Classification-Predictor Network (ClPr)

• ClPr is a family of models that couple classification outputs with secondary prediction tasks across domains such as gaze prediction, image restoration, neural architecture search, loss forecasting, and network-based classification.
• Its implementations range from multitask sequence models to dynamic pruning modules and meta-predictors, balancing improved accuracy with trade-offs like latency and routing sensitivity.
• Key insights include enhanced forecasting precision, FLOP reduction via difficulty routing, and calibrated uncertainty estimation, all validated across both average-case and tail-event performance evaluations.

Classification-Predictor Network (ClPr) is not a single standardized architecture in the arXiv literature. Rather, the label is used for several related constructions in which a classification signal is coupled to a predictive, routing, or inferential objective. In short-term gaze prediction, ClPr denotes a multitask sequence model that jointly classifies eye-movement events and predicts future gaze positions; in adaptive image restoration, it denotes a classifier that predicts restoration difficulty and selects a sparsity level; in neural architecture search, it denotes a predictor of final classification accuracy conditioned on architecture and dataset difficulty; in uncertainty estimation, it denotes a classifier paired with an auxiliary loss predictor; and in Bayesian connectomics it can refer to a classification framework whose predictor is itself a network-valued object (Melnyk et al., 8 Sep 2025, Zhou et al., 2022, Istrate et al., 2018, Gollakota et al., 27 Feb 2025, Guha et al., 2020). Taken together, these usages indicate that ClPr is best understood as a family of classification-conditioned prediction mechanisms rather than as a unique model class.

1. Terminological scope and core design idea

Across the cited literature, the common structure is a coupling between a discrete or probabilistic classification output and a second predictive task. The secondary task varies by domain: forecasting gaze trajectories, selecting a sparse restoration branch, estimating architecture accuracy, predicting incurred loss, or regressing a response from network-valued predictors. What remains stable is the use of classification-related structure to improve another inference problem, or conversely the use of auxiliary prediction to audit or route classification.

Context	Role of ClPr	Representative paper
Gaze forecasting	Joint event classification and gaze prediction	(Melnyk et al., 8 Sep 2025)
Image restoration	Difficulty classifier for dynamic N:M branch selection	(Zhou et al., 2022)
Architecture search	Train-less predictor of peak validation accuracy	(Istrate et al., 2018)
Uncertainty estimation	Classifier plus auxiliary loss predictor	(Gollakota et al., 27 Feb 2025)
Network classification	Classifier with a network-valued predictor	(Guha et al., 2020)

This breadth matters methodologically. In some settings, classification is an auxiliary supervisory signal that enriches temporal representation learning. In others, it is a routing variable that determines computational allocation. In still others, prediction is about the classifier itself: its future accuracy, its expected loss, or the effect of structured predictors on class labels. A plausible implication is that the term has evolved functionally, with its meaning determined by where the classification signal enters the computational graph and what prediction target it conditions.

2. Multitask ClPr for short-term gaze prediction

In "Short-Term Gaze Prediction: Analysis of Individual Differences, Typical and Extreme-Case Errors" (Melnyk et al., 8 Sep 2025), ClPr is a multitask time-series forecasting model for raw eye-tracking data without IMU or scene-image inputs. The task is short-term gaze prediction at a sampling rate of 90 Hz on the Meta Quest Pro, with a main prediction horizon of 44 ms and supplementary analyses at 22 ms and 66 ms. Input window sizes were treated as hyperparameters over 6–20 samples, and the final model text reports an optimized ClPr input length of 12 samples, while the computational table reports input shape $(3,16)$; this suggests minor configuration differences between optimization and benchmarking runs.

The model jointly performs event classification over time and position prediction over time. Its shared trunk consists of a Temporal Convolutional Network encoder, an LSTM module, and an attention-plus-pooling stage. It then branches into a regression head that predicts continuous position differences for the next prediction interval and a classification head that outputs categorical event labels. The forecasting target is positional differences rather than absolute gaze position,

$$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$

The best-performing modality for time-series forecasting was reported as $[ \text{velocities}, \text{heading angle} ]$, and ClPr uses $$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$ as inputs while predicting $$[ \text{positional differences}, \text{classification labels} ]$$.

