Algorithmic bias, as I learned the term, means the finite subset of models that a learning algorithm is capable of learning. For instance, a linear regression algorithm, which predicts (x,y) points given x and a training set, can only construct models of the form y = mx + b after learning the two parameters m and b. There may be other functions that fit the data better, but we cannot search all of them, or even all of those with description lengths smaller than the training data, because Kolmogorov complexity (and therefore AIXI) are not computable. It is impossible to have an unbiased algorithm like AIXI.
Ethical AI means one that won't launch an unfriendly singularity. It could also mean an algorithm that fails to predict human behavior given their ethnic group or gender. We really don't need to worry about the first case for a long time. The second case seems unavoidable, and suppressing research is likely to perpetuate racism and sexism in humans. Shaming researchers and denying that there are differences doesn't make the facts go away. On Mon, Mar 8, 2021, 11:30 AM James Bowery <[email protected]> wrote: > Here's an example of what I mean by AIXI being valuable in practice: > > NVIDIA is now coupling their terms of service with the following > injunction to AI developers that they avoid "algorithmic bias". > > Ethical AI >> NVIDIA’s platforms and application frameworks enable developers to build >> a wide array of AI applications. Consider potential algorithmic bias when >> choosing or creating the models being deployed. Work with the model’s >> developer to ensure that it meets the requirements for the relevant >> industry and use case; that the necessary instruction and documentation are >> provided to understand error rates, confidence intervals, and results; and >> that the model is being used under the conditions and in the manner >> intended. > > > Well well well... just what _exactly_ is "algorithmic bias"? Huh? Who > decides what is "algorithmic bias"? How do we prevent bias in the decision > of what constitutes "algorithmic bias"? > > Ockham's Razor as rigorously formalized by the size prior of AIXI's AIT > subtheory is useful to sweep away the egregious politicization of the > phrase "algorithmic bias". Note that this practical benefit is an > _application_ of the advance in the philosophy of science represented by > algorithmic information theory. > > Until someone comes along and offers as clearly defined (ie: Not "Not even > wrong!") philosophy of science, those wielding influence in the field of > "algorithmic bias" are being UNETHICAL in failing to advance lossless > compression as the most unbiased model selection criterion. > > > On Mon, Mar 8, 2021 at 9:42 AM James Bowery <[email protected]> wrote: > >> While I agree that the value of a theory is to be found in its >> contribution to practice, you are underestimating the value of AIXI in this >> regard. Of course, you are far from alone and there is that old saw >> "safety in numbers." >> >> I simply assumed that the premiere AGI theory, which AIXI _is_, would be >> sufficiently familiar to the denizens of an "AGI" group to sustain the use >> of acronyms for its two sub-theories. >> >> On Mon, Mar 8, 2021 at 5:33 AM John Rose <[email protected]> wrote: >> >>> So with SDT you were alluding to AIXI Sequential Decision Theory.... >>> sorry suffering from acronym overload here, need to rewind the Turing tape. >>> >> *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + delivery > options <https://agi.topicbox.com/groups/agi/subscription> Permalink > <https://agi.topicbox.com/groups/agi/T5bda8b8be25887f8-Md25f5ab14e9a7594fffa2359> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T5bda8b8be25887f8-Md55a4e50bd98aa2bf2ba1aed Delivery options: https://agi.topicbox.com/groups/agi/subscription
