[HN Gopher] Explainability is not a game
       ___________________________________________________________________
        
       Explainability is not a game
        
       Author : mitchbob
       Score  : 14 points
       Date   : 2024-07-02 21:18 UTC (1 hours ago)
        
 (HTM) web link (cacm.acm.org)
 (TXT) w3m dump (cacm.acm.org)
        
       | brigadier132 wrote:
       | ML models are just statistical function approximators right? So
       | isn't asking for "explainability" from ML models equivalent to
       | asking "why is fire hot?"? If an ML model makes a prediction the
       | reason why is because it was trained on something that made the
       | prediction more likely than others.
        
         | lgas wrote:
         | Sure, but there are different variables in the training data
         | that may influence each inference more or less. So if you can
         | observe which variables influenced the inference and how much
         | then you can (at least attempt to) construct an explanation.
        
           | mvanveen wrote:
           | In this case let's assume that the weights and biases of,
           | say, a neural network are fixed and the model is already
           | trained.
           | 
           | One way of thinking about explainability is that it deals
           | with determining for some input data how much each feature is
           | contributing to the final outcome (e.g. variable 1 and 2
           | contributed x% and y% to the final inferred value).
           | 
           | You're correct to suggest that when you backpropogate
           | residual error there are also non-linear interactions between
           | features and that will affect how much each variable is
           | contributing to the updated weight value (in fact that's kind
           | of the point of a deep network ;).
        
         | taeric wrote:
         | At very abstract points, I think you have a point; but that is
         | largely a failure to scale for any system. Consider holding a
         | rock is not dangerous. Holding a coal is. Why do you think
         | that?
         | 
         | Take that further and consider why people are hesitant to touch
         | cooking surfaces? The answer isn't "because fire is hot." The
         | answer is "because the purpose of cooking surfaces is to get
         | hot enough to cook things and without knowing how long it has
         | been unused, it is not safe to touch."
        
         | Vecr wrote:
         | That's a big problem though, mechanistic
         | interpretability/explainability is really important for the
         | whole "not wrecking everything" part of AI. You could also just
         | hope AI stalls out or goes back to more hand-tuned Bayesian
         | stuff, but hope is not a plan. Progress past standard analysis
         | has been made, so I'm not sure how important this paper is.
        
       | tomxor wrote:
       | > ML models is most often inscrutable, with the consequence that
       | decisions taken by ML models cannot be fathomed by human decision
       | makers
       | 
       | Yup, James Mickens planted this seed 6 years ago, well worth the
       | watch, even if you already agree it's brilliant cathartic humour
       | (the subject isn't really "security").
       | 
       | https://www.youtube.com/watch?v=ajGX7odA87k
        
       | mvanveen wrote:
       | I am a co-author of a patent for model explainability for credit
       | risk underwriting applications using Shapley values.
       | 
       | In fairness I haven't given this article a thorough read but my
       | initial impression is that I'm finding myself frustrated by the
       | FUD this article is attempting to spread. As my boss would often
       | remark to remind us all: _model explainability is an under-
       | constrained optimization problem._ By definition there isn 't a
       | unique explanation decomposition unless you further constrain the
       | problem.
       | 
       | Therefore, I personally find that hand-wringing around there not
       | being 100% agreement around different explanations for a model
       | inference, while definitely thought provoking and worth
       | considering, should at least account for this reality. For some
       | reason a lot of folks in the ML community seem to have come the
       | opinion that because the problem is under-constrained that means
       | that explanations shouldn't be calculated or have no utility.
       | 
       | Would you prefer a model that examines which features are driving
       | the model to deny a disproportionate number of folks of a
       | particular race or ethnicity or not, all things being equal? My
       | point is even if there are limitations to explainability I think
       | there are a lot of very real, critical scenarios where applying
       | SHAP can be of actual, real world utility.
       | 
       | Furthermore, it's not clear that LIME or other explainability
       | methods will provide better or more robust explanations than
       | Shapley values. As someone that has looked at this pretty
       | extensively in credit underwriting I'd personally feel most
       | comfortable computing SHAP values while acknowledging some of the
       | limitations and risks this article calls out.
       | 
       | Axioms such as completeness are also pretty reasonable and I
       | think there is a fair amount of real world utility to
       | explainability algorithms that derive from such an axiomatic
       | basis.
        
       ___________________________________________________________________
       (page generated 2024-07-02 23:00 UTC)