By Dr. Manuel Aparicio IV
Cognitive Computing is defined as more brain-like and more human-like thinking. This begs the question, “How do we reason?” Douglas Hofstadler and Emmanuel Sander have written Surfaces and Essences: Analogy as the Fuel and Fire of Thinking, arguing that every moment of every thought is based on analogy. Analogy is based on reasoning from past experience. We reason by similarity or dissimilarity of how a current situation is informed by memories of the past. This has been Saffron’s position for many years: memory-based representation, inspired by our brains, and similarity-based reasoning, inspired by our minds, are core to true cognitive computing.
We have described the associative representation of SaffronMemoryBase in “Your Brain is Cognitive (Not a Database)”. To further this understanding,, we describe more about Saffron’s memory-based reasoning and its mathematics. Dr. Paul Hofmann, Saffron’s CTO, will describe a “love affair” between associative memories and Kolmogorov complexity at the IBM Research Colloquium: “The Cognitive Enterprise”. Paul’s talk, “Reasoning by Similarity on Top of an Associative Memory Fabric”, provides mathematical substance to the definition of Cognitive Computing by exploiting Kolmogorov complexity. This approach has been called a Universal Cognitive Distance and is the basis for a Universal Artificial Intelligence. If Hofstadter and Sander are right (we think they are), then such a distance provides the computational solution to similarity-based, distance-based, cognitive reasoning.
The meaning of “universal” is important. It defines an approach that is non-assumptive, non-parametric, and non-functional, apposed to “fitting” a particular model by tuning it to a particular set of data. The later, traditional approach in statistics is called “data modeling”. There is an alternative approach, “algorithmic modeling”, which treats the data mechanism as unknown; it learns incrementally from the data. The Cognitive Distance is called universal because it measures similarity between ideas without the need to specify parameters.
As Paul will describe, associative memories are universal too. Rather than fitting a model to data, memories instantly remember everything. Inference using similarity is delayed until query time, thus separating the storage of the observed knowledge from making meaning of it. The combination of representing associative memories and reasoning by Cognitive Distance offers a solution to the “one algorithm hypothesis”, suggested by Jeff Hawkins of Numenta/Grok and Andrew Ng of Stanford/Google, who will also be speaking at the IBM colloquium. Saffron delivers a singular approach that is universal in its application to any domain.
Application scenarios are numerous, covering both anticipatory sense making as well as predictive decision making. As we work with our customers to further develop these applications, they are particularly excited about convergence and narrative. Cognitive Distance tells us how close one thing is to another. Convergence tells us how these things are moving closer together (or farther apart) over time. Narrative tells the story of specific actors and events converging and diverging over time. In national security intelligence, we can ask and answer, “How close is a person moving toward a weapon or location — or persons to each other in information space?” In supply chain risk intelligence, “How close are suppliers getting to competitors — or competitors to each other?” In consumer intelligence, “How close are customers moving toward channels and products — or influential consumers to each other?” Tracking entity trajectories and momentums is the definition of anticipatory sense making.
Cognitive Distance is a true distance in the mathematical definition, to describe the similarity of objects in information space and to find patterns in the data. Paul will mention the close relationship between Kolmogorov complexity and Shannon entropy, inspired by the physics of entropy applied to information, and how we have used entropy for model-free classification in a number of applications. Associative memories also have a long history with physics, going back to John Hopfield and his Hopfield Network, inspired by a class of models in physics (Ising spins) as a model of instant learning and similarity-based recall in human memory. A quantum physicist by training, Paul will describe this “match made in heaven” between an associative memory representation and similarity-based reasoning.
The IBM colloquium is a summit of Cognitive Computing interests. Saffron’s contributions include the mathematical approach to “cognitive” based on Kolmogorov complexity. As another contribution to high performance cognitive computing, SaffronMemoryBase is geared to the computation of many, many distances at Big Data scale. As the ultimate contribution, Saffron has long applied and proved the cognitive approach to enterprise customers from national security to global manufacturers and foundations with more in process.
