Barnamala is a compact convolutional network with 1.11 million parameters for 46-class handwritten Devanagari recognition on DHCD. The paper reports 99.73% accuracy, matching the nearest large-model baseline with 17.32 million parameters while using 15.6 times fewer parameters. Every tested configuration, including large teacher ensembles, reportedly reaches the same 11-error floor, with no statistically clear winner under exact McNemar tests and Wilson confidence intervals. Zero-shot accuracy on CMATERdb digits is 76.6%, rising to 97.8% after fine-tuning. Mean corruption accuracy is 75.7%, compared with 38.7% for large baselines.
No heat snapshots are available in the last 24 hours.