**Hanti Lin**

Department of Philosophy

University of California at Davis

**ika[AT]ucdavis.edu**

I am a philosopher of science and formal epistemologist at UC Davis. I did my postdoc at the Australian National University and my PhD at Carnegie Mellon University. I have spent the past two or three years working on a project that aims to justify certain kinds of inductive inferences and to make some progress in our endeavor to reply to Hume's problem of induction. To set the bar very high, my project targets specifically at inductive inferences that are fundamental to the sciences but have hitherto been quite recalcitrant---resisting

**any**justification in**statistics**,**machine learning theory**, or**formal epistemology**. I have at least two examples in mind: (a) enumerative induction to its*full*conclusion, e.g., that all ravens are black, not just that all the ravens you will observe are black; (b) causal inference*without*the (somewhat notorious) faithfulness condition or the like in machine learning. To justify those kinds of inductions, I articulate and defend an epistemological tradition that is very influential in science but often underrecognized and misunderstood in philosophy---the tradition that takes seriously the epistemic ideal**convergence to the truth**, following the footsteps of**C. S. Peirce**,**H. Reichenbach**and**H. Putnam**. See the next section for more details. You can find my**CV**here.**Main Project:**

**Learning-Theoretic Epistemology**

If I am right, the epistemological tradition I just mentioned is best reconstructed as a commitment to the following core guidelines:

(i)Inferential procedures and learning methods for tackling one empirical problem or another should beevaluatedandjustifiedin terms of a certain distinguished group of epistemic ideals, the most important of which include (but are not limited to)convergence to the truthand its various modes.

(ii)An inferential procedure is always evaluated with respect to how good it is for tackling oneempirical problemor another. For there isnosuch thing as a universally best inferential procedure, best for tackling all possible empirical problems.

(iii)Convergence to the truth has many differentmodes. They are epistemic ideals for inferential procedures to achieve where possible; some are higher epistemic ideals than some others are. When tackling an empirical problem, we should look for what modes of convergence to the truthcan be achieved, and try toachieve the best mode we can have.(iv)Taking convergence to the truth seriously doesnotimply ignoring other epistemic ideals. The epistemological tradition in question only says this:

For any empirical problemP, a learning method or inferential procedure counts as one of the best for tacklingPonly ifit satisfies this constraint: achieving the best achievable mode of convergence with respect toP.

Note that this constraint is only claimed to benecessary. Feel free to pursue any other epistemic ideals you value in addition to convergence to the truth.(v)The above constraint, although concerned only with convergence, actually can lead toshort-runepistemic constrains on what to believe, provided that we pursue certain additional epistemic ideals---especially the ideal about stable inquiry that Plato favors inMeno, and the Bayesian ideal about diachronic coherence.

Many people have contributed ideas to one of those guidelines or another. Contributors include, for example:

- philosophersC. S. Peirce,H. ReichenbachandH. Putnam;- machine learning theoristsE. M. Gold,D. Angluin, andL. Valiant;- statisticianR. Fisher, and even Bayesian statisticiansP. DiaconisandD. Freedman.

But none of them takes

*all*the five guidelines seriously. I do. I call the above epistemological tradition**learning-theoretic epistemology**, for it is most recognizable in the many branches of learning theory as studied in theoretical computer science.In this project, I will articulate the core guidelines (i)-(v) clearly, and keep distance from additional theses that are entirely optional but potentially misleading. Once that is done, I will be able to argue that learning-theoretic epistemology deserves philosophers' attention: that it is able to reply to the Keynesian worry that we are all dead in the long, that it is compatible with Bayesianism, that it is neutral between externalism and internalism, and that it can even be welcome by coherentists and evidentialists despite its reliabilist flavor.

This project has already produced something. I have been able to argue that, by taking all the core guidelines (i)-(v) seriously, we can justify certain kinds of inductive inferences that have long resisted justification in formal epistemology, statistics, and machine learning. See the following papers for details:

1. Modes of Convergence to the Truth: Steps toward a Better Epistemology of InductionThis paper aims to justify enumerative induction at itsfullstrength---a task that very few formal epistemologists (if any) have attempted before. Slides are available here.

2. The Hard Problem of Theory Choice: A Case Study on Causal Inference and Its Faithfulness Assumption, forthcoming inPhilosophy of Science.With the same justification strategy as in the preceding paper, this paper aims to justify causal inference without assuming what almost all theorists of causal discovery assume: the famous Causal Faithfulness Condition or the like.

This is a paper in statistics and machine learning theory, providing the theorems that are needed in the preceding, philosophical paper. This is joint work with Jiji Zhang.

For more details about the project, visit the project page.

My older work focuses on the cognitive and conative roles of accepting sentences or propositions, especially the roles that it can or should play in inquiry, decision-making, or linguistic understanding--even for a Bayesian agent. That constitutes the bulk of my publications so far. I am also interested in philosophy of language and logic, especially the topics about compositional, non-truth-conditional semantics which is in line with expressivism. But for now I have to focus a lot more on the main, epistemological project before I can get back to a semantics paper I have presented several times: "When 'Or' Meets 'Might': Toward Acceptability-Conditional Semantics", which is available upon request.

**Publications**

Lin, H. (forthcoming) "The Hard Problem of Theory Choice: A Case Study on Causal Inference and Its Faithfulness Assumption",

*Philosophy of Science*.*The Open Handbook of Formal Epistemology*.

*Res Philosophica*, 94(2): 207- 232.

Lin, H. (2017) “Enumerative Induction and Semi-Uniform Convergence to the Truth”, in Baltag, A., Seligman, J. and Yamada, T. (eds.)

*Logic, Rationality, and Interaction: Proceedings of the 6th International Workshop, LORI 2017*, Lecture Notes in Computer Science, Springer, 362-376.Kevin, K. T., Genin, K. and Lin, H. (2016) “Realism, Rhetoric, and Reliability”, Synthese, 193(4): 1191-1223.

Lin, H. (2016) “Bridging the Logic-Based and Probability-Based Approaches to Artificial Intel- ligence”, in Hung, T.-W. (ed.)

Lin, H. (2016) “The Meaning of Epistemic Modality and the Absence of Truth”, in Yang, C-M., Deng, D.-M., and Lin, H. (eds.)

Lin, H. (2013), "Foundations of Everyday Practical Reasoning",

*Rationality: Constraints and Contexts*, Amsterdam: Elsevier.*Structural Analysis of Non-Classical Logics*, Berlin: Springer-Verlag.Lin, H. (2014) "On the Regress Problem of Deciding How to Decide",

*Synthese*191: 661- 670.Lin, H. and Kelly, K. T. (2013) "Comments on Leitgeb's Stability Theory of Belief",

*Logic Across the University: Foundations and Applications*, Studies in Logic Vol. 47, London: College Publications.*Journal of Philosophical Logic*, 42(6): 831- 862.

Lin, H. and Kelly, K. T. (2012), “A Geo-logical Solution to the Lottery Paradox”, Synthese 186: 531- 575.