Djelal Osman

Explainable AI Enthusiast | PhD Physics @Cambridge_Uni | Probability Theory, Machine Learning, Numerical Optimisation, Graph Theory, Applied Mathematics

[Djelal Osman @djelal]

@10x_er A good amount of intuition for formal verification overlaps with the concept of why strongly (and semantically) typed code is better than validation, a la @lexi_lambda 's "Parse, don't validate" blogpost: https://t.co/nRO59yePuq.

[Djelal Osman @djelal]

@getjonwithit That's true, I stated the implication too loosely. Decidability doesn’t imply completeness. My point is perhaps better illustrated by the fact that arithmetic truth is outside effective formal procedure. PA is c.e. but incomplete, while true arithmetic is complete but not c.e.

[Djelal Osman @djelal]

@lakens I fail to understand the question, "which comes first, the hypothesis or observation?", with respect to probabilistic inference. Probabilities make no assertion about the causal relationship between any sets of propositions; only about their logical relationships.

[Djelal Osman @djelal]

@rasbt I saw a presentation by Naftali Tishby in 2017/18 where he explained that the process of stochastic gradient descent will, after the point of plateau in error, lead to more efficient representations of the inputs across all layers (the information goes up when continuing SGD).

[Djelal Osman @djelal]

@neuralreckoning Computation is ill-defined, but typically is used to refer to a process that makes decisions w.r.t. some information. Computation likely refers to the process that "generates" a probability distribution (the answer) given information (possible outcomes + constraints + question).

[Djelal Osman @djelal]

@AzizaS Kinetic Inductance Detector?

[Djelal Osman @djelal]

@QiaochuYuan This is fascinating. In regards to the need for distance, direction, rotations/translations as fundamental, what are your views on the modern field of Geometric Algebra (developed by David Hestenes from earlier Clifford Algebra)? Do you feel this might fit the bill to that end?

[Djelal Osman @djelal]

@ProfLaurenRoss In Bayesian inference, one must first define a set of exhaustive propositions. Given these propositions, the job of Bayesian analysis is to give only a logically consistent result - not a result true to the "real world". Is there an analogous philosophy/model in causal inference?

[Djelal Osman @djelal]

@MechncOfMeaning @yudapearl Just to be a bit more rigorous, when I say that only P(rc) and P(r) measured, I mean that only P(rc) is measured, and the scaling factors P(r) or P(c) are both equivalent, which is why P(rc)/P(r) is symmetric given the knowledge of the observer in this example.

[Djelal Osman @djelal]

@MechncOfMeaning @yudapearl In order to make "direct" inference regarding P(r | c), one must observe information pertaining to P(r | NOT(c)). This can only happen if P(rc) and P(r NOT(c)) are observed. Even then, only logical inferences are made, no physical relations between r and c in this case. 3/3

[Djelal Osman @djelal]

@togelius It might be more common that someone finds the current interpretation of the evidence to be implausible / inconsistent (given some other logic), actively seeking evidence to just disprove the existing theory. Adding a better explanation is more difficult, and likely less common.

[Djelal Osman @djelal]

@gro_tsen The real Bitcoin was the friends we made along the way.

[Djelal Osman @djelal]

@lakens One would hope that a person does research because they want to answer a question. What difference does it make if nobody else wanted to know that answer? Even less importantly, if the answer had no utility? Research with predictable utility is half already-answered anyway.

[Djelal Osman @djelal]

@paulg This is due to the way AI research is pursued, and the reason we have had comparatively no fundamental progress in explainable AI. It is the wrong way to do it, the product of seeking sexy results at the expense of real understanding, and a detriment to true long term progress.

[Djelal Osman @djelal]

@divbyzero It came out amazing! How beautiful!

[Djelal Osman @djelal]

@divbyzero (When you do print it, I mean)

[Djelal Osman @djelal]

@divbyzero Please post it! I'm excited to see it!