Jet flavor classification in high-energy physics with deep neural networks

Classification of jets as originating from light-flavor or heavy-flavor quarks is an important task for inferring the nature of particles produced in high-energy collisions. The large and variable dimensionality of the data provided by the tracking detectors makes this task difficult. The current state-of-the-art tools require expert data reduction to convert the data into a fixed low-dimensional form that can be effectively managed by shallow classifiers. We study the application of deep networks to this task, attempting classification at several levels of data, starting from a raw list of tracks. We find that the highest-level lowest-dimensionality expert information sacrifices information needed for classification, that the performance of current state-of-the-art taggers can be matched or slightly exceeded by deep-network-based taggers using only track and vertex information, that classification using only lowest-level highest-dimensionality tracking information remains a difficult task for deep networks, and that adding lower-level track and vertex information to the classifiers provides a significant boost in performance compared to the state of the art.

Figure 1

Distributions in simulated samples of high-level jet flavor variables widely used to discriminate between jets from light-flavor and heavy-flavor quarks.

Figure 2

Top: Distribution of the number of tracks associated to a jet in simulated samples. Bottom: Distribution of the number of vertices associated to a jet in simulated samples, before and after removing tracks which exceed the maximum allowed value of 15.

Figure 3

Feedforward neural network architecture. In the first layer, connections of the same color represent the same value of the shared weight. The others layers are fully connected without shared weights.

Figure 4

Architecture of the long short term memory networks as described in the text.

Figure 5

Architecture of the outer recursive networks as described in the text.

Figure 6

Signal efficiency versus background rejection (inverse of efficiency) for deep networks trained on track-level, vertex-level or expert-level features. The top panel shows the performance for $b$-quarks versus light-flavor quarks, the bottom panel for $b$-quarks versus $c$-quarks.

Figure 7

Signal efficiency versus minimum jet $p_{\mathrm{T}}$ relative to light quarks (top) or charm quarks (bottom). In each case, efficiency is shown for fixed values of background rejection for networks trained with only expert features or networks trained with all features (tracks, vertices and expert features).

Figure 8

Signal efficiency versus minimum jet pseudorapidity relative to light quarks (top) or charm quarks (bottom). In each case, efficiency is shown for fixed values of background rejection for networks trained with only expert features or networks trained with all features (tracks, vertices and expert features).

Figure 9

Rejection of light quarks (top) or charm quarks (bottom) versus minimum jet $p_{\mathrm{T}}$. In each case, rejection is shown for fixed values of signal efficiency for networks trained with only expert features or networks trained with all features (tracks, vertices and expert features).

Figure 10

Rejection of light quarks (top) or charm quarks (bottom) versus minimum jet pseudorapidity. In each case, rejection is shown for fixed values of signal efficiency for networks trained with only expert features or networks trained with all features (tracks, vertices and expert features).

Figure 11

Distributions of expert-level features for heavy-flavor and light-flavor classes. Also shown are distributions of light-flavor and charm jets surviving network threshold selections chosen to given rejection of 10 and 50, for networks using only expert information and networks using expert information in addition to lower-level information.