The learning curves for given architecture of Artificial Intelligence
Each particular technology that poses ability to learn will produce next learning curves. By looking at this curves we can suggest if particular architecture achieved the human-level abilities.
Let’s look closer at each part.
List of skills — all possible tests or tasks that at least one human can pass. Imagine this pile of skills: the whole variety from distinguishing geometry shapes to speech recognition, all possible games from mastering chess up to mastering Go, from making music up to new scientific discoveries.
The difference between people and machine is that human can acquire any of these tasks by choosing right curriculum and machine cannot. No need to reinvent something. The curriculum helps to acquire well-studied knowledge and abilities using the most effective methods.
Also, humans have the huge list of common abilities that help to acquire more complex and specific skills. For example, the ability to read, write and speak gives the infinite power of learning any other subject: building planes, play piano, quantum physics.
List of skills axis. From here it’s possible to assume that more common skills have to be acquired at first — to help acquire more specific fields of skills. With this principle, we can build kind of sequential list of skills — from most common to the most interesting from the science point of view. These skills were reflected on an image as human-level skills.
Time of learning axis defines time consumption for learning with given computational or other resources. The problem here is that higher level skills require more computational power. Because of it, learning time becomes huge.
There are two curves of possible learning. The first half of curve A(from point O to point N) is similar to B to some extent. With one difference — curve B may have architecture limit and in this sense are same to A or may not have it. The second option leads us to the topic of technological singularity.
What to do to create Human-level Intelligence?
Basing on this graph there are two complementary ways to achieve human-level intelligence:
- Decrease learning time. Such sequential list of skills has to be created where each learned skill will help to acquire next one with higher efficiency has to be created. Such curriculum will make learning curve grow faster.
- Push up the architectural limit by developing the park of architectures. Okey, it’s obvious. But by using the proposed view we can measure the progress. And measuring makes the development of AI real objective science.
So each project that aims to create Artificial Intelligence may be considered under proposed scheme. And each project narrows this approach by proposition own hypothesis and vision. Proved hypothesis becomes new the basis of research. And the science of Artificial Intelligence emerges.
Don’t miss my new article:
Periodic Table of AI at the Edge of Chaos_This article is a compilation of ideas and insights that are beneficial for Artificial General Intelligence development…_medium.com
Call to Action:
- Pr_ess the heart_ or share the article if you found this article interesting
- Follow me on Medium or Twitter so you see my upcoming articles in your feed.
- Answer the question: What is your interest in AI?