ILNet: Diverse Architectures in Machine Learning

• ILNet is an acronym used for various machine learning architectures, each tailored to specific tasks such as infrared detection, trajectory prediction, and object localization.
• The infrared ILNet employs a U-shaped encoder–decoder with Interactive Polarized Orthogonal Fusion, Dynamic One-Dimensional Aggregation, and Representative Block to enhance small target detection.
• Other ILNet variants incorporate inverse learning attention for multi-agent forecasting, localization-quality supervision for semi-supervised detection, and interpretable rule-based logic for classification.

ILNet is a reused acronym in recent machine-learning literature rather than a single, unified architecture. In arXiv publications, it denotes at least three distinct named models—an infrared low-level network for salient infrared small target detection, a trajectory prediction architecture with Inverse Learning attention, and an “Improving Localization net” for semi-supervised object detection—while related work also uses ILNet/ILN for inertial localization networks and for an interpretable logical network formulation built from AND/OR/NOT concepts (Li et al., 2023, Zeng et al., 9 Jul 2025, Rossi et al., 2022, Zhang et al., 21 Jul 2025, Perreault et al., 11 Aug 2025).

1. Nomenclature and major uses

The acronym has been expanded differently across subfields, and the resulting models address different data modalities, objectives, and inductive biases. In computer vision, ILNet may refer to a salient infrared small target detector that emphasizes low-level cues, or to a trajectory forecasting model that performs inverse temporal reasoning over multi-agent interactions. In semi-supervised detection, “IL-net” abbreviates “Improving Localization net” and denotes an auxiliary localization-quality classifier integrated into a Mean Teacher pipeline. Related literature further uses ILNet as a general inertial localization network concept, and as a shorthand for an interpretable logical neural network architecture (Li et al., 2023, Zeng et al., 9 Jul 2025, Rossi et al., 2022, Zhang et al., 21 Jul 2025, Perreault et al., 11 Aug 2025).

Name in literature	Domain	Defining components
ILNet	Salient infrared small target detection	IPOF, DODA, RB
ILNet	Multi-agent trajectory prediction	IL Attention, DAS
IL-net	Semi-supervised object detection	IoU classification head, pseudo-label filtering
RepILN	Inertial localization	RepBlock, TSSA, SA-GCU
ILNet as instantiated in Neural Logic Networks	Interpretable classification	AND/OR/NOT concepts, DNF rules

A common source of confusion is that these models do not share a common architecture despite the shared acronym. This suggests that “ILNet” functions primarily as a field-local naming convention whose meaning must be resolved from context.

2. Infrared low-level network for salient infrared small target detection

In "ILNet: Low-level Matters for Salient Infrared Small Target Detection" (Li et al., 2023), ILNet is an infrared low-level network for infrared small target detection (SIRST). The method is motivated by the observation that typical targets are smaller than about $15\times 15$ pixels, are low-contrast due to low signal-to-clutter ratio, and have weak color and texture cues. Because high-level semantics are seldom present, repeated down-sampling and abstraction in deep CNN layers cause target signals to attenuate or vanish in deeper layers; the paper terms this phenomenon deep abatement. The model therefore reframes SIRST as salient object detection of intensity-salient micro-regions rather than semantic segmentation of meaningful objects.

The network adopts a U-shaped encoder–decoder architecture built with ReSidual U-blocks as in $U^2$Net. The encoder extracts multi-stage features $\mathbf E_i$, and the decoder produces multi-stage outputs $\mathbf D_i$. At each stage, an Interactive Polarized Orthogonal Fusion module bidirectionally fuses aligned encoder and decoder features. The fusion is “polarized” because it separates spatial and channel branches, “orthogonal” because those branches operate on complementary axes, and “interactive” because each branch uses both $\mathbf E_i$ and $\mathbf D_i$ to compute attention or gating. The module output is

$$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$

Inside IPOF, ILNet inserts Dynamic One-Dimensional Aggregation layers, designed to adapt the depth of lightweight 1D convolutions to the stage-wise embedding dimension. The empirical mapping is

$$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$

with $n=b=2$ in all experiments. The stated role of DODA is to avoid underfitting high-channel stages and overfitting low-channel stages while keeping aggregation low-cost.

After stage-wise decoding, ILNet uses a Representative Block to aggregate decoder outputs globally and to enhance deep outputs using shallow outputs, which carry more low-level detail. The enhancement factor is

$$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$

and the final segmentation map is

$U^2$0

The paper describes RB as “ensemble learning”-inspired because it dynamically allocates weights for shallow and deep layers.

