TEJ KUMAR

This sounds interesting, and Google is not the only company working on this problem either. I wondered how Google is addressing this challenge? I began exploring and what I found was very interesting, especially in terms of the future of technology.

When thinking about the idea of automation, there is a lot of catch-22. On the one hand, I don’t want to be a lazy person or someone who doesn’t do as much, I want to be frugal and use the cheapest services.

Which is the wrong way to think, because the essence of automation is to stream data in the closest and most efficient way possible. Google does this from time to time using its artificial intelligence and machine learning capabilities.

In recent years the company has been experimenting with more than just the machine learning technology. It has also been trying to implement artificial intelligence into more places, like search and information search where it shows you specific results. As well as this it was recently working with a branch of academia called Brainlab on a deep learning model.

They created a more complex approach than the one used by Google’s Deep Learning and Machine Learning labs. At first the idea was to build a neural network that could identify things as speech. It took over a hundred hours of human training and a week of human training to learn its domain. For that it is worth noting that there’s a lot of human intervention.

So what Google’s approach for organizations like Google or Minds, a startup incubator that Google Ventures has invested in. They basically removed the human from the loop. However, it is going to take time before the model is truly reliable enough to go to production. These systems were created to aid manual work by processing information in a way that doesn’t require a lot of human input.

Even then it requires a lot of accuracy to really make a significant difference in how work is done. However, what Google may have accomplished is what is called an “adversarial attack.” For a simple explanation, an attacker produces a probability that something will happen (a label), and then it makes a prediction for it. For Google’s machine learning model, the fact that it needs human input means that it is more likely to be subject to an adversarial attack.

Google’s solution is to avoid adversarial attacks. It was been using a neural network network that only predicts the probability of an event. If the network predicts anything, or it is vulnerable to an attack, Google will output a label like “Warnings” or “erroneous”. But if the prediction gets it wrong, Google’s software will warn the user. This is done to prevent a situation where the human gets to live and work in front of an adversary and is victim of a malicious attack.

Their solution is a lot like the method they use to create bad actors. Here is the irony for organizations that take the time to do this analysis is that by creating such programs to protect their products is as if they want to increase their chances of getting attacked, or are creating malicious hackers.

But Google doesn’t have something perfect yet. Automation tools aren’t able to improve the level of quality of the training, it’s still mixed signals. For example one training has corrected 20% errors, another has corrected up to 100%. In reality it’s probably very unlikely that it will correct anywhere near the amount of errors.

To fix this Google has developed a multiple million prediction record engine. Even though it has the ability to be more efficient than Google’s Deep Learning effort, it is probably not in a position to support a real production like Google’s.