top of page

Intuition Learning - Way to AGI

Updated: Oct 24, 2023

If you're a CXO, founder or investor - follow me on LinkedIn & Twitter, or join my newsletter on my website here. I share latest simplified AI research and tactical advice on building AI products.


Please note this is my original research work. Do not use without authorization and citation.



Abstract


In this paper I propose a new learning architecture called as Intuition Learning. Current learning methods like supervised learning, unsupervised learning, deep learning, and reinforcement learning, even with so much progress and hard work, are only able to simulate a narrow set of real life human behaviours. For progressing towards generalized artificial intelligence (AGI) we need a type of learning which (1) can learn things on own without any training (2) can learn things beyond human knowledge. None of this is possible with existing approaches. To achieve these goals, a new way of learning called as Intuition Learning (IL) is proposed here. It uses a knowledge graph as memory with data nodes capturing data (information) and skill nodes capturing behaviours (like driving, eating). During sleep phases or naps or live-on learning these models run random currents through a random series of nodes and tries to add skills possible through these nodes. This approach on analysis proves sufficient for achieving general intelligence on it's own satisfying both (1) and (2).



Proposal


To achieve Artificial General Intelligence (AGI) we need an approach which satisfies following two conditions -

  1. models can learn things on own without any training

  2. models can learn things beyond human knowledge

Unless these two conditions are satisfied, AGI is not possible. Why ? Because unless (1) is satisfied model will not be able to keep up with advancement of world events, shift in culture and evolution of information. And (2) is needed because even humanity itself is not generalized enough in absolute reality.


Supervised learning fails both (1) and (2) because it needs new labelling of datasets for learning new skills and also learning skills which humans do not know is not possible.


Deep Learning also fails both (1) and (2). To learn a totally new skill (like diving) for a neural network (like ChatGPT), you need to train from scratch and labelled datasets are needed. Even if (1) is satisfied using an ensemble approach where specialised models work under a central brain, (2) is not possible. Neural networks lack imagination !


Reinforcement learning similarly fails both (1) and (2). An agent can not learn skills on its own belonging to an environment which it has not seen.


Lastly, Unsupervised learning methods also fail both (1) and (2) since algorithms need to be pre-defined for learning a known human skill.


Thus, current existing approaches are limited by their very definition to reach AGI level intelligence.


Therefore, here a new learning method known as Intuition Learning (IL) is proposed. IL utilises concept of memory using knowledge graphs and generates skills based on self-run currents on random node paths. This simulated human way of learning but at the same time also automatically accelerates it.


As more and more information is added (in data nodes) and skills are added (in skill nodes), more and more skills become possible. This happens during sleep or nap phases when system dedicates full power to learning on latest graph periodically. For high compute power systems, even continuous learning is possible in this way.


This approach achieves (1) by this current propagation(intuition currents) method and (2) is possible because of absence of limitations of human skills and knowledge.


Thus, this system can keep scraping web data and enriching it's data nodes and continuous learning of new skills based on (even new type of) data is guaranteed.


Notably this approach is also very beneficial in terms of training time and cost. Since it's a most generic approach, manual training for each domain separately is not required. And once learnt this knowledge can be transferred to all other IL systems using Knowledge Graph node transfer. IL systems will grow on their own with time.

Training cost and time are reduced since days or months of GPU computations are not required since this is much faster (and simple) way of learning.


It also improves accessibility in production for companies since even low compute power systems can also perform IL learning, albeit needing more sleeps or naps.


Lastly, explainability is achieved due to simple and transparent learning process.


Thus, Intuition Learning approach is desirable in terms of all learning achievements, time, money, transparency and accessibility to all.



Risks


Intuition Learning will give rise to autonomous AI very fast so it is important to establish control mechanisms in such systems from day one.



Future Work


Next I am working on implementing this architecture and in future run experiments for industrial applications.






7 views

Comments


bottom of page