Mushir Akhtar

Research profile

Research summary

Mushir Akhtar is a first-author machine learning researcher at the Indian Institute of Technology Indore whose 2026 output, cataloged by Evolve from arXiv, spans three evenly weighted strands: core machine learning, robust randomized neural networks, and evaluation methodology for medical imaging AI. In the imbalanced-learning strand, BARGE combines a bounded prior-adjusted score with reliability-guided embedding geometry, drawing on proper scoring rules and density-power divergence; Evolve's verdict credits clean-data gains over logit adjustment while noting a modest edge under noise (70.0 vs 72.3 MBE average) and missing noisy long-tailed baselines. A parallel line hardens broad learning systems against label noise, first through a wave loss and then in IFW-BLS, which adds intuitionistic fuzzy sample weighting and a bounded asymmetric wave loss — judged sound, though grid-best test reporting inflates its gains. The third strand is a first-authored audit of four medical vision-language models for chest X-ray tuberculosis screening, showing that zero-shot rankings, reliability and thresholds often fail to port across cohorts and prompts. The bibliometric snapshot lists 7 distinct co-authors (collaboration score 41/100) and a mean reproducibility score of 55/100 across three scored papers, each rated medium readiness.

Research keywords

Highlights

• Auditing four medical vision-language models for chest X-ray tuberculosis screening

  A first-authored evaluation-methodology audit that goes beyond benchmark scores: zero-shot rankings, reliability and thresholds often fail to port across cohorts and prompts, with explicit attention to the negative spectrum in TB screening.

• BARGE: bounded adjustment with reliability-guided embedding for noisy, imbalanced learning

  Couples a bounded prior-adjusted score with reliability-guided embedding geometry, drawing on proper scoring rules and density-power divergence; Evolve notes clean-data gains over logit adjustment but a modest noisy edge (70.0 vs 72.3 MBE average) and no noisy long-tailed baselines.

• IFW-BLS: dual-robust broad learning system with intuitionistic fuzzy wave loss

  Pairs intuitionistic fuzzy sample weighting with a bounded asymmetric wave loss to harden broad learning systems against label noise; judged sound overall, though grid-best test reporting inflates the reported gains.

• Wave-loss robust broad learning system for classification under data uncertainty

  A first-authored broad learning system paper applying a wave loss for classification under data uncertainty; the catalog entry carries no field, keywords, verdict or slug, making it the least-documented item in the record.

Timeline

1. 2026Published a first-authored robust broad learning system using a wave loss for classification under data uncertainty.
2. 2026Extended the broad learning line with IFW-BLS, adding intuitionistic fuzzy sample weighting and a bounded asymmetric wave loss.
3. 2026Released BARGE for imbalanced learning with noisy labels, combining bounded prior-adjusted scores with reliability-guided embeddings.
4. 2026First-authored an audit of four medical vision-language models for chest X-ray tuberculosis screening, examining cross-cohort and cross-prompt portability.

Papers in the Evolve catalog

• 2026-09-18first author

  Beyond Benchmark Scores: Auditing Medical Vision-Language Models for Chest X-Ray Tuberculosis Screening

• 2026-09-14first author

  Bounded Adjustment with Reliability-Guided Embedding for Imbalanced Learning with Noisy Labels

• 2026-09-02first author

  IFW-BLS: Dual-Robust Broad Learning System with Intuitionistic Fuzzy Wave Loss

Frequent co-authors