Training uses deep supervision with binary cross entropy at each decoder stage. No IoU, focal, or explicit orthogonality losses are used; BCE is the only training loss, with weight decay $U^2$1 as regularization and BN/LN inside modules. Images are resized to $U^2$2; training uses Adan with initial learning rate $U^2$3, batch size $U^2$4, AMP, and $U^2$5 epochs in PyTorch 1.10.0 on NVIDIA Jetson AGX Xavier. The paper reports three scales: ILNet-S, ILNet-M, and ILNet-L.

Quantitatively, ILNet-L reports $U^2$6 IoU, $U^2$7 nIoU, $U^2$8 $U^2$9, and $\mathbf E_i$0 $\mathbf E_i$1 on NUAA-SIRST; on IRSTD-1K it reports $\mathbf E_i$2 IoU, $\mathbf E_i$3 nIoU, $\mathbf E_i$4 $\mathbf E_i$5, and $\mathbf E_i$6 $\mathbf E_i$7. The paper states that ILNet-L surpasses ISNet on nIoU and reduces false alarms substantially on both datasets. Ablations show that each of IPOF, DODA, and RB helps, that full-stage IPOF fusion improves target-level metrics with small overhead, and that shallow-to-deep enhancement is decisively better than deep-to-shallow enhancement. ILNet-S is reported at approximately $\mathbf E_i$8M parameters and approximately $\mathbf E_i$9 GFLOPs, which the paper characterizes as suitable for edge deployment. Visual analysis attributes clearer target boundaries, fewer miss detections, and fewer false alarms to the low-level-centric design, and the paper states that gains are more pronounced on the larger IRSTD-1K dataset (Li et al., 2023).

3. ILNet for multi-agent trajectory prediction with Inverse Learning attention

In "ILNet: Trajectory Prediction with Inverse Learning Attention for Enhancing Intention Capture" (Zeng et al., 9 Jul 2025), ILNet is a multi-agent motion forecasting model for interaction scenarios. The method takes $\mathbf D_i$0 interacting agents observed for $\mathbf D_i$1 steps and predicts $\mathbf D_i$2 future modes $\mathbf D_i$3 over $\mathbf D_i$4 steps, together with probabilities $\mathbf D_i$5. The central motivation is that factorized attention over temporal and agent axes is efficient but static and forward-only, so it captures only obvious and immediate behavioral intentions and lacks explicit spatio-temporal coordination across neighboring moments.

The model introduces Inverse Learning Attention and Dynamic Anchor Selection. The encoder first embeds agent motion, shape, category, and relational edge features, together with an HD-map graph. Mode queries interact with the map via multi-head graph cross-attention, and then pass through three attention layers: Ego-Agent Temporal Attention, Agent Future Attention, and Agent History Attention. The future-attention step conditions the current moment on interacting agents at $\mathbf D_i$6 inside the observed history window; the history-attention step then inversely reasons back to $\mathbf D_i$7. The resulting “proposed intention” is fused as

$\mathbf D_i$8

The paper explicitly distinguishes this from inverse reinforcement learning: the “inverse” step concerns temporal information flow within the observation window rather than reward recovery.

After IL Attention, ILNet applies repeated factorized graph self-attention blocks over agents, historical time, and modes. Proposal trajectories are then refined by Dynamic Anchor Selection, which learns trajectory change keypoints per mode from local correlations across history, mode, and future-time axes. Historical and proposal trajectories are embedded, local $\mathbf D_i$9D correlations are extracted with lightweight Conv2D blocks, and an MLP with sigmoid predicts normalized anchor positions. These anchors are used to reparameterize proposal trajectories into anchor-centered polar coordinates for a second refinement stage. The final output is

$\mathbf E_i$0

Training uses winner-takes-all mode selection, Huber regression loss on both proposal and final predictions, and a cross-entropy loss on final mode probabilities. The paper reports $\mathbf E_i$1, $\mathbf E_i$2 heads, $\mathbf E_i$3, cosine learning-rate scheduling, weight decay $\mathbf E_i$4, $\mathbf E_i$5 epochs, and training on $\mathbf E_i$6 RTX 4090. On INTERACTION, the model is evaluated with minJointADE@6 and minJointFDE@6; on Argoverse it reports minADE@6, minFDE@6, Miss Rate@6, and Brier-minFDE@6, together with diversity metrics on validation.

