274 lines
4.8 KiB
Plaintext
274 lines
4.8 KiB
Plaintext
<subsection Why is the trivial Model so good?>
|
|
|
|
<frame>
|
|
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>Now lets focus a bit more on the trivial model</e>
|
|
<e>in it, I compare just the angular part to 0 (its mean)</e>
|
|
<e>and as you see on the left side, the distribution for tops is way more complicated (logarithmic color coding!)</e>
|
|
<e>so since comparing to zero=approximating this radius, tops are clearly classifiable using this</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="rmeanangle3"></i>
|
|
</que>
|
|
</split>
|
|
|
|
</frame>
|
|
|
|
<frame>
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>compare this to the distribution in #p_t#</e>
|
|
<e>basically no preference</e>
|
|
<e>even switches depending on the displayed particle</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="meanpt1" f2="meanpt7"></i>
|
|
</que>
|
|
</split>
|
|
</frame>
|
|
<frame>
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>This related to another problem</e>
|
|
<e>if you would train a working autoencoder no longer on qcd data, but on top, it would still consider tops more complicated</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="recinv"></i>
|
|
</que>
|
|
</split>
|
|
</frame>
|
|
<frame>
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>This you can see best in aucmaps</e>
|
|
<e>These show the auc as a function of the particle id and the current feature</e>
|
|
<e>blue color = qcd data is simpler</e>
|
|
<e>red color = top data is simpler</e>
|
|
<e>white color = no preference</e>
|
|
<e>a perfectly working network would be darkblue if trained on qcd and darkred if trained on top</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="xqcdmap.png" f2="xtopmap.png"></i>
|
|
</que>
|
|
</split>
|
|
</frame>
|
|
<frame>
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>you can subtract those maps</e>
|
|
<e>here more different=more red</e>
|
|
<e>basically no difference in angular data</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="deltamap" wmode=True></i>
|
|
</que>
|
|
</split>
|
|
</frame>
|
|
|
|
<frame>
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>you have the same problem of adding d-distributions as you have in the scaling case</e>
|
|
<e>so you could ask yourself if adding something to the angular data actually helps</e>
|
|
<e>comparing the only angular data to the general data, you see that it in fact hurts the auc (even though just a bit)</e>
|
|
<e>this effectively means, my current network does not use #p_t# at all</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="angularscale"></i>
|
|
</que>
|
|
</split>
|
|
</frame>
|
|
|
|
<frame>
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>But again, this does not mean, that there is no information in pt</e>
|
|
<e>in fact, you see in these aucmaps, that the pt part is actually red where it should be red and blue where it should be blue</e>
|
|
<e>so how about using only #p_t#</e>
|
|
<e>you obviously lose quality</e>
|
|
<e>also training an autoencoder to get an high auc in pt is not yet trivial</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="xqcdmap.png" f2="xtopmap.png"></i>
|
|
</que>
|
|
</split>
|
|
</frame>
|
|
|
|
<frame>
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>multiplicative scaling does not really work</e>
|
|
<e>best network reaches an auc of about #0.78# which is about the same, as QCDorWhat gets for minimally mass decorrelated networks</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="trivialptscale"></i>
|
|
</que>
|
|
</split>
|
|
</frame>
|
|
<frame>
|
|
<split>
|
|
<que>
|
|
Benefits
|
|
<list>
|
|
<e slight improvement in network quality>
|
|
<e allows training to actually be inverted>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
Problems
|
|
<list>
|
|
<e>you basically split your training into a network with a good auc, and one that learns (hopefully) non trivial stuff</e>
|
|
<e less effective loss scaling>
|
|
</list>
|
|
</que>
|
|
</split>
|
|
So maybe you could do the same with some different preprocessing (one that does not just give you trivial information)
|
|
</frame>
|
|
<frame>
|
|
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>Easiest Transformation: no Transformation (4 vectors)</e>
|
|
<e>so</e>
|
|
<l2st>
|
|
<e>Energy</e>
|
|
<e>#p_1#</e>
|
|
<e>#p_2#</e>
|
|
<e>#p_3#</e>
|
|
</l2st>
|
|
<e>trained on qcd, but prefers top!</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="badmap"></i>
|
|
</que>
|
|
</split>
|
|
|
|
</frame>
|
|
<frame>
|
|
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>Why is that so?</e>
|
|
<e>maybe just a bad network</e>
|
|
<e>compare metrics (defining distance in topK)</e>
|
|
<e>basically require the network to learn the meaning of #phi# and #eta# itself</e>
|
|
<e>so without, no concept of locality, meaning no useful graph</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="badmetrik" f2="goodmetrik"></i>
|
|
</que>
|
|
</split>
|
|
|
|
</frame>
|
|
<frame title="How to solve this">
|
|
|
|
<list>
|
|
<e>add Dense Network infront of the TopK</e>
|
|
<l2st>
|
|
<e>better, but still not good</e>
|
|
</l2st>
|
|
<e>run TopK still on preprocessed Data</e>
|
|
<l2st>
|
|
<e>good, but numerical problems</e>
|
|
<l3st>
|
|
<e>require to go to 4 particles and less training data</e>
|
|
</l3st>
|
|
</l2st>
|
|
</list>
|
|
|
|
</frame>
|
|
<frame>
|
|
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e>same good reconstruction in #p_1# and #p_2#</e>
|
|
<e>makes sense, since #Eq(p_t**2,p_1**2+p_2**2)#</e>
|
|
<e>but apparently Energy and #p_3# prefer tops</e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="goodmap"></i>
|
|
</que>
|
|
</split>
|
|
|
|
</frame>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<ignore>
|
|
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e></e>
|
|
<e></e>
|
|
<e></e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
|
|
</que>
|
|
</split>
|
|
|
|
<frame>
|
|
|
|
<split>
|
|
<que>
|
|
<list>
|
|
<e></e>
|
|
<e></e>
|
|
<e></e>
|
|
</list>
|
|
</que>
|
|
<que>
|
|
<i f="none"></i>
|
|
</que>
|
|
</split>
|
|
|
|
</frame>
|
|
|
|
<frame>
|
|
|
|
<list>
|
|
<e></e>
|
|
<e></e>
|
|
<e></e>
|
|
</list>
|
|
|
|
</frame>
|
|
|
|
|
|
</ignore>
|
|
|
|
|
|
|
|
|