Many of you have noticed, and inquired about differences between the "Yes Z", "No Z" results on the main Time-Machine app "My Predictions" page, and scores that are shown in the Leaderboard. These stats are currently calculated in two different ways, and a third method for the performance charts. I'll explain:
When you complete a trial, you see 2 confidence scores at the bottom of your photo (when you view feedback). The score on the left is called the "intended" score and the score on the right is called "displacement score". Intended score is the confidence score that reflects how close your RV transcript matched the photo that you see. The Displacement score is how close your RV transcript matched all of the other 9 photos in that prediction. "Displacement" refers to how we, at times, lose our "focus" and actually perceive elements of other photos connected to the prediction, but not actually the photo for the "intended" trial number.
You can view your statistics for each intended and displacement in the performance section of Time-Machine. Below is an example of the correct and incorrect (red and green lines) scores for Intended.
Below is an example of the same plot, but for all displacement trials. You can see that in both cases, the green line is above the red line as we want, and in this case, BOTH ending green line values are statistically significant > 1.8.
So what is the best way to decide what your actual prediction is using intended and displacement scores? It turns out that the most predictive method is a mix of both, but the mix percentage is becoming more accurate as we collect more data. It is a algo that is being optimized as time goes on and more data comes in. Because the mix is not exactly known yet, we just make the stats on the main page a 50% mix of both intended and displacement.
However, we kept the leader board at 100% intended.
That's why the two are different. Neither one is wrong, and at times, one is better than the other. To be consistent, we will be changing the leader board to match the main predictions page and make it a 50-50 mix of both intended scores and displacements scores.
コメント