Human-AI feedback loop

AI safety is very real and it’s very important that every AI practitioner, AI researcher and organizations/teams implementing or using AI should give AI-safety the first priority rather than an afterthought. Let’s try to analyse this thought in the upcoming sections.

The feedback loop

Whenever a machine learning model is being trained, including a “feedback loop” is helpful as it lets the algorithm make minute corrections to it’s predictions overtime. In most of the cases during the training phase this feedback loop is usually another algorithm.

What’s the problem with this approach?

The problem with this approach is that we still don’t fully understand why a ML algorithm makes a particular decision, the way it does. So, being overly dependent on an algorithm to provide feedback to another algorithm would not be a wise decision. A possible solution to this problem would be to involve a human at some point in the feedback loop, so that the machine and human can work together in order to rectify any kind misjudgements on the machine’s part.

Can you give some examples of how this can be done?

An example would be including a human driver to judge a self learning car’s decision to make a particular driving manoeuvre or correcting the car’s judgment on differentiating a pedestrian from a tree. Another situation where Human-AI feedback loop would be helpful can be in document processing products/solutions where machine learning is being used and where training data is in short supply. Here a human can help point out mistakes made by algorithms during training time and can minimize the need for huge amount of data.

Conclusion

If we want to build safe, reliable AI then the focus should be first on including humans in the feedback loop and designing AI which can learn from humans and even teach humans in turn.

A nice article on including humans-in-the-loop can be read here and here.

If you happen to like this article then do give a clap and share this across. If you have some comment or thought to share in this topic then do share the same in the comments section.