It is therefore necessary to make aware of leaders and decision-makers by a non-technical approach to AI, as did Omar bin Sultan Al Olama, Minister of State for Artificial Intelligence of the United Arab Emirates, with the University of Oxford through a one-year program that trained 94 government officials (fontes in French and in English).

Andrew Ng’s latest MOOC (AI for everyone) gives the key information needed to set up this type of training as well as an online AI PlayBook Transformation that allows any organization to think about their own AI mutation strategy.

The content of this MOOC is free and here are the key elements of the week 1.

Credit: all images in this post come from the MOOC Andrew Ng, AIfor everyone.

Key points of the week 1

. Value creation in all sectors
. AIs, not an AI
. AI → Machine Learning → Supervised Learning
. Why now? Big Data and GPU
. Acquire training data
. Problems with the data
. Differences between Machine Learning and Data Science
. Deep Learning
. Characteristics of an AI company
. 5 steps to becoming an AI company
. How to know if the AI ​​can be used?
. Tips for getting a powerful DL model

Tips for a trainer

This week’s content 1 contains all the essentials for understanding AI, its capabilities, limitations, and how to implement it in a company’s automated processes to become an AI company (the content of the week 2 allows to go more in details on the methodology of use of the AI ​​in the projects).

The trainer must present the AI ​​in a Top-Down way starting with its impact in the short and medium term in terms of additional value creation. Indeed, by discovering the predictions of value creation in his sector, the participant will have a personal interest in understanding the AI, and therefore an additional motivation.

The course must be based on simple and comprehensible examples for all.

At the end of the course, after having presented the reality of AI in the form of Machine Learning models (and in particular Deep Learning), the trainer will present the 5 stages of transformation of a company into an AI company, then multiply the examples of what AI can do and what it can not do.

Value creation in all sectors

$ 13 trillion by 2030 (Mckinsey, September 2018) with almost $ 1 trillion in retail (the least affected sectors will be those with high manual value as a hairdresser or surgeon).

AIs, not an AI

If AI combines both ANI (Artificial Narrow Intelligence) and AGI (Artificial General Intelligence), the latter does not exist yet (and will not exist for very long). What we know how to develop today are AIs, each specialized in one task.

AI → Machine Learning → Supervised Learning

In the vast majority of cases, the AI ​​is actually a model of Machine Learning (ie a model that learns from examples) whose learning is done in a supervised way (Supervised Learning). It can thus be summarized that an AI model is very often a model that learns to give an output B (prediction) from an input A (given).

Why now? Big Data and GPU

Today we have a lot of data (up to Big Data) and the computational ability to analyze it (GPU). It is therefore possible to train models of deep neural networks (Deep Learning) whose performance is superior to other models and grow with the number of training data.

Acquiring Training

Data (dataset) for training an AI model is divided into 2 groups: structured data and unstructured data. In the first group, we find all the data in tabular form and in the second, text, voice and image data. In both cases, each piece of data is a pair (A, B), where B is the target / label of A. For example, a set of numbers about a house (A) is targeted at its price (B) and the image of a cat (A) has the label “cat” (B).

To obtain a dataset which will allow to train an AI model for a particular task, there are 3 possibilities: to create it by labeling it (manual labeling), to obtain it by data recording (from observing behaviors) or to download it (download from websites / partnerships).

Problems with the data

If having data is essential to lead an AI model, its acquisition must be done in cooperation between the AI team and the team concerned by the project (IT if it is data on the technique, Marketing if it is promotional data and sales, etc.). Indeed, the AI ​​team can guide the project team on the nature, quantity and quality of the data to be acquired to improve the performance of the AI model (eg, recording every minute and not only every 10 minutes, images in similar quantity in different categories, etc.).

In practice, it is necessary to pre-process the data before using it to drive an AI model because if the data contains errors (garbage in), the model will not be able to perform (garbage out). Errors can be incorrect labels, false or missing values ​​as well as data of a different nature.

Differences between Machine Learning and Data Science

Many organizations already use Data Science to analyze their data in order to extract the main characteristics (ex: importance of this or that parameter on the value of the target, general trends, etc.). Business Intelligence (BI) tools also allow them to interact visually with them (slide deck). The Learning Machine generates meanwhile predictive application (software) able to answer a single question. In summary, Data Science can visualize the past and Machine Learning can predict the future.

Deep Learning

The Machine Learning which is a subset of the AI ​​itself contains subsets of which the most powerful is the Deep Learning (DL). Originally inspired by neural networks of the brain, DL models are composed of several layers of computational units (artificial neurons), each capable of detecting more and more complicated characteristics of a training dataset. In fact, thanks to training from a dataset, a DL model creates a representation of the world that it can then apply to any new data A similar to those of the dataset to predict B.

Characteristics of an AI company

As having a website is not enough to become an Internet business, it is not enough to use one or more DL models to become an AI company. The organization’s organization chart, its data acquisition and storage strategy, and all its automated processes must take into account AI.

5 steps to becoming an AI company

How do I know if AI can be used?

An imperfect rule of thumb that you can use to decide whether or not to use an AI model in an existing process is: everything you could do in less than a second, you could probably automate it using a DL model of the Supervised Learning type.

Tips for a successful DL model

About the author: Pierre Guillou is a consultant in artificial intelligence in Brazil and France. Please contact him via his Linkedin profile.