Influenced by media and scientific papers that report mainly the progress in AI, people tend to think that AI can do anything. Well.. that is not true. Despite the significant advancements, there are still many limitations where AI is not as effective as humans. Before you start to work on an AI projects, you should be familiar with the strengths and limitations of AI:

AI is good at:AI is not good at:
– Spam filtering
– Language translation
– Speech recognition (audio to text)
– Self-driving car
– Visual inspection (e.g. find defect products in manufacturing)
– Online advertising
– Categorisation
– Detection (cars on the road)		– Market Research
– Showing empathy
– Reading intentions (e.g. interpreting gestures)

Ok, you want to get hands-on experience and work on an AI project. And although everything sounds really cool, it’s preferable to start with something feasible, something you can execute successfully.

How to choose a feasible problem:

1) Choose a simple concept ( that would require < 1 sec of human thought.)

2) Go with a topic you have lots of data available for

Machine Learning project workflow

  1. Collect relevant data : audio files, images, etc.
  2. Train a model the A to B mapping, (e.g. image classification, image detection). Iterate.
  3. Deploy the model  (then fine-tune with the collected data)

Case study A: How to build a Speech recognition system

  1. Collect audios with people saying the trigger words (e.g. “Alexa” and “Hello”)
  2. Teach the model to map the audio to the correct word “Alexa” or “Hello”
  3. Deploy. Fine-tune for accent for example and iterate..

Case study B: How to build a self driving car

  1. Collect images showing positions of other cars. Mark the cars with a rectangle to differentiate them from the background.
  2. Train the model to detect cars on images.
  3. Deploy the model – maintain and update.

Data Science project workflow

  1. Collect Data : e.g. user behaviour
  2. Analyse data to get insights
  3. Suggest hypothesis and actions (deploy changes, iterate)

Case studies

  • Optimising manufacturing line and reducing number of defect products (e.g. finding optimal conditions)
  • Optimising sales funnels