A preprint describing a new way to build randomized neural networks reports higher average accuracy than corresponding baseline models in both binary and multiclass tests drawn from 71 benchmark datasets downloaded from the UCI repository. The set contained 32 binary-classification datasets and 39 multiclass datasets. The residual-guided framework was integrated into RVFL, ELM and BLS and compared with each model’s corresponding baseline.
On binary tasks, average accuracy was 82.92% for the residual-guided RVFL version, compared with 80.92% for baseline RVFL. The corresponding ELM figures were 82.76% and 80.82%, while BLS recorded 82.78% against 81.65%. On multiclass tasks, residual-guided RVFL reached 75.94%, compared with 73.33%; ELM reached 75.35% against 73.34%; and BLS reached 76.20% against 74.96%.
A staged search through hidden features
The method changes how the hidden layer is assembled. Here, hidden features are the intermediate signals used before the final prediction, and the residual is the error left by the current model. The proposal greedily constructs hidden features in stages, using a closed-form criterion based on how much that residual can be reduced. At each stage, it generates random candidate units, scores them by the exact reduction in a ridge-regularized training objective, selects the top candidates, and refits the readout, the final output layer, in closed form.
That process has a formal guarantee, but the guarantee is narrower than the accuracy result. The authors prove that adding the selected hidden features cannot increase the ridge-regularized training objective, so the objective is non-increasing from one stage to the next. This is a statement about the model’s training objective. It does not mean that held-out accuracy or generalization will improve at every stage.
Accuracy was not the only reported gain
The reported differences also extended to stability and ranking. For binary tasks, the reported standard-deviation measure fell from 8.91 to 7.88 for RVFL, from 8.36 to 7.76 for ELM, and from 6.24 to 5.75 for BLS. The average rank fell from 1.97 to 1.03, from 1.88 to 1.13, and from 1.91 to 1.09, respectively. The standard deviation is reported across cross-validation folds, so the lower figures indicate less variation in those fold results; the lower average ranks indicate better placement in the method comparison.
The multiclass summary showed the same direction. Reported standard deviation decreased from 10.43 to 9.11 for RVFL, from 10.12 to 8.99 for ELM, and from 7.98 to 7.17 for BLS. Average rank decreased from 1.96 to 1.04, from 1.94 to 1.06, and from 1.86 to 1.14, respectively. The report gives these as aggregate summaries, with no confidence intervals or formal significance tests.
A benchmark result with clear limits
Models were evaluated with five-fold cross-validation and grid-based hyperparameter search. For final reporting, the configuration with the highest mean accuracy was selected, and standard deviation across folds was also reported. The comparison therefore combines mean accuracy with fold-to-fold variability, rather than relying on accuracy alone.
The paper also gives a conditional claim about computational scaling. With candidate-pool size and block size held fixed, the authors state that total computational cost grows approximately linearly with the hidden-unit budget. The supplied text reports no empirical runtime table, so readers are not given a separate timing result alongside the scaling claim.
One supplementary breast-cancer illustration offered a concrete view of the stagewise behavior. As the stage index and hidden-feature budget increased, the training objective remained strictly non-increasing, while residual energy decreased steadily. Because this illustration uses one dataset, it shows the pattern in that example and does not establish generalization.
The evidence remains limited to the benchmark tasks tested. The experiments covered binary and multiclass classification on UCI benchmarks. Regression, multilabel and semi-supervised applications were not evaluated and are described only as future extensions. Within that scope, the reported comparisons were RVFL, ELM and BLS against their corresponding baseline models.
The front matter states that the work was accepted at the 2026 IEEE World Congress on Computational Intelligence. It does not report a funding source.
Paper data and sources
Original title: Residual-Guided Randomized Neural Networks
Authors: Mushir Akhtar, M. Tanveer, Mohd. Arshad
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text