Event labels derive from a modified MNH algorithm adapted from EyeLink 1000 usage at 1000 Hz to 90 Hz Quest Pro data. Evaluation focuses on fixations, saccades split by amplitude, and post-saccadic periods ("ceps"), with small saccades defined as $< 2^\circ$, large saccades as $> 2^\circ$, short fixations as $< 400$ ms, long fixations as $> 400$ ms, and ceps defined as the 110 ms interval after saccade offset. Smooth pursuit is not part of these RAN tasks. The reported regression objective is the mean Euclidean distance over forecasted steps,

$$e_t = \sqrt{(x_t-\hat{x}_t)^2 + (y_t-\hat{y}_t)^2},\qquad L_{\text{pred}} = \frac{1}{T}\sum_{t=1}^{T}\|\hat{g}_t-g_t\|_2.$$

The paper does not explicitly report the classification loss or multitask weighting, although it notes Adam optimization with learning rate $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$0, batch size $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$1, sliding-window sampling, and PyTorch Lightning training management.

The empirical picture is event dependent. On average, LSTM was most accurate overall and generally best for fixations and saccades, whereas TF and ClPr were best on average for ceps. For extreme-case behavior measured via $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$2, TF was most accurate for ceps and event-agnostic tails, LSTM was best for large saccades and often fixations, and ClPr was competitive but not typically the top $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$3 model. ClPr performed better as the prediction horizon decreased, and subject-level rankings differed between $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$4 and $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$5, especially for ceps and small saccades. Computationally, ClPr has 467.72k parameters, 5.15 MMac, and 0.6855 ms inference time per input sample on an NVIDIA RTX A4000 with CUDA 12.2. The principal trade-off identified is that the classification step adds latency and label dependency, making deployment in strict real-time pipelines more demanding.

3. ClPr as a difficulty predictor in dynamic image restoration

In "ClassPruning: Speed Up Image Restoration Networks by Dynamic N:M Pruning" (Zhou et al., 2022), ClPr corresponds to the paper’s Classifier Module. Here the function of ClPr is not to classify semantic content but to predict restoration difficulty and thereby select an appropriate sparse subnetwork. The input is a degraded image $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$6, the classifier outputs a softmax vector $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$7, and the selected class

$$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$8

determines which restoration branch $$\Delta x_{t+PI} = x_{t+PI} - x_t,\qquad \Delta y_{t+PI} = y_{t+PI} - y_t.$$9 processes the image. The paper uses $[ \text{velocities}, \text{heading angle} ]$0 capacity tiers: dense $[ \text{velocities}, \text{heading angle} ]$1, medium-sparse $[ \text{velocities}, \text{heading angle} ]$2, and high-sparse $[ \text{velocities}, \text{heading angle} ]$3, with easy inputs routed to the most pruned branch and hard inputs to the dense branch.

The ClPr architecture is a lightweight CNN comprising four Conv-BatchNorm-ReLU blocks, global average pooling, and a fully connected layer followed by softmax. For initial supervised pretraining, the model uses pseudo labels constructed differently by task. In deblurring, degraded samples $[ \text{velocities}, \text{heading angle} ]$4 are formed by averaging adjacent blurry frames,

$[ \text{velocities}, \text{heading angle} ]$5

In denoising, deraining, and dehazing, the degradation residual $[ \text{velocities}, \text{heading angle} ]$6 is binned by variance and rescaled according to the median variance of a class,

$[ \text{velocities}, \text{heading angle} ]$7

The pretraining loss is cross-entropy,

$[ \text{velocities}, \text{heading angle} ]$8

End-to-end optimization introduces two additional loss terms. The entropy term

$[ \text{velocities}, \text{heading angle} ]$9

sharpens confidence and stabilizes convergence, while the cost term

$$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$0

encourages the use of cheaper branches when reconstruction is adequate. Reconstruction is handled by the weighted $$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$1 objective

$$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$2

and the final objective is

$$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$3

Training proceeds in three stages: pretrain ClPr with cross-entropy on pseudo-labeled data, freeze ClPr and train the restoration module with $$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$4, then unfreeze all parameters and finetune with $$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$5.

The restoration branches are derived by dynamic N:M fine-grained structured pruning. For each contiguous group $$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$6 of $$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$7 weights, the top-$$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$8 magnitudes are retained and the remainder are zeroed:

$$[ \text{velocities}, \text{heading angle}, \text{classification labels} ]$$9

Because pruning is non-differentiable, the method uses SR-STE for backpropagation, and because different pruning ratios induce different feature statistics, it assigns separate BatchNorm parameters to each branch via Variable BatchNorm. Empirically, the paper reports that ClassPruning can help existing methods save approximately 40% FLOPs while maintaining performance. Representative examples include DnCNN on BSD68, where FLOPs fall from 87.28G to 50.88G while PSNR changes from 31.63 to 31.59 dB at $$[ \text{positional differences}, \text{classification labels} ]$$0, and SRN on GoPro, where FLOPs fall from 217.2G to 131.8G with PSNR changing from 30.21 to 30.18. The main limitation is routing sensitivity: misclassification of difficulty can send hard inputs to overly sparse branches and degrade restoration quality.