The reported test-set results are state-of-the-art within the paper’s comparison set. On Argoverse, ILNet reports minFDE6 $\mathbf E_i$7, MR6 $\mathbf E_i$8, B-minFDE6 $\mathbf E_i$9, minADE6 $\mathbf D_i$0, and approximately $\mathbf D_i$1M parameters. On INTERACTION, it reports minJointFDE6 $\mathbf D_i$2 and minJointADE6 $\mathbf D_i$3; the paper states that this improves over HPNet by $\mathbf D_i$4 FDE and $\mathbf D_i$5 ADE with fewer parameters in that setup. Ablations on INTERACTION validation show that inverse modeling with FA then HA yields $\mathbf D_i$6 FDE/ADE, compared with $\mathbf D_i$7 for the TA-only baseline, and adding DAS yields $\mathbf D_i$8. DAS adds only $\mathbf D_i$9M parameters on INTERACTION and $$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$0M on Argoverse, with small latency increases. The paper further reports improved robustness under randomly masked history and stronger gains in complex scenarios such as intersections and roundabouts (Zeng et al., 9 Jul 2025).

4. IL-net for semi-supervised object detection

In "Improving Localization for Semi-Supervised Object Detection" (Rossi et al., 2022), IL-net stands for Improving Localization net and is integrated into a Mean Teacher or Unbiased Teacher semi-supervised object detection pipeline. The setting maintains a Student detector and an EMA-updated Teacher. The Teacher predicts pseudo-labels on weakly augmented unlabeled images, and the Student is trained on labeled data and on strongly augmented unlabeled data using those pseudo-labels. The paper argues that standard pseudo-labeling by confidence thresholding is insufficient because confidence is not strictly tied to localization uncertainty.

The empirical motivation is explicit. The paper reports that pseudo-boxes with IoU $$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$1 represent about $$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$2 of all pseudo-boxes but contribute to more than $$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$3 of classification errors. It also reports that unsupervised regression loss contributes $$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$4–$$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$5 of the total loss, while unsupervised classification contributes at most $$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$6. IL-net addresses this by adding an auxiliary localization classification task that predicts whether a box is high quality with respect to IoU. The head is attached to the ROI stage of Faster R-CNN with FPN and ResNet-50, and computes

$$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$7

where $$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$8 is the shared ROI feature, $$\mathrm{IPOF}_i(\mathbf{E}_i,\mathbf{D}_i)=\vartheta_1\!\big(\mathbf{A}_i^C(\mathbf{E}_i,\mathbf{D}_i)+\mathbf{A}_i^S(\mathbf{E}_i,\mathbf{D}_i)\big).$$9 contains class scores, and $$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$0 contains box regression deltas. The corresponding loss is focal loss,

$$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$1

with focusing parameter $$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$2.

The total Student loss is written as

$$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$3

with supervised and unsupervised decompositions

$$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$4

For supervised IoU targets, proposals are foreground if $$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$5, and are labeled high-quality when $$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$6 with $$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$7. During pseudo-label generation, Teacher boxes are filtered first by the localization-quality threshold $$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$8 and then by the class-confidence threshold retained from Unbiased Teacher:

$$C_i' = 2^{\,2n\,(L_i-1)+b}, \qquad L_i = \left\lceil\,1+\frac{\log_2 C_i' - b}{2n}\,\right\rceil,$$9

with $n=b=2$0. The supervised and unsupervised branches otherwise follow the standard UT protocol, including weak augmentation by random horizontal flip and strong augmentation by color jittering, grayscale, Gaussian blur, and cutout patches.

The training setup uses COCO 2017 with $n=b=2$1 labeled images, validation on minival, $n=b=2$2 iterations, $n=b=2$3 Tesla P100 GPUs, and SGD with learning rate $n=b=2$4, momentum $n=b=2$5, and weight decay $n=b=2$6. The paper sets $n=b=2$7 and finds the best performance at $n=b=2$8. Under this setting, the full ILNet configuration reports AP $n=b=2$9, AP50 $$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$0, AP75 $$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$1, AP$$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$2 $$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$3, AP$$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$4 $$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$5, and AP$$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$6 $$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$7, versus the Unbiased Teacher baseline at AP $$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$8, AP50 $$\boldsymbol{\gamma}=\mathrm{Sigmoid}\!\left(\mathrm{Conv}\!\Big(\mathrm{Cat}\big[\hat{\mathbf{D}^l_{(a+1)/2}},\dots,\hat{\mathbf{D}^l_{a}}\big]\Big)\right),$$9, AP75 $U^2$00, AP$U^2$01 $U^2$02, AP$U^2$03 $U^2$04, and AP$U^2$05 $U^2$06. The reported gain is $U^2$07 AP. Ablations further state that the IoU head alone improves AP by about $U^2$08 to $U^2$09, that IoU-based filtering adds a smaller additional gain, that $U^2$10 is the best training threshold, and that overly large $U^2$11 harms performance (Rossi et al., 2022).

