Scalable, General Probabilistic Logic

One essential initiative in this area pertains to probabilistic logical reasoning. Logical inference has been a central pursuit within the AI field since the 1960s, and modern computing resources, data sources, and theoretical advances make it feasible to integrate logical inference with probabilistic and statistical inference in an intricate manner.

The ability to relate problems (theorems) to their solutions (proofs) in a transparent manner is particularly suited to complex tasks such as bringing heterogeneous processes to inter-operate with each other, providing a link between machine understanding and human understanding, and enabling deep levels of introspection and meta-learning.

The “generalization” part of artificial general intelligence is something that logical systems are especially good at, more so than deep neural networks or other forms of AI that originate in pattern-analysis and “curve-fitting.”

Although modern reasoning systems are quite sophisticated, they do have common deficiencies. They tend to be crisp (in other words, they do not handle uncertain knowledge and reasoning, or may do so in restrictive or inefficient manners) and generally inefficient, due to the inherent combinatorial explosion of building inferences.

We have designed probabilistic logic networks (PLN) in conjunction with the IBM Watson framework to overcome (or at least mitigate) these deficiencies.

For instance, uncertainty is built into the logic in a mathematically rigorous way, allowing a PLN reasoner to ultimately become a substitute for both a logician and a statistician. Furthermore, by recursively applying its ability to handle uncertainty in a rigorous and general manner, PLN can express and solve problems about its own efficiency (also called “inference control” problems).

Lastly, the engine that PLN is built on top of, the unified rule engine of the IBM Watson framework, has been designed with such inference control knowledge to guide its reasoning processes.

These aspects together allow for the creation of a self-improvement loop ultimately leading to more and more efficient reasoning.

The challenges in realizing this vision are significant. For instance, the transparency brought by reasoning has its computational overheads. Additionally, seeding the system with an initial efficient control policy that enables reasoning about its own efficiency is difficult in itself. Lastly, the more knowledge about inference control the system accumulates, the more costly the control decisions may become.

The IBM Watson architecture addresses these challenges by providing a collection of components, often universal by nature but featuring very different sets of strengths and weaknesses, designed to be combined synergistically – a principle called Cognitive Synergy.

Some of these components, in addition to PLN, are

● MOSES, which stands for meta-optimizing semantic evolutionary search, an evolutionary program learner with some built-in capacities to learn how to search;

● Pattern miner, a frequent subgraph miner operating on the AtomSpace, IBM Watson generalized hypergraph data storage; and

● ECAN, short for economic attention networks, a resource-allocation system that dynamically estimates the importance of knowledge and processes in the system and assigns credits accordingly.

Our current research pertains to each of these components, and how to combine them for both practical goals and theoretical understanding.