4. Train-less ClPr for neural architecture search

In "TAPAS: Train-less Accuracy Predictor for Architecture Search" (Istrate et al., 2018), the ClPr concept is instantiated as a model that predicts the peak validation accuracy a neural architecture would achieve on a dataset without training that candidate architecture. Formally, the predictor maps an architecture encoding $$[ \text{positional differences}, \text{classification labels} ]$$1 and a dataset-difficulty scalar $$[ \text{positional differences}, \text{classification labels} ]$$2 to a predicted accuracy,

$$[ \text{positional differences}, \text{classification labels} ]$$3

The system combines three components: Dataset Characterization, a Lifelong Database of Experiments, and the Train-less Accuracy Predictor.

Dataset difficulty is summarized by the Dataset Characterization Number (DCN), defined as the peak accuracy of a small Deep normalized ProbeNet trained for 10 epochs:

$$[ \text{positional differences}, \text{classification labels} ]$$4

The DCN is used both to condition the predictor directly and to filter relevant past experiments from the Lifelong Database of Experiments according to

$$[ \text{positional differences}, \text{classification labels} ]$$5

The predictor itself is iterative. It uses layer-wise and sub-network encodings, initializes with random-guess accuracy $$[ \text{positional differences}, \text{classification labels} ]$$6, and updates accuracy predictions layer by layer:

$$[ \text{positional differences}, \text{classification labels} ]$$7

The architecture encoding includes layer type, output/input height ratio, output/input depth ratio, number of weights, total number of layers so far, inference FLOPs estimate, inference memory estimate, and a running accuracy field. TAP uses two stacked LSTMs, with 50 hidden units in the first and 100 in the second, followed by a sigmoid output layer. It is trained with RMSprop at learning rate $$[ \text{positional differences}, \text{classification labels} ]$$8, batch size 512, HeNormal initialization, and feature-wise standardization. The training database is substantial: 19 datasets, 800 sampled architectures per dataset, and incremental training of subnetworks to supply accuracy targets.

The reported contribution is speed and cross-dataset calibration. TAPAS predicts in fractions of a second, exceeds 100 networks per second on a single GPU, and supports simulated evolution with 20,000 mutations in 400 seconds on one GPU. In the reported searches, the best discovered networks reach 93.67% accuracy for CIFAR-10 and 81.01% for CIFAR-100 after verification training. Scenario C, which evaluates unseen-dataset prediction across 11 real datasets, shows that including DCN and pre-filtering the experiment database by $$[ \text{positional differences}, \text{classification labels} ]$$9 improves mean squared error, Kendall’s Tau, and $< 2^\circ$0 relative to variants without DCN or without pre-filtering. In this usage, ClPr is a meta-predictor over classifier architectures rather than a component inside a single end-task classifier.

5. ClPr as classifier–loss-predictor coupling

In "When does a predictor know its own loss?" (Gollakota et al., 27 Feb 2025), ClPr denotes a classifier paired with an auxiliary loss predictor. The classifier $< 2^\circ$1 outputs a predictive distribution, while the loss predictor $< 2^\circ$2 estimates the loss that $< 2^\circ$3 will incur on an input. For a binary predictor $< 2^\circ$4 and a proper loss $< 2^\circ$5, the paper uses the decomposition

$< 2^\circ$6

with

$< 2^\circ$7

The baseline loss predictor is the self-entropy predictor,

$< 2^\circ$8

which corresponds, in the multiclass cross-entropy setting, to the entropy of the predicted distribution.

Loss prediction quality is measured by expected squared error on realized losses, and improvement over the self-estimate is quantified by the advantage

$< 2^\circ$9

The theoretical contribution is a tight connection between nontrivial loss prediction and multicalibration. Multicalibration error is defined as

$> 2^\circ$0

and the main theorem states

$> 2^\circ$1

Accordingly, a loss predictor that substantially beats the classifier’s own entropy estimate witnesses a multicalibration failure, and a multicalibration violation can be converted into a loss predictor with nonzero advantage.

