SciPost Submission Page
ML-Based Top Taggers: Performance, Uncertainty and Impact of Tower & Tracker Data Integration
by Rameswar Sahu, Kirtiman Ghosh
Submission summary
| Authors (as registered SciPost users): | Rameswar Sahu |
| Submission information | |
|---|---|
| Preprint Link: | scipost_202403_00033v1 (pdf) |
| Date submitted: | 2024-03-24 07:35 |
| Submitted by: | Sahu, Rameswar |
| Submitted to: | SciPost Physics |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approach: | Phenomenological |
Abstract
Machine learning algorithms have the capacity to discern intricate features directly from raw data. We demonstrated the performance of top taggers built upon three machine learning architectures: a BDT that uses jet-level variables (high-level features, HLF) as input, while a CNN trained on the jet image, and a GNN trained on the particle cloud representation of a jet utilizing the 4-momentum (low-level features, LLF) of the jet constituents as input. We found significant performance enhancement for all three classes of classifiers when trained on combined data from calorimeter towers and tracker detectors. The high resolution of the tracking data not only improved the classifier performance in the high transverse momentum region, but the information about the distribution and composition of charged and neutral constituents of the fat jets and subjets helped identify the quark/gluon origin of sub-jets and hence enhances top tagging efficiency. The LLF-based classifiers, such as CNN and GNN, exhibit significantly better performance when compared to HLF-based classifiers like BDT, especially in the high transverse momentum region. Nevertheless, the LLF-based classifiers trained on constituents' 4-momentum data exhibit substantial dependency on the jet modeling within Monte Carlo generators. The composite classifiers, formed by stacking a BDT on top of a GNN/CNN, not only enhance the performance of LLF-based classifiers but also mitigate the uncertainties stemming from the showering and hadronization model of the event generator. We have conducted a comprehensive study on the influence of the fat jet's reconstruction and labeling procedure on the efficiency of the classifiers. We have shown the variation of the classifier's performance with the transverse momentum of the fat jet.
Current status:
Reports on this Submission
Strengths
1. The authors demonstrate that enhancing high-level BDT taggers with the classification scores of GNN/CNN classifiers trained on low-level features improves the tagging performance while mitigating systematic uncertainties from MC generators.
2. They show that adding HL tracking information to the taggers also improves performance.
3. The authors give thorough explanations and lots of details on the procedures they used.
Weaknesses
1. The way the manuscript is structured makes it difficult to read.
2. The authors do not discuss how the MC generator uncertainties present in coming from the low-level features propagate into the composite BDTs.
3. No uncertainties for the ROC curves are provided.
Report
Overall, the work in this paper is interesting and novel. However, the paper is very difficult to read. Information that should go together is sometimes scattered in different places, and in some parts, information is repeated, resulting in an overload that is hard to follow.
I suggest the authors carefully restructure the paper to increase clarity. I believe the content of the paper merits publication, but not in its current format. However, if the authors address this issue and also answer my questions below, I would be happy to recommend it for publication.
1. When discussing the jet images for the CNN training, in the example you rescale the images with 16x16 (arising from the resolution of the detector) up 64x64 pixels. Can you elaborate here? Why not use a CNN for the 16x16 image?
2. For the GNN / CNN training, is there a particular reason why you train the model for 32 epochs with such small batch-sizes (16 and 32)? Why not train for more epochs with larger batch-size? For the CNN a small batch-size can also degrade the results when using BatchNorm layers.
3. How exactly is the b-tagging performed for the BDT? Notice that the b-tagging efficiency should degrade substantially for higher jet pT's (above 200 GeV) when using vanilla b-taggers. Are you including this in your b-tagging?
4. None of the ROC curves have uncertainties. Have you checked if your results change substantially when training the models multiple times with different random seeds?
5. All of the composite classifiers are simple BDTs that use a CNN or GNN score as one of its HLFs. Wouldn't the reduction in MC generator uncertainty be somehow related to the BDT rank of the score feature? Is there any simple way of quantifying this reduction?
6. Any reason why you did not stack CNN + GNN + BDT?
Requested changes
1- Shorten introduction substantially by removing information that is not directly relevant for this work. The conclusion should be used as a template.
2- In the Dataset section I would recommend putting the details of the MC simulations into an appendix.
3 - There is no need to provide details on LorentzNet in the Model section. It is enough to cite the original reference. Same with the CNN sub-section.
4 - Move validation sub-sections in sec IV add to an appendix
5 - "Centralize" all results in a single section (currently results are scattered in two sections), and try to just keep the more relevant results in the bulk and the rest in an appendix
Recommendation
Ask for major revision
Strengths
1 - Thorough study of the role of high and low level features in ML approaches to top-tagging.
2 - Exploration of composite classifiers as a way to include both high and low level features in classification.
3 - Nice observation on the effect on systemic uncertainties when using the composite classifiers on top of deep neural nets.
Weaknesses
1 - The discussion of both the setup and the results is much too long, there is a lot of repetition of points.
2 - The most significant result appears to be the simultaneous reduction in systematic error and improvement in tagging performance that comes from using composite classifiers, but this is not studied in so much detail. The authors say that this further work will be done in the future.
