With so much media at present using the term “Artificial Intelligence” (AI), I wanted to provide a brief overview of what is meant by the term “Cognitive Artificial Intelligence” (COG).
AI is being used as a catch-all phrase for any software system that performs a function deemed to be clever.
Phrases such as “learning” or “understanding” are often coupled within the same information, although the truth is most often a little less glossy and not entirely consistent with what we as humans may normally refer to as learning or understanding.
You may have heard of “Deep Learning”. Deep Learning is actually another catch-all phrase relating to a collection of algorithms. It is commonly associated with actions that require some form of pattern matching and usually with very large volumes of data. Through the application of these various algorithms, Deep Learning processes have rapidly improved the reliability and resolution of the actions that implement it.
Speech recognition is one such example. The improvements in accuracy were so dramatic when Deep Learning methods were applied, that some companies reported as much as a 60% overnight improvement in accuracy.
Xuedong Huang, Microsoft's Chief Speech Scientist recently stated that their system is almost on par with human level speech recognition. A truly amazing achievement.
This is where things get a little murky though.
The classification of Artificial Intelligence
Speech recognition performs one task and one task only. A stream of audio data is fed into the speech recognition service and it figures out the words that have been said. It then returns the recognised words in order (to form a sentences) in text form so the words can be further analysed by software. That is it.
Of course, this is an over simplification of what is going on in the background, but the main point being here that the speech recognition services has no understanding of what the words mean. No understanding of the intent from the person who spoke, no indication of the overall context that the words apply to. It just recognises the words and returns them.
As the software developer, you could receive the words back and then perform a “keyword” match in the sentence to execute some action.
- “Turn the lights on.”
- “Play music”
- “Play hip-hop music”
- “What is the capital of France?”
While this can be effective in some applications tightly bound to a specific set of functions, this isn’t intelligent and also doesn’t demonstrate “knowledge”. As a result, the same input mechanics can’t be applied to other actions.
This is what a lot of existing “AI” systems do, and although one could consider them to be clever, we would not consider them to be intelligent.
Cognitive Intelligence is about replicating the mechanics we humans use to interact with people and the world around us. This could be in the capacity to learn, or how we reason and understand various concepts. It is not about trying to replicate the human brain, but about creating incredibly flexible solutions that we can interact with easily, without needing to know keywords or magic phrases.
Strong AI is sometimes referred to as Artificial General Intelligence.
The purpose of building a COG is to interact with humans in a manner that is consistent with being human. This includes emotions, understanding who and what objects are, what their purposes and functions may entail. It also includes the areas such as getting things wrong and learning from mistakes.
It can’t be done however, without the use of the existing services previously mentioned. So while those systems don’t represent intelligence from a cognitive stand point, they are vital to the creation of a cognitive solution.
The figure below shows a simplified view of an overall architecture where the components containing AI are used to serve a COG intelligence system.
A cognitive artificial intelligence solution is a different approach to working with returned data from various systems. Again with speech recognition as the example, a COG not only analyses the words, the phrases, the structure of the sentences, but it looks to resolve all of that in a layer of understanding.
The COG extracts context or applies these words to context. It combines data input from vision systems and emotion detection (and much more) all at the same time to determine not just what is being said, but how it is being said, and then how to respond accordingly, and why some responses are appropriate at one point while not being appropriate at another.
The COG is an extremely complex challenge. It involves so much data of different types being moved around and used in such subtle ways that it is often difficult to test specific functionality. Unlike conventional software programs that execute blocks of code that can be tested against, the COG runs on dynamic logic. This means that while the software system itself can be running completely fine with no errors at all, the logic that is the result of all the data translations could be completely broken or not make sense at all.
A request to turn some lights on, could result in a blender being turned on instead, because the logic connects the spoken words to the function, rather than any hard coded instruction.
One of the greater challenges is with the COG being able to “program” itself dynamically. This is where the learning part comes in. And where the real dangers of "Strong AI" reside.
Achieving a self-developing solution is not trivial but doable. Controlling what it learns and how it considers that knowledge (good, bad, right, wrong) is extremely hard. But without this, a COG system will never evolve in its intelligence or possible functional usability. Therefore a balanced approach between curated and automated learning needs to be taken to ensure safety.
Bringing it all together.
In wrapping up, a cognitive artificial intelligence system has the capability to understand, reason and learn with similar processes that we as humans naturally develop. It is also the product of many other AI based components which feed the cognitive architecture.
Being able to recognise speech in audio was an incredible challenge that some really smart people have worked on for decades. They should be recognised and applauded for the enormous successes they have achieved, which now make cognitive artificial intelligence possible.
Detecting words was one thing – but figuring out what those words actually mean and what to do with it may just be an even greater challenge.
This is cognitive artificial intelligence.