Yuval Abrams
I work on philosophy of science, philosophy of law, and epistemology, and their relations to metaphysics and moral and political philosophy. I am also very interested in rational choice.
Ph.D. CUNY Graduate Center
LL.M. NYU
LL.B./B.A. Hebrew University of Jerusalem
Supervisors: David Papineau
Ph.D. CUNY Graduate Center
LL.M. NYU
LL.B./B.A. Hebrew University of Jerusalem
Supervisors: David Papineau
less
Related Authors
InterestsView All (72)
Uploads
Papers by Yuval Abrams
Closer look at the various justifications for Probabilism (Dutch Books, Representation Theorems, Convergence, Joyce’s argument for Accuracy), the claim will be made that they only apply in empirical domains where confirmation is possible. Non empirical claims cannot be Dutch-Booked, do not converge, etc. It follows, I argue, that not only are non-empirical hypotheses not confirmable on Bayesian grounds, but that there is something mistaken in viewing our credences in them as credences in the relevant sense. Simply put: there are no arguments for such ‘credences’ to be probabilities, since the arguments for probabilism fail in respect to them. An agent whose ‘credences’ in such hypotheses failed to obey the probability calculus is not thereby irrational .
The argument will be established by demonstrating that the adequacy conditions that White, in his (2006) criticism of Dogmatism, suggests a theory of justification must meet are irrelevant to deciding between empirically equivalent hypotheses. Bayesian confirmation is ineffective in attempting to justify skeptical or anti-skeptical hypotheses and that the assignment of distinct credences to hypotheses that are metaphysically distinct but empirically equivalent is unwarranted, regardless of which justification the Bayesian endorses for assigning credences in general. Credences should be assigned to the entire family of empirically equivalent hypotheses without a specific assignment to the individual hypotheses. Confirmation, therefore, is only relevant to the entire family of hypotheses. Skeptical and anti-skeptical hypotheses with the same empirical content, cannot be decided between using Bayesian machinery: if we are to justify one as against the other, that justification must be independent of Bayesian criteria.
Talks by Yuval Abrams
threshold, if any, that constitutes BARD-proof. We reconstruct BARD-proof such that it admits of the desired mathematical precision, but we do not explicate BARD-proof as mere shorthand for some sufficient threshold probability (or credence). Two major insights are offered to cash this out. The first tells us when a belief state (whether a credence or full belief) is reasonable. The second describes the reasonable dynamic evolution of belief states during a trial. For convenience, we speak as Bayesians (while remaining agnostic between its subjective and objective varieties.)
On the first, ruling out relevant alternatives as unreasonable requires attention to higher-order evidence about the belief-formation process that the trier-of-fact used to arrive at a guilty verdict. The probability of guilt is sufficient only if (i) the trier-of-facts have higher-order evidence with respect to first-order evidence adduced at trial such that their beliefs caused by that first-order evidence are not themselves believed to be caused by flawed or otherwise unreasonable cognitive processes; and (ii) that belief-formation process is safe, in the sense that the agent could not have easily been wrong in similar situations (Williamson 2000: 147).
The second key insight is predicated on an understanding that a simple model of conditionalization starting from a prior probability function is insufficient to capture reasonable doubt, since absent cases of deductive certainty, there will always be a set of priors, starting from which, and conditional upon evidence adduced at trial, doubt regarding guilt would be warranted and, therefore, prima facie reasonable. Accordingly, the BARD test should trace back to all possible priors and ask: of all the possible starting points (priors) that lead to doubt, is there at least one of them that is reasonable?
What is reasonable is captured by the first insight we offer. Unreasonable doubt is doubt that even if it is fully rational (such that it obeys Kolmogorov’s axioms and updates without error on the evidence
via conditionalization) is nonetheless unreasonable because its initial credence assignment was unreasonable.
We take it as plain beyond peradventure that a probability threshold for BARD-guilt is at most a necessary, not sufficient, condition for conviction. One could discover what it is and still not know whether to acquit or convict in a trial. By contrast, our two insights offer a sufficient condition for BARD-proof.
Closer look at the various justifications for Probabilism (Dutch Books, Representation Theorems, Convergence, Joyce’s argument for Accuracy), the claim will be made that they only apply in empirical domains where confirmation is possible. Non empirical claims cannot be Dutch-Booked, do not converge, etc. It follows, I argue, that not only are non-empirical hypotheses not confirmable on Bayesian grounds, but that there is something mistaken in viewing our credences in them as credences in the relevant sense. Simply put: there are no arguments for such ‘credences’ to be probabilities, since the arguments for probabilism fail in respect to them. An agent whose ‘credences’ in such hypotheses failed to obey the probability calculus is not thereby irrational .
The argument will be established by demonstrating that the adequacy conditions that White, in his (2006) criticism of Dogmatism, suggests a theory of justification must meet are irrelevant to deciding between empirically equivalent hypotheses. Bayesian confirmation is ineffective in attempting to justify skeptical or anti-skeptical hypotheses and that the assignment of distinct credences to hypotheses that are metaphysically distinct but empirically equivalent is unwarranted, regardless of which justification the Bayesian endorses for assigning credences in general. Credences should be assigned to the entire family of empirically equivalent hypotheses without a specific assignment to the individual hypotheses. Confirmation, therefore, is only relevant to the entire family of hypotheses. Skeptical and anti-skeptical hypotheses with the same empirical content, cannot be decided between using Bayesian machinery: if we are to justify one as against the other, that justification must be independent of Bayesian criteria.
threshold, if any, that constitutes BARD-proof. We reconstruct BARD-proof such that it admits of the desired mathematical precision, but we do not explicate BARD-proof as mere shorthand for some sufficient threshold probability (or credence). Two major insights are offered to cash this out. The first tells us when a belief state (whether a credence or full belief) is reasonable. The second describes the reasonable dynamic evolution of belief states during a trial. For convenience, we speak as Bayesians (while remaining agnostic between its subjective and objective varieties.)
On the first, ruling out relevant alternatives as unreasonable requires attention to higher-order evidence about the belief-formation process that the trier-of-fact used to arrive at a guilty verdict. The probability of guilt is sufficient only if (i) the trier-of-facts have higher-order evidence with respect to first-order evidence adduced at trial such that their beliefs caused by that first-order evidence are not themselves believed to be caused by flawed or otherwise unreasonable cognitive processes; and (ii) that belief-formation process is safe, in the sense that the agent could not have easily been wrong in similar situations (Williamson 2000: 147).
The second key insight is predicated on an understanding that a simple model of conditionalization starting from a prior probability function is insufficient to capture reasonable doubt, since absent cases of deductive certainty, there will always be a set of priors, starting from which, and conditional upon evidence adduced at trial, doubt regarding guilt would be warranted and, therefore, prima facie reasonable. Accordingly, the BARD test should trace back to all possible priors and ask: of all the possible starting points (priors) that lead to doubt, is there at least one of them that is reasonable?
What is reasonable is captured by the first insight we offer. Unreasonable doubt is doubt that even if it is fully rational (such that it obeys Kolmogorov’s axioms and updates without error on the evidence
via conditionalization) is nonetheless unreasonable because its initial credence assignment was unreasonable.
We take it as plain beyond peradventure that a probability threshold for BARD-guilt is at most a necessary, not sufficient, condition for conviction. One could discover what it is and still not know whether to acquit or convict in a trial. By contrast, our two insights offer a sufficient condition for BARD-proof.