Intelligent machines using machine learning algorithms are ubiquitous, ranging from simple data analysis and pattern recognition tools to complex systems that achieve superhuman performance on various tasks. Ensuring that they do not exhibit undesirable behavior—that they do not, for example, cause harm to humans—is therefore a pressing problem. We propose a general and flexible framework for designing machine learning algorithms. This framework simplifies the problem of specifying and regulating undesirable behavior. To show the viability of this framework, we used it to create machine learning algorithms that precluded the dangerous behavior caused by standard machine learning algorithms in our experiments. Our framework for designing machine learning algorithms simplifies the safe and responsible application of machine learning. [...] Let D, called the data, be the input to the ML algorithm. For example, in the classification setting, D is not a single labeled training example but rather all of the available labeled training examples. D is a random variable and the source of randomness in our subsequent statements regarding probability. An ML algorithm is a function a, where a(D) is the solution output by the algorithm when trained on data D. Let Θ be the set of all possible solutions that an ML algorithm could output. Our framework mathematically defines what an algorithm should do in a way that allows the user to directly place probabilistic constraints on the solution, a(D), returned by the algorithm. This differs from the standard ML approach wherein the user can only indirectly constrain a(D) by restricting or modifying the feasible set Θ or objective function f. Concretely, algorithms constructed using our framework are designed to satisfy constraints of the form Pr(g(a(D)) ≤ 0) ≥ 1 – δ, where g: Θ → ℝ defines a measure of undesirable behavior (as illustrated later by example) and δ ∈ [0, 1] limits the admissible probability of undesirable behavior. Using our framework for designing ML algorithms involves three steps: 1) Define the goal for the algorithm design process. [...] Note that this is in contrast to the standard ML approach: In the standard ML approach, Eq. 1 defines the goal of the algorithm, which is to produce a solution with a given set of properties, whereas in our framework, Eq. 2 defines the goal of the designer, which is to produce an algorithm with a given set of properties. [...] 2) Define the interface that the user will use. The user should have the freedom to specify one or more gi that capture the user’s own definition of undesirable behavior.[...] 3) Create the algorithm. [...] In practice, designers rarely produce algorithms that cannot be improved upon, which implies that they may only find approximate solutions to Eq. 2. Our framework allows for this by requiring a to satisfy only the probabilistic constraints while attempting to optimize f; we call such algorithms Seldonian. [...] The Seldonian algorithms and applications we present below are illustrations to show that it is possible and tractable to design Seldonian algorithms that can tackle important problems of interest. [...] We must therefore provide the user with a way to tell our algorithm the statistic to be bounded, without requiring the user to provide the value, g(θ), of the statistic for different solutions θ (see step 2 above). To achieve this (14), we allow the user to specify a sample statistic g^(θ,D), and we define g(θ) to be the expected value of this sample statistic: g(θ) = E[g^(θ,D)], where E denotes expected value. Continua su https://science.sciencemag.org/content/366/6468/999 ___ Alcune osservazioni da profano (sperando di essere corretto ed imparare qualcosa di nuovo :-D) La fairness del modello selezionato è - probabile, per cui questo framework non può essere considerato una garanzia - definita sulla base di una o più funzioni (gi) definite ex-ante (la cui definizione è sotto il controllo del designer) - tali funzioni sono a loro volta statistiche (per esempio "l'errore medio nelle predizioni per uomini è entro un ε = 0.05 da quello delle predizioni per le donne"). Se ho capito, l'ottimizzazione Seldoniana
argmax a ∈ Af(a)s.t. ∀i∈{1,...,n},Pr(gi(a(D))≤0)≥1−δi
seleziona gli algoritmi che minimizzano la probabilità di ciascuno dei vincoli desiderati in isolamento (entro un certo errore), ma non è detto che minimizzino le combinazioni degli stessi (ad esempio potrebbero minimizzare il bias MEDIO sul sesso e il bias MEDIO sull'etnia o il bias MEDIO sull'età, ma svantaggiare o avvantaggiare comunque le donne anziane di etnia rom). La questione è rilevante perché l'articolo prevede di lasciare all'utente la scelta delle definizioni di "fairness" rilevanti FRA quelle fornite dal progettista. Di conseguenza questo interessantissimo approccio (che sostanzialmente rende probabili alcuni meta-caratteristiche "secondarie" durante la selezione della funzione che approssima l'obiettivo ignoto) non può essere usato per produrre decisioni sull'uomo in quanto comunque imperscrutabile. Giacomo