Report
The paper studies the use of high and low level features in top-tagging at the LHC. In doing so, the authors address an important problem in the field, that systematic uncertainties grow as the machine-learning methods rely on lower level features from MC generators. They show that using a composite classifier combining a deep neural network and a BDT to incorporate both the high and low level features, that it may be possible to both increase the performance of the tagger and to reduce the systematic uncertainty arising from the MC generator used to create the training samples. Unfortunately, this point isn't studied enough to make a strong claim here. A more detailed analysis, as in 2212.10493, would be valuable, and I look forward to seeing this in the authors future work. The authors also have interesting findings on a comparison between the classifier efficiency, top-tagging efficiency, and the truth-level identification.
With the requested changes below, I would recommend this paper for publication in SciPost Core.
Requested changes
1 - The introduction should be shortened to include the main points only.
2 - The results sections can also be shortened/combined to avoid too much repetition of the results and discussion.
3 - Can you comment on the importance of the b-tag information when comparing the performance of the BDT, CNN/GNN, and composite classifiers?
Recommendation
Accept in alternative Journal (see Report)
Strengths
1. Detailed explanations included for procedures implemented in this work.
2. Comprehensive study of different types of classifiers.
3. Nice summaries of main results in tables.
Weaknesses
1. Important definitions are sometimes buried in the text.
2. Key points are often lost in lengthy discussions.
Report
This is an interesting and rigorous study which compares the performance of different classifiers in the task of top tagging. The authors include convincing explanations to motivate the various aspects of this study (e.g. using composite models). The use of composite models which use both HLFs and LLFs is particularly interesting in how they are robust against differences in Monte Carlo simulations. However, the level of detail included in this work often leaves many important definitions and key points buried in the text. This made it a challenge for me to follow along without having to constantly review past sections. Nevertheless, this study made several important points and should be published after a significant revision is done to improve the clarity of the work. Also, the conclusion is well-written and captured all the main points of the study.
Requested changes
1. The introduction is quite lengthy and repetitive (e.g. repetition of the different architectures). Please make the introduction more concise and highlight the main points of the paper. Using bullet points to emphasize the main points and then including related details might be a good way to achieve better clarity. (Something similar to what is done in the conclusions)
2. It would be helpful for important definitions (e.g. the various kinds of efficiencies) to be explicitly defined outside the long paragraphs. This makes it easier for the reader to refer to them and ensures that these definitions are not buried in the discussion.
3. p6: In the last paragraph, you mentioned how preprocessing is meant to improve the classifier performance. However, later in Sec IVB you say that the preprocessing for the CNN smears the invariant mass distribution and results in a poorer performance. Does this mean that the CNN may be able to achieve a better performance with a different preprocessing choice?
4. p7: Please include descriptions of the preprocessing step in each row in Fig 2.
5. p16: Please enlarge the size of Fig 4.
Typos/formatting:
6. p3: 2nd paragraph line 5 - should be “fat jet” instead of “fatjet”
7. p3: 3nd paragraph - should be “Section II” instead of “section II”
8. p4: 1st paragraph - “ATLAS” instead of “Atlas”
9. p6: 2nd paragraph on right column - “second dataset” instead of “Second dataset”
10. p8: 1st paragraph - “note:” instead of “note;” (colon not semicolon)
11. p10: Footnote 14 - “parton” instead of “paron”
12. p10: Last paragraph, right before footnote 16 in text - “eps^tag_s” instead of “1/eps^tag_s”?
13. p13: 2nd paragraph - “at least” instead of “atleast”
Recommendation
Ask for minor revision
Author: Rameswar Sahu on 2024-05-13 [id 4484]
(in reply to Report 1 on 2024-05-03)
We express our gratitude to the referee for the thorough review of our manuscript. Below, we provide a summary of our responses to the comments made by the referee:
The referee writes: "The introduction is quite lengthy and repetitive (e.g. repetition of the different architectures). Please make the introduction more concise and highlight the main points of the paper. Using bullet points to emphasize the main points and then including related details might be a good way to achieve better clarity. (Something similar to what is done in the conclusions)"
Our response: Following the referee's comment, we have made several changes to the introduction section, which will be reflected in a future version of our paper. These changes include: 1. We have changed the formatting of the first paragraph on the right-hand side column of page 2 and listed the classifiers using bullet points. 2. We have removed the third and 10th-12th sentences from the second paragraph on the right-hand side column of page 2. 3. We have removed the third sentence from the first paragraph on the left-hand side column of page 3. 4. We have moved the 5th sentence from the same paragraph to footnotes. 5. Except for the last sentence, we have removed all other texts from the second paragraph on the left-hand side column of page three. We have added the last sentence to the second paragraph.
These changes make the introduction section mode concise without affecting the main idea of our work.
The referee writes: "It would be helpful for important definitions (e.g. the various kinds of efficiencies) to be explicitly defined outside the long paragraphs. This makes it easier for the reader to refer to them and ensures that these definitions are not buried in the discussion. "
Our response: We concur with the referee's recommendation and have added a paragraph in the section "CLASSIFIER PERFORMANCE" listing the definitions of the various efficiencies used in our analysis.
