Interaction of Crowdsourced models with OpenCog Hyperon in the GCN

Interaction of Crowdsourced models with OpenCog Hyperon in the GCN

OpenCog Hyperon is a new faster version of OpenCog that uses embeddings to ground logic to external processes for improved neuro-symbolic reasoning. SingularityNET's premium AI software, OpenCog, consists of a knowledge graph and a collection of cooperating AI agents reading to and writing from the knowledge graph in a cognitively synergistic manner. Agents of OpenCog Hyperon write to the knowledge base in Metta, a dependent type language with DSLs for various AI types such as probabilistic logic networks, neural networks, evolutionary computation, pattern matching, and so on. AI-DSL is an overarching DSL for AIs that allows SingularityNET services to communicate with one another. Metta is also a functional language that employs folding, but both folding and dependent typing are methods of performing proofs and problem reductions. The knowledge graph is a concise, manipulable representation: an atomspace, which is a probabilistic dependent type metagraph. Because a type is gradually determined, the atomspace is probabilistic and is treated as a distribution of possible types. It is a metagraph because its links can attach to other links, nodes can be graphs, and nodes and links can both be typed. Logic is implicit in the atomspace paths via a Heyting algebra. The knowledge graph can keep track of its own implementation, allowing for self-reflection and, thus, improvement. It begins with an explanation because the reasoning is transparent and serves as its own explanation, as opposed to having to find one in black box methods such as neural networks.

OpenCog comes with AI modules that act synergistically, including the MOSES probabilistic evolutionary computation module that creates logic programs, a stochastic pattern matcher and a probabilistic logic rule engine. All of these do the same type of thing in different ways. For example, at Rejuve, MOSES has been used on gene expression data to find which genes are expressed before a gene switch has turned on at around age 60 and which genes are expressed after, to better characterize aging¹⁰ MOSES can find such patterns and store them in the knowledge base, but the pattern matcher and the probabilistic logic engine can as well, albeit with different sets of strengths. The AI modules operate on the same atomspace and thus complement each other's efforts. The neural embeddings, in particular, sub-symbolically ground symbolic logic in environmental stimuli and connect the knowledge graph to the other neural models of our Generative Cooperative Network.

Rejuve has used word2vec embeddings of the atomspace created with a deepwalk algorithm for creating novel, promising hypotheses regarding the SNPs unique to supercentenarians. SNPs, or Single Nucleotide Polymorphisms, are variations in the genetic code. SNPs are found in about one out of every 1000 locations that differ. Statistical methods are commonly used to find SNP correlations; however, by including all of the hypergraph's relations, we have a more holistic space that may find more indirect correlations than traditional methods. Thus, Hyperon improves the GCN embeddings while in turn, the embeddings improve Hyperon.

MOSES has also been applied to personalized medicine for chronic conditions of aging. For example, MOSES has taken genetic expression profiles, covariate conditions, and treatments in 15 different studies to predict treatment outcomes, to recommend as the next in a sequence of trials in n of 1 studies.

Figure 2. OpenCog Hyperon Cognitive Architecture

¹⁰ Goertzel et al (2020). Embedding Vector Differences Can Be Aligned with Uncertain Intensional Logic Differences. Artificial General Intelligence: 13th International Conference.