My Research Summary

Research Overview

Can we achieve the predictive power of deep learning without its computational burden and training instability?

My research addresses this question through Randomized Neural Networks (RdNNs)—a theoretically grounded learning paradigm that replaces iterative gradient-based optimization with closed-form solutions. Unlike deep neural networks that require millions of parameters, specialized GPUs, and careful hyperparameter tuning, RdNNs randomly initialize hidden-layer parameters and learn only the output weights. This leads to fast training, scalability, and guaranteed global optimality, making them particularly suitable for data-limited and resource-constrained environments.

My PhD research advances randomized learning along three complementary directions: interpretability, robustness and scalability, and real-world biomedical validation.

References:

[1] M.Sajid, M. Tanveer and P. N. Suganthan (2025). “Ensemble Deep Random Vector Functional Link Neural Network Based on Fuzzy Inference System,” in IEEE Transactions on Fuzzy Systems, v. 33, no. 1, pp. 479-490, DOI: https://doi.org/10.1109/TFUZZ.2024.3411614 [I.F.: 11.9]

[2] M.Sajid, A. K. Malik, M. Tanveer and P. N. Suganthan (2024). “Neuro-Fuzzy Random Vector Functional Link Neural Network for Classification and Regression Problems,” in IEEE Transactions on Fuzzy Systems, v. 32, no. 5, pp. 2738- 2749, DOI: https://doi.org/10.1109/TFUZZ.2024.3359652 [I.F.: 11.9]

Interpretable Randomized Neural Networks

Although RdNNs are computationally efficient and analytically tractable, they are often regarded as non-interpretable black-box models due to their random hidden representations. A central focus of my research has been to embed interpretability directly into randomized learning architectures, rather than relying on post-hoc explanation techniques.

To achieve this, I developed neuro-fuzzy RVFL models and their ensemble deep variants, which integrate an intermediate fuzzy inference layer encoding human-interpretable IF–THEN rules. This structured layer anchors subsequent network computations to a semantically meaningful representation, enabling intrinsic interpretability while preserving the efficiency of randomized learning.

Extensive experiments on 80+ benchmark datasets demonstrate that these models achieve competitive predictive performance while providing transparent feature-level explanations. On breast cancer diagnosis datasets, the models autonomously identified clinically relevant tumor-related features, aligning closely with established medical biomarkers and achieving up to 95% classification accuracy. This line of work establishes that interpretability and performance need not be competing objectives in efficient tabular learning systems [1, 2].

Robust and Scalable Randomized Learning

A second core theme of my research focuses on improving the robustness of RdNNs under noise, outliers, and class imbalance, conditions that frequently arise in real-world data. Conventional RdNN formulations rely heavily on squared-error objectives and uniform sample treatment, making them sensitive to corrupted or atypical observations.

I addressed these limitations by redesigning how data are represented, weighted, and penalized during learning. One key contribution introduces granular computing into RdNNs [3], where samples are grouped into compact representative granules. Learning from these granules emphasizes stable regions of the data distribution while suppressing noise and outliers, simultaneously improving robustness and scalability. In practice, this approach reduces the effective training set by 50–70%, alleviating matrix-inversion bottlenecks without degrading performance.

At the objective-function level, I developed robust loss functions (including Wave loss and XG loss [4, 5]) that introduce bounded and asymmetric penalization of errors. These formulations improve robustness to outliers by 12–18% on noisy datasets while maintaining closed-form solvability. I further incorporated intuitionistic fuzzy weighting and graph-based embeddings to model sample reliability, class imbalance, and intrinsic data geometry, enabling adaptive, uncertainty-aware learning [3, 6, 7, 8].

References:

[3] M. Sajid, A. Quadir, and M. Tanveer (2025). “GB-RVFL: Fusion of Randomized Neural Network and Granular Ball Computing,” in Pattern Recognition, Elsevier, v. 159, p. 111142, DOI: https://doi.org/10.1016/ j.patcog.2024.111142 [I.F.: 7.6]

[4] M. Sajid, A. Quadir, and M. Tanveer (2024). “Wave-RVFL: A Randomized Neural Network Based on Wave Loss Function,” in The 31st International Conference on Neural Information Processing (ICONIP), v. 15287, pp. 242-257, DOI: https://doi.org/10.1007/978-981-96-6579-2_17 [Core Rank B Conference]

[5] M. Akhtar, A. Kumari, M. Sajid, A. Quadir, M. Arshad, P. N. Suganthan and M. Tanveer (2025). “Towards Ro bust and Inversion-Free Randomized Neural Networks: The XG-RVFL Framework,” in Pattern Recognition, Elsevier, p. 112711, DOI: https://doi.org/10.1016/j.patcog.2025.112711 [I.F.: 7.6]

[6] M. A. Ganaie∗, M. Sajid∗, A. K. Malik and M. Tanveer (2024). “Graph Embedded Intuitionistic Fuzzy Random Vector Functional Link Neural Network for Class Imbalance Learning,” in IEEE Transactions on Neural Networks and Learning Systems, v. 35, no. 9, pp. 11671–11680, DOI: https://doi.org/10.1109/TNNLS. 2024.3353531 [I.F.: 8.9] ∗ M. Sajid and M. A. Ganaie contributed equally to this work.