The paper distinguishes prediction-only, input-aware, and representation-aware loss predictors, depending on whether $> 2^\circ$2 includes only the output score, the input features, or also internal or external representations. It also identifies blind spots: for strictly proper losses such as squared loss and binary cross-entropy, $> 2^\circ$3 at $> 2^\circ$4, so violations at those scores are not exposed by that loss. Empirically, experiments on UCI Credit Default and Bank Marketing show a robust positive correlation between subgroup multicalibration error and the efficacy of training a loss predictor. In this formulation, ClPr is not a routing or multitask forecast model, but an uncertainty-estimation construct in which prediction targets the classifier’s own future loss.

6. Network-valued predictors and Bayesian ClPr

In "High Dimensional Bayesian Network Classification with Network Global-Local Shrinkage Priors" (Guha et al., 2020), the phrase corresponds to a classification framework in which the predictor is an observed network. For subject $> 2^\circ$5, the response is binary, $> 2^\circ$6, and the predictor is an undirected weighted network $> 2^\circ$7 represented by an adjacency matrix $> 2^\circ$8. Stacking the upper-triangular entries yields an edge feature vector $> 2^\circ$9, and the classification model is logistic regression,

$< 400$0

Equivalently, with a symmetric coefficient matrix $< 400$1,

$< 400$2

The central methodological contribution is a network-aware global-local shrinkage prior. Edge coefficients are centered at a low-rank nodal interaction term $< 400$3, while residual edge effects are subject to either a network lasso prior (BNLC) or a network horseshoe prior (BNHC). In the paper’s notation,

$< 400$4

with node activity indicators

$< 400$5

and spike-and-slab node priors for the latent vectors $< 400$6. This construction allows the model to identify influential nodes and edges rather than treating all edges as unrelated coefficients.

Posterior computation uses Polya–Gamma augmentation for the logistic likelihood, giving conditionally Gaussian updates for $< 400$7. The paper also provides asymptotic guarantees: under stated regularity, sparsity, and scaling conditions, the posterior concentrates around the true model even when the number of edges grows faster than sample size. Empirically, simulation studies report that BNLC and BNHC outperform competitors such as lasso, Bayesian lasso, Bayesian horseshoe, and the Relion et al. network classifier in AUC, edge-selection error, and coefficient estimation. In a brain connectome application with $< 400$8 subjects and $< 400$9 regions of interest, 10-fold cross-validation AUCs were reported as 0.617 for BNLC, 0.598 for BNHC, 0.532 for lasso, 0.466 for Relion, 0.461 for Bayesian lasso, and 0.484 for Bayesian horseshoe. In this usage, the “predictor network” is literal: classification is performed from network-valued covariates, and the ClPr framework is designed to recover interpretable node and edge effects.

7. Comparative perspective and recurring trade-offs

Despite their heterogeneity, these ClPr formulations share several recurrent trade-offs. First, adding a classification signal can improve another predictive objective, but it introduces extra supervision, architectural complexity, or routing sensitivity. In gaze forecasting, the event-classification branch appears beneficial for post-saccadic average accuracy, yet adds latency and depends on label quality. In ClassPruning, the classifier enables substantial FLOP reduction, but misrouting can induce artifacts. In loss prediction, an auxiliary predictor exposes failures of multicalibration, but only insofar as the chosen loss and feature class can witness them. In Bayesian network classification, network-aware structure improves inference and interpretability, but at the cost of more elaborate priors and posterior computation.

Second, ClPr systems frequently mediate between average-case and tail behavior. The gaze paper emphasizes the divergence between $> 400$0 and $> 400$1 subject rankings, showing that strong average performance does not guarantee robustness in extreme cases (Melnyk et al., 8 Sep 2025). The loss-prediction paper makes a formally analogous point in uncertainty estimation: a predictor’s own entropy can be reliable on average while still concealing subgroup-specific multicalibration failures (Gollakota et al., 27 Feb 2025). This suggests that many ClPr designs should be evaluated not only by aggregate error but also by stratified, event-conditional, subgroup-conditional, or tail-sensitive criteria.

Third, the term itself remains context dependent. In one research line it names a multitask sequence model; in another, a difficulty router; in another, an accuracy meta-predictor; in another, a classifier-plus-auditor pair; and in another, a network-based Bayesian regression framework. The strongest unifying description is therefore functional: a Classification-Predictor Network is a system in which classification outputs, class labels, or classification-oriented structure are explicitly used to condition prediction, resource allocation, uncertainty estimation, or structured inference.