Distilling Neural Networks | Two Minute Papers #218

die foto's colors this is to the papers wit cadeau journée fire zesde latest explosion in een jaar research virtually no field of science remains aantal spanjolet works these are amazing tools that helps of problems waar de solochain zou easy to identify met difficult to explain for instance wie low back flip een missie van het mathematical in de familie of driekwart forces rotations en torque is much more challenging nu al networks xl at the skies of test provided that we can still i'm a large number of training samples of twee piek in site disney euro networks with a more and more layers and more neurons within the layers is hier scopa de final decision dit pension wordt nu rond haar activated by er rimpels de jaar highly efficient hallo ever tried to understand how a decision is be made by the networks is going to be flessen nummer is specially troublesome in de network use as from cancer dat wie i do not have access to any sort of explanation mee euro niazi eksept we dat trappelen consideration this piece of work' is debat distillation which means dat weet ik en jool at work and tried to express iets in een working in the form of a decision tree is take into consideration de series of variables and provide een klier roadmap to the decisions based on their voor een stagiair useful in using the age en de maand of real time of people to try to get further than likely to play video games or this finding hoe ze get de loon van de bank bijzonder each location and inca ja this sounds great hall over de main issue is that decision tree is een natte substituut voor nuon adwords de series 6 dat we hebben generation versus interprata billen tea tree dag situation which means that she is dat providers decisions over vinden training dna en generalized pulley en de warmste waar easy to interpret marina shirt zal in order to break out of this tradle situatie en via ideal of this piece of art is de eerste your network to build a decision tree by taking een input deden set voor training tryin to generate more training day dag de fellows dus een prak peace and 4 all this to de seizoen 3 gerard families als voor de testikel problem of identifying widgets in die hem die stenen set and each seizoen is meant to cut de number of output options in half and shows en wielewaal dat we can vary effectief lid perform the classification and only for decisions van migraines input and not only daad maar dit also shows waar dit is looking for voor een stans hier we gaan zie dat de final decision before weekendploeg worden input numbers week of eet het looks for the presence of a tiny er ja de joins the end of this week to my knees and different visualisatie of de qonect for game beters argues that in your network weekly tries to distinguish uw kapsel strategies one waar de player start plegen van de inner and one of the operation of the board dit is zo'n let the streets perform better than traditional decision tree what's more the only slightly worse than the corresponding euro net worth what can explain their decisions much more clearly en waar ofzo faster in summary the rate of progress in machine learning research is sloerie insane these days and i'm all for piepers dutch writer providers en greater understanding of what is happening kan door de hut i'm loving this idea in particular we hadden earlier episode houten supercharge the decision tree wirefree boosting is er interested in learning more about it de link is available in de video description tag for watching god' folio generous support must see you next time