AI is worth learning and not too tricky – if you can get your head around key frameworks

Starting with: TensorFlow

Little girl concentrates hard on a book she's reading

M3 The hype around AI promises interesting work and fat paychecks, so no wonder everyone wants in. But the scarcity in talent means that researchers, engineers and developers are looking for ways to pick up new skills to get ahead.

The best framework to get stuck into if you want to be employable in AI is probably TensorFlow, Barbara Fusinska, a data scientist and freelancer for Katacoda, a learning platform that provides tutorials in languages commonly used by software engineers, told The Register.

Fusinska zipped through the online Katacoda lessons and offered advice during a lecture at our Minds Mastering Machines conference in London, UK, earlier this month.

The term "democratizing AI" can be fuzzy and is often thrown around by companies, but open source frameworks like Google’s TensorFlow, Facebook’s Caffe, PyTorch or Amazon’s MXNet are a concrete example that shows how anyone can use the tools.

It can be confusing to decide which framework to learn. Each one has its own advantages and disadvantages, and depending on what metrics and benchmarks are tested TensorFlow isn’t always the best, but it is the one most commonly used.

“This means that it has the biggest, thriving community that you can turn to for help and support,” Fusinska said.

TensorFlow isn’t the easiest of languages, and people are often discouraged with the steep learning curve. There are other languages that are easier and worth learning as well like PyTorch and Keras.

It’s not always best to tackle AI purely by playing around with code even if you are a strong programmer.

“Experimenting with pre-built systems and APIs will give you results, but if you don’t understand how the numbers were produced, you won’t interpret them well. People are smart enough to realise how they work over time, but it’s a lot easier if you start with a little theory,” she said. “Go away, and learn some maths - like matrices - and try to understand how machine learning and neural networks work or there will be a longer road ahead of you”.

It’s helpful to learn the different architectures and types of neural networks so you know how they can be used.

There are a range of online resources starting at various difficulties such as courses from Coursera taught by Andrew Ng, or nanodegree programs at Udacity, and Fast.ai offers a practical class for coders. Many developers and researchers are also keen to share knowledge in blog posts.

The best way to learn is to pick a project. “Try to do something real like taking ImageNet to build a classifier,” Fusinska recommended. “But don’t pick something too hard, if you have no idea where to start then the problem is probably too hard to start with.”

“Make a list of what don’t know and what you need to know and then go away and learn it.” ?


Biting the hand that feeds IT ? 1998–2017