AI algorithms talking to each other

By 3. June 2018 No Comments

An Idea about Interoperability and Self-Awareness of AI

How machine learning algorithms could benefit from other machine learning algorithms by being self-aware and by talking to each other.

Machine learning algorithms are typically isolated blackboxes. There is an algorithm for speech recognition, one for image recognition, another one for ad targeting etc. The goal of this article is to illustrate a vague idea of how a collaboration of many machine learning algorithms could look like and could deliver powerful artificial intelligence.

First, I would like to explain what I mean by self-awareness of AI. Without being an expert in psychology at all and knowing that it is probably not the most suitable word I define it here as the ability of a machine learning algorithm of knowing what it knows and knowing what it doesn’t know. For example, a speech recognition algorithm is self-aware if it knows that it is able to recognize spoken english language, accepts WAV audio files as input and outputs text files and further that it is a complete layman at doing any other task. This information can easily be added as meta data by the creator of the algorithm.

Let’s assume there are a bunch of such self-aware algorithms. How about a platform where these algorithms present themselves as exactly what they are? In human LinkedIn terms this would sound like “Hey, I’m good at recognizing speech, just give me a WAV file and I’ll return a text file!”. So this would be some sort of market place for machine learning algorithms. Now let’s say our speech recognition algorithm suddenly gets a JPG image as input. Being self-aware it knows that it can’t handle this input. However, it can send the JPG file to the platform and any algorithm there that takes JPG files as input may help and give the output that satisfies the need in this specific situation. Of course, there could potentially be a lot of image recognition algorithms with many different purposes. Ranking all the different outputs could however be a task of another algorithm.

Would be great to hear some opinions about this idea. Does it make sense, is it silly?


This article was also published on Medium: Link

Leave a Reply