After understanding AI and how to develop AI projects by driving Machine Learning models (and Deep Learning in particular), week 3 of Andrew Ng’s new course (AI for everyone) tells you the steps to take to start using AI in your company (association and a public organization, too) to turn it into an AI business.
List of articles “AI for everyone”
- AI for everyone (week 1)
- Building AI projects (week 2)
- Building AI in your company (week 3)
- AI and Society (week 4)
Now that you have a general knowledge of AI thanks to the week 1 of Andrew Ng’s new course (AI for everyone) and you know how to train a machine learning model through week 2, you can start using AI in your business to turn it into an AI business!
Based on the content of his AI Transformation PlayBook, Andrew Ng teaches us this week 3 how to set up his own AI mutation strategy.
The content of this MOOC is free and here are the key elements of the week 3.
Credit: all images in this article come from the Andrew Ng MOOC, AI for everyone.
Tips for a trainer
This week’s content contains all the essential elements for developing AI in an organization (association, public organization, company). Since the principles and methods are mostly independent of the nature of the organism, we will use the term company in this article.
The trainer should present the course content in a Top-Down manner starting with very concrete case studies such as the development of an AI model for an intelligent speaker that can react to a phrase like “Hey device, tell me a joke”. This method of teaching will help participants (especially those with no technical knowledge) to become interested in the course content.
Key points of the week 3
. Case study: Smart speaker
. Case study: Self-driving car
. The different functions in an AI team
. AI Transformation Playbook
. AI pitfalls to avoid
. Taking your first step in AI
Case study: Smart speaker
Company’s activities are often complex. An example of a complex AI project is the development of an AI model for a smart speaker that can react to a sentence like “Hey device, tell me a joke”.
Indeed, there are 4 unique processes that make up this action and therefore as many AI or software models to use. This set of processes is called the “AI pipeline” and usually there is one AI team per process.
The “AI pipeline” is even more complex (ie, more unique processes), so the final action can vary depending on what the user says as giving a duration. It is then necessary to train an additional AI model to recognize this duration.
The following slide summarizes the 4 main processes of this example as well as the numerous commands for each of which it is necessary to train different AI models.
Case Study: Self-driving car
There are 3 main steps for a self-driving car to decide on its route and speed, but there are many other localization processes that are necessary for the decision.
Most are based on training an AI model that can detect other cars, pedestrians, road lines, road signs, work in progress, traffic lights, etc. All of this location data will then allow a software “Motion planning” to decide the route and speed of the car.
The different functions in an AI team
For complex AI projects, it can have 100 or more people in the AI team because each AI model can request a dedicated group of people. Here are the main functions of such a team:
- Software Engineer: 50% or more of the team, in charge of developing softwares.
- ML Engineer: responsible for creating and training ML models.
- ML Researcher: the state of the art is evolving very quickly in the ML. The MLR is in charge of following this evolution, doing research and possibly publishing the results of its research.
- Applied ML Scientist: it is an intermediate function between the 2 others; the AMS will be in charge of adapting already published models to the specific projects of its company.
In addition to these 4 functions, an AI team may need the following specialists: Data Scientist, Data Engineer and AI Product Manager.
However, few companies can start with such a complete AI team. The following slide proposes to start with a reduced team up to 1 person who starts in ML/DL with online courses for example (the important thing in AI is to start a project without waiting to have an ideal environment!):
AI Transformation Playbook
As the head of Google Brain and Baidu’s AI department, Andrew Ng has gained experience on how to transform a business into AI. In order to share his knowledge on this subject, he published online the AI PlayBook Transformation that allows any organization to think about their own AI mutation strategy and implement it in 5 steps:
1. Execute pilot projects to gain momentum
Andrew Ng at Google preferred start applying Deep Learning in Google Speech projects then in Google Maps projects before interest Google Online Adverstising department that is clearly the most important at Google (ie, the one that brings the most revenue and on which Google has been developed).
This strategy minimizes the risk of failure to introduce AI in a company by gaining experience on non-priority projects, with a quick return (6 to 12 months maximum project) and with a an internal AI team or not depending on the possibilities of the company.
By choosing pilot projects with a high probability of success, you can then convince the rest of the company to use AI, especially on projects with a strong interest in the company.
2. Build an in-house AI team
Andrew Ng recommends creating an AI Unit at the same hierarchical level as the BUs (Business Unit) and acting as a provider for each BU. Indeed, it will be difficult for each BU to have within it an AI skill.
In addition, this AI Unit can develop a common AI platform for BUs that will make it easier to share the models and data needed to drive ML/DL models.
This AI Unit may be under the direction of a Chief Technical Officer (CTO), Chief Internet Officer (CIO), Chief Data Officer (CDO) or even a Chief Audit Officer (CAIO) which is an increasingly senior position. created in companies.
Finally, Andrew Ng recommends that this AI Unit be initially funded by the CEO of the company before acting as a resource center (ie, selling its services to the BUs) to facilitate its establishment.
3. Provide broad AI training
Training the technical teams of the company to the AI is necessary (about 100h) but not enough. It is necessary to train the whole of the company with contents adapted to their functions (to foresee a sensitization to AI of approximately 4h for the management team and approximately 12h for the heads of departments).
Important: before developing your own content, the CLO (Chief Learning Officer) must study with AI experts the many online resources including MOOCs and encourage employees to study them.
4. Develop an AI strategy
Andrew Ng prefers to place this step in position 4 according to his experience. Indeed, the company, especially the management team, must gain experience on the specifics of AI before defining its own strategy (for example: does the company have to buy data ?, create free services generating data?, etc.).
He proposes to define his business strategy based on the “Virtuous Cycle of AI”. It’s actually starting with a first version of a product to acquire users and therefore data that will improve the product and create a virtuous loop. Of course, this product must correspond to one of the key activities of the company, preferably a niche to avoid being in competition with major companies in its sector or even with the giants of the Web (Andrew Ng takes the example of BlueRiver Technology, which was purchased for $ 300 million after developing a weed-detecting ML model, allowing them to be killed without spraying the entire planting of chemicals: the BlueRiver Technology team started taking pictures of weeds with their smartphone …).
When it comes to AI, data acquisition (as well as storage in a unified platform) is a strategic goal for a company that wants to use AI. It is also necessary according to Andrew Ng to think about positioning his activity on an online network/platform. Indeed, this will generate a lot of data specific to your business and the use of AI models will further increase the value of this data for your business (examples of companies like Uber and Facebook).
Of course, your AI strategy must match (at least initially) the goals of your business. If, for example, one of its priority objectives is cost reduction, then you have to use AI to achieve this goal.
5. Develop internal and external communications
If your company uses AI, you must communicate externally to your investors, your partners, your customers and even to the public authorities, especially if you work in a sector regulated. This communication will ensure the value of your business, benefit its image and attract new talent. Do not forget to also communicate internally to reassure your employees about your AI strategy and also to associate them with it.
AI pitfalls to avoid
Andrew Ng informs us about 5 pitfalls to avoid when a company starts using AI but the tip number 1 is: start a first project without waiting, which will allow you to gain experience that will be useful to you for the second project and so on.
Taking your first step in AI
The important is to start and the amount of online resources on AI, ML and DL allows you to do it without delay and without waiting to recruit an AI specialist. Andrew Ng completes this tip with the following list:
About the author: Pierre Guillou is a consultant in artificial intelligence in Brazil and France. Please contact him via his Linkedin profile.