The library. Every paper below is real and readable in minutes — abstract, plain words, honest results (including the negative ones), and what each one taught me.
Part-of-speech tagging for Bangla — a low-resource language whose main benchmark, Microsoft IL-POST, is severely class-imbalanced — using contextual embeddings from three Bangla BERT models. On macro-F1, which weights all 30 classes equally, a neural network beats a decision tree overall but wins by feasting on the majority classes; the tree scores lower yet spreads its errors more evenly across the tagset. That asymmetry motivates an inverted distillation: read the class counts in the tree's leaf nodes as a probability distribution and distill that "dark knowledge" from the tree into the neural student — a de-biasing signal rather than a compression trick.
A comprehensive review of the state of the art in quantum machine learning, paired with hands-on classification experiments: two first-generation hybrid quantum-classical models — a quanvolutional neural network on a gate-based simulator and a continuous-variable quantum neural network on a photonic simulator — trained on MNIST and compared head-to-head against classical baselines of comparable size on accuracy and convergence. Both hybrids trail their baselines; the margins, and where each model loses, are the contribution.
I fine-tuned a pretrained RetinaNet (ResNet backbone, feature pyramid network, focal-loss classification head) on the BCCD microscopy dataset to detect red blood cells, white blood cells, and platelets — 364 images, 4,888 annotations, three classes, split 255/73/36 train/val/test. Only the classification head was modified, from 80 classes to 3; everything else transferred from pretrained weights. The detector reached mAP 0.876 at IoU 0.5 and 55.25% at IoU 0.50:0.95 on the held-out test split, running entirely on a 4 GB laptop GPU. The technical result is ordinary; the argument the paper builds around one omitted augmentation is the part worth reading.
Signature-based intrusion detection can't see attacks it has no signature for; its coverage of unseen attacks is zero by construction. This work reframes intrusion detection as anomaly detection — learn the shape of normal traffic so well that anything abnormal announces itself — and uses knowledge distillation to compress that capability toward something deployable. On the CICIDS2017 benchmark we train four classical supervised models, select the strongest (a decision tree) as a teacher, and distill it into a neural student. The teacher performs near-ceiling; the student loses roughly 22 macro-F1 points in the transfer. We report the failure and its mechanism as the main finding: a decision tree makes an excellent classifier and a poor teacher, for the same structural reason.