The referee writes: "p6: In the last paragraph, you mentioned how preprocessing is meant to improve the classifier performance. However, later in Sec IVB you say that the preprocessing for the CNN smears the invariant mass distribution and results in a poorer performance. Does this mean that the CNN may be able to achieve a better performance with a different preprocessing choice?"
Our response: The motivation behind image data preprocessing is to remove any irrelevant information from the data. It helps the machine to learn useful information and improves the training efficiency. Take the translation preprocessing step as an example. At the LHC, there is no preferred direction transverse to the beamline, and physics should be invariant to azimuthal rotation in the transverse plane. This means the distribution of a jet image at a certain azimuthal coordinate should not differ from its distribution at a different azimuthal coordinate. This motivates us to translate the jet images to be centered at the origin of the azimuthal coordinate, i.e., phi=0.
Data preprocessing may come at a cost if the preprocessing step requires transformations that do not respect the symmetry of the data or lead to distortion of the information in the original data. An example of the former can be the translation of the image in the rapidity(eta) direction. We know that a translation in eta corresponds to a Lorentz boost along the beam line. If we use a variable like the energy of the calorimeter cells as pixel intensity, then a naive translation in the eta direction will change the invariant mass of the jet. However, we can choose to use variables like the transverse momentum, which are invariant under longitudinal boost, and keep the invariant mass information intact.
An example of preprocessing that can alter/distort the data is the rotation step. The motivation behind the rotation preprocessing step is the observation that the radiations inside a jet are approximately symmetric about the jet axis in the eta-phi plane. The problem arises because the jet images are composed of discretized pixels, and any rotation other than a factor of "pi/2" cannot be exact. To perform rotation by arbitrary angles, we usually follow some interpolation methods, and this step can smear out the invariant mass information of a jet. As far as we know, there is no well-defined solution to this problem, though some literature recommends removing this preprocessing step completely (see, for instance, "arXiv:1701.08784").
Note that it is also an option to not use any preprocessing steps in the analysis. However, this requires a large data sample so that the classifier can effectively learn on its own which information is actually important for the classification task and which is redundant. We do not follow this procedure because of the huge computational cost involved.
As far as top tagging is concerned, most analyses in the literature follow the four-stage preprocessing involving translation, rotation, reflection, and normalization. For our analysis, we followed the above preprocessing choice to set up a framework that is easy to relate to other similar analyses in the literature.
The referee writes: "p7: Please include descriptions of the preprocessing step in each row in Fig 2."
Our response: We have included a brief description of the preprocessing steps corresponding to each row of Fig 2 in the figure caption. The details of these preprocessing steps are in the text in the last paragraph of the subsection titled CNN.
The referee writes: "p16: Please enlarge the size of Fig 4."
Our response: We have made the figure larger in the manuscript.
In points 6 through 13, the referee has pointed out a few typos/formatting errors in the manuscript. We would like to thank the referee for pointing out these typos. We have corrected these errors in the manuscript.
Author: Rameswar Sahu on 2024-05-21 [id 4501]
(in reply to Report 2 on 2024-05-15)We extend our thanks to the referee for their thorough examination of our manuscript. Below, we summarize our responses to the referee's comments:
We concur with the referee's first two points and plan to implement them in a future version of our paper. As per the third point,
The referee writes:
"3 - Can you comment on the importance of the b-tag information when comparing the performance of the BDT, CNN/GNN, and composite classifiers?"
Our response:
B-tagging information plays a significant role in separating top jets from QCD jets. A quantitative estimate of its importance can be inferred from the method-specific ranking of input features for the calorimeter-based BDT classifier shown in Table VIII of our paper. However, with the inclusion of track-based features, the importance of this variable decreases as the tracking information is more important in the discrimination of quark jets (top decay product) from gluon jets (a major component of the QCD background jets) (more details provided in the paper). Similarly, in the case of the composite classifiers, due to the presence of other important features, the importance of the b-tagging information is masked.
As for the CNN and GNN-based classifiers, as they are trained directly with the jet-constituents four-momentum as input, they do not have direct access to the b-tagging information. However, one may argue that since the classifiers already have the information of all constituents, they can use this information to gain knowledge about the b-jet inside the top-jet. But this does not work. As shown in the correlation plots in Figure 11 of our paper, the b-tagging information has little to no correlation with the CNN/GNN score. This indicates that the CNN/GNN score and b-tag variable carry independent information. This behavior of not recognizing some important high-level features of the input data (in this case, the b-tagging information) is a well-known problem with low-level feature-based classifiers like CNN/GNN. As far as we know, three different strategies exist in the literature that can address this problem. One of these methods is the construction of composite classifiers, as demonstrated in our paper. The second strategy is to introduce the high-level features as additional global variables in CNN and GNNs. Finally, the third method is a more recent one and advocates using an attention mechanism that forces the CNN/GNN to learn more about these high-level features (See arXiv:2403.11826 for more on the second and third strategies).