[7] M. Sajid, A. K. Malik and M. Tanveer (2024). “Intuitionistic Fuzzy Broad Learning System: Enhancing Robustness Against Noise and Outliers,” in IEEE Transactions on Fuzzy Systems, v. 32, no. 8, pp. 4460- 4469, DOI: https://doi.org/10.1109/TFUZZ.2024.3400898 [I.F.: 11.9]

[8] M. Tanveer, A. Mishra, M. Sajid, and A. Quadir (2025). “BLS-CIL: Class Imbalance Broad Learning System Via Dual Weighting and Layer Trimming”, in Pattern Recognition, Elsevier, DOI: https://dx.doi.org/ 10.2139/ssrn.5237998”.

References:

[9] M. Sajid, R. Sharma, I. Beheshti and M. Tanveer (2024). “Decoding Cognitive Health Using Machine Learning: A Comprehensive Evaluation for Diagnosis of Significant Memory Concern,” in WIREs Data Mining and Knowledge Discovery, v. 14, no. 5, p. e1546, DOI: https://doi.org/10.1002/widm.1546 [I.F.: 11.7]

[10] M. Tanveer, M. Sajid, M. Akhtar, A. Quadir, T. Goel, A. Aimen, S. Mitra, Y-D Zhang, CT Lin and J. Del Ser (2024). “Fuzzy Deep Learning for the Diagnosis of Alzheimer’s Disease: Approaches and Challenges,” in IEEE Transactions on Fuzzy Systems, v. 32, no. 10, pp. 5477- 5492, DOI: https://doi.org/10.1109/TFUZZ. 2024.3409412 [I.F.: 11.9]

Real-World Validation: Neurodegenerative Disease Diagnosis

An important aspect of my work is demonstrating that theoretically grounded robustness translates into practical performance on challenging real-world problems. I validated my models on Alzheimer’s disease (AD) diagnosis, a domain characterized by severe class imbalance, high-dimensional noisy features, and limited labeled data.

Using both MRI-derived tabular features and EEG time-series data, I addressed multi-class classification of Alzheimer’s disease, mild cognitive impairment, and cognitively normal populations. The proposed RdNN frameworks achieved stable and reliable performance across disease stages by integrating adaptive sample weighting and geometry-preserving graph embeddings. I further explored complex-valued feature representations that capture complementary spectral information, leading to 6–9% accuracy improvements on ADNI benchmarks [9, 10]. These results confirm that randomized learning frameworks can deliver robust, interpretable, and computationally efficient solutions in high-stakes biomedical applications.

References:

[11] A. Quadir, M. Sajid, and M. Tanveer (2025). “Granular Ball Twin Support Vector Machine,” in IEEE Transactions on Neural Networks and Learning Systems, v. 36, no. 7, pp. 12444- 12453, DOI: https://doi.org/10.1109/TNNLS.2024.3476391 [I.F.: 8.9]

[12] R. Mishra, M. Sajid, M. Tanveer (2025). “SpectRA: Parameter-Efficient High-Rank Singular Vector Adaptation for Large Language Models”, [Under Review]

Kernel Methods and Efficient Foundation Models

Kernel methods form an important part of my research, particularly for learning robust decision boundaries and modeling nonlinear relationships in high-dimensional spaces. My work focuses on designing efficient and theoretically grounded kernel-based frameworks, carried out through collaborative research and supervision of Master’s theses. A key direction involves Restricted Kernel Machines (RKMs), inspired by Suykens’ work on deep RKMs. We developed the first one-class RKM for anomaly detection (published at ICONIP 2024) and subsequently proposed the first Twin RKM for multiview learning (IJCNN 2025).

I have also contributed to enhancing Twin Support Vector Machines (Twin-SVMs) by improving their robustness, scalability, and efficiency. This includes the Granular Ball Twin-SVM (published in IEEE TNNLS [11]) and a Least Squares Twin-SVM variant (Pattern Recognition). More recently, I have explored unifying kernel methods and randomized neural networks through granular computing, combining the strengths of both paradigms; this work is currently under revision.

More recently, I have begun exploring parameter-efficient fine-tuning strategies for large language models (LLMs), focusing on low-rank and structured adaptation to enable task-specific learning with minimal computational overhead. This direction reflects my broader interest in scalable, controllable, and optimization-aware AI systems, spanning both classical machine learning and modern foundation models [12].

Page updated

Google Sites

Report abuse