Training machines sans bias will augment humans: Amazon Web Services executive

Last year, Facebook Artificial Intelligence Researchers (FAIR) had to shut down one of its Artificial Intelligence (AI) systems after chatbots started speaking in their own language, defying the codes provided. There have been other instances too where anomalies in the AI models were noticed.

However, according to Olivier Klein, Head of Emerging Technologies, Asia-Pacific, at AWS, which is retail giant Amazon’s Cloud arm, a Machine Learning (ML) model will always operate the way you’ve trained it.

“If you train a model with a bias, you would end up with a biased model. You continuously need to train and re-train your ML model and the most important thing is that you need some form of feedback from the end-consumers,” Klein said.

“I think there are certain elements of human emotions like empathy that cannot be automated. There will be scenarios where it makes sense to automate and give customers better experiences. ML is absolutely not about replacing humans but enhancing the experiences,” he explained.

The success depends on the data points, or observations, that you put into the deep learning or neural networks.

“If your data points are very small and minimal, you probably end up with a model that is not really doing what you want it to do. So, it always goes back to what’s the data that you’re collecting and what are you training ML models on,” explained Klein.

At AWS, he and the team are busy adding unbiased data inputs to ML models, building services around those for enhanced consumer experience.

“We keep training and retraining the ML models and optimising those. Take, for example, our Amazon Rekognition service. It has a variety of different capabilities like object detection, object recognition, sentiment detection, etc. One of it is also facial recognition,” said the AWS executive.

Scientists have developed an AI system that can identify people on video, detecting their age and gender more quickly and accurately. This has become the basis for offline detection systems in Android mobile apps, per researchers from Higher School of Economics in Russia. Modern neural networks detect gender on videos with a 90 pc accuracy. The situation with age prediction is much more complicated. Traditional neural networks consider discrete values of age. In each of the video frames, the network estimates the probability of the person in the image being of a certain age. Like, if in 30 pc of the frames the top prediction of the network is a person’s age as 21 years, and in 10 pc as 60 years, its conclusion will be as follows: with a probability of 30 pc this person is 21, and with a probability of 10 pc, he or she is 60. Due to various conditions of observations or even slight head rotation, prediction of the same person’s age in different video frames varies in the range of 5 years, plus or minus.