News for hackers. Stuff that matters.
Re:The Computational Theory of Mind (2015) by YeGoblynQueenne on Thursday December 19, 2019 @09:26AM (#21835139)
>> The first argument emphasizes learning (Bechtel and Abrahamsen 2002: 51). A vast range of cognitive phenomena involve learning from experience. Many connectionist models are explicitly designed to model learning, through backpropagation or some other algorithm that modifies the weights between nodes. By contrast, connectionists often complain that there are no good classical models of learning [4.2 Arguments for connectionism].

There is no special need for a specific model of "learning" in a classical setting. Given an inference procedure, such as induction, adbuction or deduction, that can derive new facts and rules in a logical language from observations and a pre-existing theory (i.e. a pre-existing set of facts and rules), all it takes to "learn" is to store the newly derived facts and rules to a database.

I mean "learning" in the sense of Mitchell's definition of _machine_ learning, as (informally) the ability of a system to improve its performance from experience. In this sense, a system that starts with a database of logical facts and rules and adds new facts and rules derived from new observations is "learning".

You can find many examples of learning in a classical, logic setting in the early ('70s and '80s) machine learning literature, particularly with propositional logic learners such as decision list and decision tree learners, the most famous of which are J. Ross Quinlan's ID3 and C4.5 decision tree learners. The field of Inductive Logic Programming studies learning in First-Order Logic languages, especially logic programming languages such as Prolog and Answer Set Programming, and includes early systems such as Shapiro's Model Inference System, Quinlan's FOIL (First-Order Inductive Learner, essentialy a relational version of ID3), Muggleton's Progol and Srinivasan's Aleph (based on inverse entailment), and more recently ASP learners such as ASPAL (Mark Law), or Statistical Relational Learning techniques, e.g. by De Raedt, Kerstig, Getoor, Taschar and others; etc etc.

Bottom line- there is a huge body of work on learning in a classical, logic setting. There is no serious objection that "there _are_ good classical models of learning". Such models are all over the place in machine learning. In fact, they tend to be the most carefully characterised models of machine learning.

There are no comments on this comment yet.