The 7 most important things to consider before you embark on your AI projectJulie WhitingFebruary 21st 2019
The current generation of ML-based AI systems are amazing at doing one job, that is all they can do. So, when getting to grips with AI and how it can be used within your business, it is important to keep this fact in mind. What do we need to do to put these technologies together so that they work efficiently and achieve the goal in mind?
We had a chat with one of our Solution Architects, Ed Yau, who talked us through what Artificial Intelligence helps achieve in today’s business and technological world, as well as how you can apply it, and who you’ll need to help implement it correctly and efficiently. So, let’s get down to it!
Is it possible that AI can do things just as well as or even better than humans? Quite simply, yes! There is a long list that includes image recognition,voice transcription, star craft, driving, reviewing x-rays, predicting demand for warehouse parts, and more. What’s even better though, is the potential to apply this sort of technology into a business environment in order to raise your organisation’s productivity. However, we need to bear in mind that whilst the current generation of ML-based AI systems are amazing at doing one job, that is all they can do. So, when getting to grips with AI and how it can be used within your business, it is important to keep this fact in mind. What do we need then in order to put these technologies together so that they work efficiently and achieve the goal in mind? Imagination and experience are the main ingredients to this task- as these technologies can only do one thing well, you need to think about how (certainly in the short-term) you can get AI to be thought about as people enhancers, NOT people replacers.
So, we’ve thought about how we should see AI, but how can you apply this to your business projects? NUMBER ONE: To implement AI successfully, you need three things; people who can tell you the difference between what is feasible and what is sci-fi, a range of candidate business problems that you want to apply AI to, and as much data as you can get your hands on that relate to these business problems you are facing. Once you have all three, you can then start to have a conversation and understand what can be done. However, acquiring all three doesn’t automatically guarantee you a successful project.
OK, so how would you start your first AI project? Welcome to NUMBER TWO: What we would recommend, before you spend your entire R&D budget on a huge AI project, is to get some consultancy to help you explore some ideas. It's common sense to start small, do a discovery exercise that audits your data, and understand where the opportunities for this application are within your business. We have the benefits here at cloudThing, through working with large enterprises, that we understand what data can and cannot be used, or what you need to do if you want to keep your compliance people-happy.
NUMBER THREE: Who should you hire to help you with AI implementation etc? Admittingly, this project is quite a Dev-centric perspective, but in our opinion, most of the main AI systems entail 95% software engineering and 5% AI. Didn’t expect that figure, did you?! So, do we really need loads of personnel to help with this? The answer, yes. To make this AI outcome effective and efficient, it means that a detailed discovery exercise must be undertaken, along with a mixture of skills including some data analysts, data architects (potentially), data scientists and solution architects who understand AI. Without these people, there will not be enough knowledge and skill in order to make AI worth it in your business. So, in order to obtain these attributes, you need to go to a software consultancy and find some developers- what a hassle. However, this is where cloudThing can help as we have all of these talents under one roof- there is light at the end of the tunnel!
NUMBER FOUR: What technology should you use? All the cloud vendors these days have a rich range of AI-enabled PaaS products- Microsoft made an early move here and Microsoft Azure has one of the strongest line-ups. With an AI PaaS product, you can treat your AI as a black box in the cloud and make use of the Azure Cognitive services that come pre-trained on million-entry datasets. They've been around for a while, but Microsoft still demonstrate these services at every opportunity as they have a big ‘Wow’ factor. I think of Cognitive services as a way of letting us blur the line between the machine-technological world and the real world on your applications to create a more interesting UX. These days, there are also things like CustomVision or CustomSpeech that you can train with your own data in your specialised domain, but if you really want to customise this to your business, or you have a seriously large user case for AI, these services will cost you a lot to run. So that's when you're into the likes of pytorch, keras, tensorflow or plain old python with Numpy.
Python is what most Machine Learning (ML) projects use, but these days on the Microsoft stack, from SQL Server to .NET, you can get ML libraries. Therefore, programming a language isn't really a barrier, the real barrier is understanding how to apply these libraries to your business problem.
NUMBER FIVE: How do you minimise risk and generate ROI, as these are key questions within the success of a business? The main thing is to start small and pick a project where there is a clear and demonstrable business benefit that you can build a strong business case for. Then you can apply really strong project governance to monitor the results of your project, working in cycles with a process like CRISP-DM in defined phases, so you can keep on justifying your continued investment, or not as the case maybe! As there are risks within all projects, you must stay aware and understand the implications of these risks and how is best to minimise them.
NUMBER SIX: Data; is it really all about the data? How much data do I need? This is a simplification, but ML-based AI as we see being used, is a form of Inductive logic. I.e. it learns from seeing multiple examples from your data and is then able to make inferences on what is likely to come next. The more data you have, the more likely it is that ML-based AI will be able to help you achieve something meaningful. But even if you only have a few hundred rows of data, you will be able to do something. The only disadvantage of not having much data is that your outcomes may not be as strong and therefore, may be more fragile.
NUMBER SEVEN: Last, but not least- Compliance! GDPR- the 4 letters that we should all be familiar with as it’s still fresh. In a nutshell, GDPR states that you need consent from users to process their data and that any automated decision-making routines need to be explicable if challenged. For some black-box AI systems, this may present some issues. Work is being done in this area but you still have to be aware of the issues. Another ‘hot potato’ is Ethics. If you are an Asimov fan, you will know about his three Laws of Robotics. An example that can be used is last year when Amazon announced that they were shutting down their AI recruitment tool as they realised that it was favouring male CVs for technical jobs. You'd think it'd be quite easy to strip this out. But when you read about how it worked, you realise why it was difficult to eliminate. It was looking at past hires of that person and so then assigned a better ranking to CVs with words that featured on that data set. What was this AI recruitment tool doing? It was actually down-grading attendees of particular women's colleges. And I'm no linguist, but apparently there are some verbs generally more favoured by male job applicants - those like 'executed' and 'captured', and it was this element that the AI was also picking up on.
These steps and acknowledgements are all important in implementing AI and understanding what it is there to do, first and foremost, as well as knowing which personnel are required and what the actions needed are, and the positive and negative implications that can occur.