Artificial Intelligence Stack Exchange is a question and answer site for people interested in conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment. It only takes a minute to sign up.
Sign up to join this community
OpenCog is an open source AGI-project co-founded by the mercurial AI researcher Ben Goertzel. Ben Goertzel writes a lot of stuff, some of it really whacky. Nonetheless, he is clearly very intelligent and has thought deeply about AI for many decades.
What are the general ideas behind OpenCog? Would you endorse it as a insightful take on AGI?
I'm especially interested in whether the general framework still makes sense in the light of recent advances.
OpenCog is a project with the vision of creating a thinking machine with human-level intelligence and beyond.
In OpenCog's introduction, Goertzel categorically states that the OpenCog project is not concerned with building more accurate classification algorithms, computer vision systems or better language processing systems. The OpenCog project is solely focused on general intelligence that is capable of being extended to more and more general tasks.
OpenCog's knowledge representation mechanisms are all based fundamentally on networks. OpenCog has the following knowledge representation components:
AtomSpace: it is a knowledge representation database and query engine. Data on AtomSpace is represented in the form of graphs and hypergraphs.
Probabilistic Logic Networks (PLN's): it is a novel conceptual, mathematical and computational approach to handle uncertainty and carry out effective reasoning in real-world circumstances.
MOSES (Meta-Optimizing Semantic Evolutionary Search): it implements program learning by using a meta-optimization algorithm. That is, it uses two optimization algorithms, one wrapped inside the other to find solutions.
Economic Attention Allocation (EAA): each atom has an attention value attached to it. The attention values are updated by using nonlinear dynamic equations to calculate the Short Term Importance (STI) and Long Term Importance (LTI).
OpenCog lists 14 competencies that they believe AI systems should display in order to be considered an AGI system.
Perception: vision, hearing, touch and cross-modal proprioception
Actuation: physical skills, tool use, and navigation physical skills
Memory: declarative, behavioral and episodic
Learning: imitation, reinforcement, interactive verbal instruction, written media and learning via experimentation
Reasoning: deduction, induction, abduction, causal reasoning, physical reasoning and associational reasoning
Planning: tactical, strategic, physical and social
Attention: visual attention, behavioural attention, social attention
Motivation: subgoal creation, affect-based motivation, control of emotions
Emotion: expressing emotion, understanding emotion
Modelling self and other: self-awareness, theory of mind, self-control
Social interaction: appropriate social behavior, social communication, social inference and group play
Communication: gestural communication, verbal communication, pictorial communication, language acquisition and cross-modal communication
Quantitative skills: counting, arithmetic, comparison and measurement.
Ability to build/create: physical, conceptual, verbal and social.
In my opinion, OpenCog introduces and covers important algorithms/approaches in machine learning, i.e. hyper-graphs and probabilistic logic networks. However, my criticism is that they fail to commit to a single architecture and integrate numerous architectures in an irregular and unsystematic manner.
Furthermore, Goertzel failed to recognize the fundamental shift that came with the introduction of deep learning architectures so as to revise his work accordingly. This puts his research out of touch with recent developments in machine learning
While my knowledge of OpenCog is very limited, you could say that yes, it does still make sense and it is insightful. I'm not certain regarding all of the components of OpenCog but I do know that at least one component is relevant (I think it's part of the MOSIS component).
This component is very similar to Numenta's hierarchical temporal memory which is based more on computational neuroscience than plain math; however, I would consider Nupic a more relevant project in terms of neroscience though both are attempting to emulate components of the brain. In my opinion, such projects are far more impressive than what's going on with typical convolutional neural nets, RNNs, etc. which are too loosely related to what goes on in the brain to be said to be computational neuroscience.
That's not to say that things like ANNs, GAs, etc etc are useless for AGI. We don't really know since we don't have an example of one.