5. Related acronymic extensions: inertial localization and interpretable logic

The inertial localization literature uses ILNet or ILN in a broader class sense rather than as the exact title of the 2023 infrared model or the 2025 forecasting model. "RepILN: Reparameterized Inertial Localization Network" states that ILNet is a general class of inertial localization networks that map IMU streams to trajectories, and presents RepILN as a reparameterized ILNet for IoT constraints (Zhang et al., 21 Jul 2025). The backbone contains eight RepILN Blocks across four stages, where each block consists of a multi-branch RepBlock followed by SA-GCU. During training, the RepBlock computes

$U^2$12

using $U^2$13D convolutions with kernel sizes $U^2$14 and $U^2$15; at inference, branches are structurally reparameterized into a single $U^2$16-kernel convolution with folded BN parameters. Long-range temporal dependencies are modeled by Temporal-Scale Sparse Attention, and fused with local fine-grained features through the Sparse Attention Gated Convolutional Unit. Training predicts velocity with mean squared error and reconstructs position by discrete integration. On RoNIN, RIDI, RNIN-VIO, and TLIO, the paper reports lower ATE and RTE than RoNIN-ResNet; for example, on RoNIN it reports ATE $U^2$17 versus $U^2$18 and RTE $U^2$19 versus $U^2$20, together with a $U^2$21 parameter reduction relative to RoNIN-ResNet in the train model and approximately $U^2$22 in the deploy model.

A distinct use appears in "Neural Logic Networks for Interpretable Classification," where ILNet is described as an Interpretable/Logical Neural Network instantiated by AND, OR, and NOT operations over human-understandable concepts (Perreault et al., 11 Aug 2025). In that formulation, the network learns a disjunctive normal form program composed of an AND layer of rules and an OR output layer, with signed weights indicating whether a concept or its negation is necessary or sufficient for a higher-level concept. The differentiable forward equations are

$U^2$23

for AND concepts, and

$U^2$24

for OR concepts. The method adds biases $U^2$25 and $U^2$26 to account for unobserved necessary and sufficient conditions, uses factorized IF-THEN rule modules for categorical and continuous inputs, and applies discretization, retraining, pruning, and coverage-based bias adjustment after gradient-based learning. The reported empirical results include $U^2$27 F1 on tic-tac-toe with exactly $U^2$28 rules of size $U^2$29, approximately $U^2$30 F1 on Chess (KRKPA7), approximately $U^2$31 F1 on Monk2, and approximately $U^2$32 F1 on the Kidney dataset, together with interpretable extracted rules (Perreault et al., 11 Aug 2025).

6. Comparative themes and disambiguation

Across these papers, the shared acronym conceals markedly different methodological commitments. The infrared ILNet emphasizes low-level saliency preservation and shallow-to-deep enhancement; the trajectory-prediction ILNet emphasizes inverse temporal reasoning and learned anchor refinement; IL-net for semi-supervised detection emphasizes auxiliary localization-quality estimation and pseudo-label filtering; RepILN emphasizes structural reparameterization, sparse temporal attention, and compact inference on IMU streams; the logical ILNet formulation emphasizes explicit rule structure, signed logical weights, and post hoc rule extraction (Li et al., 2023, Zeng et al., 9 Jul 2025, Rossi et al., 2022, Zhang et al., 21 Jul 2025, Perreault et al., 11 Aug 2025).

A common misconception is therefore to treat ILNet as a single lineage. The literature instead uses the acronym for several unrelated architectures spanning infrared vision, autonomous-driving forecasting, semi-supervised detection, inertial navigation, and interpretable classification. This suggests that precise identification requires the paper title, expansion of “IL,” and application domain rather than the acronym alone.

A plausible implication is that the recurrence of the name reflects a shared design preference for problem-specific inductive structure rather than any shared implementation. In the surveyed works, that structure takes the form of low-level IR feature preservation, future-to-history intention inference, localization-quality supervision, hardware-friendly single-path deployment, or explicit DNF logic. The term “ILNet” is therefore best understood as a polysemous label whose technical meaning is entirely context dependent.