Jun Otsuka (Kyoto University, [email protected]) 2024/10/26 From Confirmation 
 to Generation Rethinking Science through Collective Predictive Coding 

Contents: Philosophical Implications of CPC-MS • CPC-MS: From the Con fi rmatory to Generative view of Science • Implications on 1. Scienti fi c Objectivity 2. Scienti fi c Progress 3. Scienti fi c Methodology 

Traditional view of science aka the “received view” of philosophy of science • Focus on the “context of justi fi cation” • Scienti fi c knowledge is characterized not by how it is discovered/conceived • but rather by how it is justi fi ed: - Objective evidence - Logical reasoning (e.g. HD method) - Statistical methods (Bayesian or Frequentist) • The con fi rmatory view of science • Goal: accumulation of con fi rmed hypotheses 

CPC-MS: the overall picture • Science is a collaborative activity in which • each agent makes observation • updates her internal representation • writes a paper as a sample from • reviewer assesses based on his • if accepted, it is incorporated into the global posterior distribution • This whole process represents the Metropolis-Hasting algorithm for approximating the true posterior k o q(zk |o) w q(w|zk) l w q(w|zl) q(w|o) p(w|o) o q(zk |o) q(w|zk) q(w|zl) q(w|o) p(w|o) ≈ 

1. Implications on scientific objectivity Criticism of the received view (Longino, Kitcher) • The received view • Scienti fi c method is practiced by a single individual. • Objectivity is warranted by individual scientists’ following the method. • Social objectivity (Longino 1990, Kitcher 1993) • “The objectivity of scienti fi c inquiry is a consequence of this inquiry’s being a social, and not an individual, enterprise (Longino, p. 67).” - Intersubjective criticisms (peer review, replications) - Heterogeneity & diversity - Situatedness (di ff erent motivations, agenda, and biases) 

1. Implications on scientific objectivity CPC-MS and social objectivity • CPC-MS as a model of scienti fi c interactions • Paper-writing → Review → Incorporation into the posterior • Convergence without consensus • Convergence of the posterior happens even if - individual scientists disagree each other, i.e., - they have incorrect posterior, • Diversity and openness promote convergence • Ergodicity of Markov Chain → will explore all possibilities in the long run q(w|o) ≈ p(w|o) q(zk |o) ≠ q(zl |o) q(zk |o) ≠ p(zk |o) 

1. Implications on scientific objectivity CPC-MS and social objectivity • CPC-MS as a model of scienti fi c interactions • Paper-writing → Review → Incorporation into the posterior • Convergence without consensus • Convergence of the posterior happens even if - individual scientists disagree each other, i.e., - they have incorrect posterior, • Diversity and openness promote convergence • Ergodicity of Markov Chain → will explore all possibilities in the long run q(w|o) ≈ p(w|o) q(zk |o) ≠ q(zl |o) q(zk |o) ≠ p(zk |o) 

2. Implications on Scientific Progress The problem of scienti fi c progress • Progress = accumulation of (con fi rmed) knowledge? • Di ffi culties 1. Incommensurability between scienti fi c revolutions (Kuhn) • No common standard between old/new paradigm. • How to say that the new paradigm is “better” than the old one? 2. Pessimistic induction • Many scienti fi c theories ended up being refuted. So no accumulation after all? 

2. Implications on Scientific Progress Progress in predictive performance • The generative view: prediction > knowledge • Posterior as means for prediction (posterior predictive distribution) • Bayesian updates minimize the predictive loss • Paradigm shift (Hayashi) • Singular models → discontinuous “jumps” among singular points • Radical changes in representation • But predictive performance gets better (so comparable). p(˜ o|o) = ∑ w p( ˜ o|w)p(w|o) q(z|o) 

3. Implications on Scientific Methodology From “foolproof algorithms” to “communication tools” • Received view: scienti fi c methods as algorithms to truth • Used by each individual scientist • Bring about objective conclusions when used properly • Generative view: scienti fi c methods as communication tools • Material and methods: clarify the assumptions/logic • Statistics: allows peer-reviewers to judge the quality of evidence • → communication platform for scienti fi c exchange 

Conclusions Implications of CPC-MS 1. Scienti fi c Objectivity • CPC-MS as a model of Social Objectivity • Diversity promotes convergence w/o consensus 2. Scienti fi c Progress • Progress in terms of predictive performance, even between di ff erent paradigms 3. Scienti fi c Methodology • Statistics as “communication tools” that facilitate the MHNG