Survivorship bias is a mathematical concept but one that has important real-life implications. My favorite example is a well-known incident from World War II. Many U.S. bombers were being shot down and so there was a study to analyze the planes to determine where to put additional armor.
The study examined bombers that had damage and the recommendation was to put additional armor where most of the damage occurred.
Luckily mathematician Abraham Wald, who was a member of the Statistical Research Group (SRG) at Columbia University, observed that this was the exact wrong thing to do. By examining damage to planes that survived, you only learn where planes can successfully absorb damage. Planes that had been shot down were not available to examine.
So, the surviving planes gave incorrect information with regards to the question at hand.
Here’s another example. One of the endlessly-spewed ridiculous notions is that companies that have women in leadership positions have better performance than those without. The problem is that the data only looks at surviving companies. It’s entirely possible (and quite likely) that companies with women in leadership positions fail at a much higher rate than those without, again leading to survivorship bias. (And that’s not even considering the correlation-causation issue).
Another example concerns successful people in competitive fields. There are many stories of how successful businessmen achieved their success by virtue of hard work and grit. While it’s undoubtedly true that most successful people did work very hard, there’s no record of the countless men who worked hard but failed — their data is lost. In this case survivorship bias leads to being overly optimistic about the relative contributions of work vs. luck.
The implication with regards to machine learning is that you should carefully analyze your training data to make sure that there’s no survivorship issue.
Some recent startup companies that will not survive. A pen-pencil-laser combination. A digital assistant for your plant. A skanky dating app. A phone case that makes coffee.
.NET Test Automation Recipes
Software Testing
SciPy Programming Succinctly
Keras Succinctly
R Programming
2026 Visual Studio Live
2025 Summer MLADS Conference
2026 DevIntersection Conference
2025 Machine Learning Week
2025 Ai4 Conference
2026 G2E Conference
2026 iSC West Conference
Your truth is so massiv, thats why I am here. The other perspective that you take impressed me more and more. And if I follow your words, the picture from the plane right gets a very different meaning.
The bible tolds a story about the woman, but the political correctness never heard of? So I am with you and try to think carefully about some cases.
To get the turn to ML, The hardest part on the repair-connect contructor was to make sure, that the damaged and cracked connections gives the same result like a common NN, if they act like exactly this NN. That was a nightmare, because the “correct” implementation was not so good as a “wrong” implementation, so on the point as the result was syncronised, the feeling was bad for me, so a backup was needed. 🙂
Anyway, the first try was not clearly better or not, so there is no clear answer at this time whats the benefiit is of the idea. On the other side, there was some few setups where the common NN failed and the repair-connect was getting a solid result.
Take an other technic, output support. My first experience with NNs was a youtube tutorial with java, a multiclass perceptron on MNIST, the result was amazing for me, over 90 percent accuracy!
And on every new episode the youtuber was adding more features, a deep NN, then with all common activations, momentum so on, just all the basic stuff. But the results with all these techniques was no so good, the cost to reach the result of a multiclass perceptron was very expensive.
So I just add a multiclass perceptron to my DNN and the output supported NN converges much faster like a linear network, but its a non linear NN with a linear support, so what is it? is it still a non linear NN, or not, or both?
So I wish I had time for a hundred lives to work on ML.
Excellent. The idea of not adding to armor to the bomber aircraft looks looks so counter-intuitive at first. But, when we think a little deeply, the truth presents itself.