ML results

Handwritten Alphanumeric model

Iteration 1:

Model details	Dataset	Accuracy achieved
Model trained and inference on	Existing dataset	99.80%
Model inference on	NIST dataset	62.70%
Average across all datasets	-	81.25%

Reasoning - As model is trained on the existing dataset, doesn't perform good on NIST dataset

Iteration 2:

Model details	Dataset	Accuracy achieved
Model trained on	Existing dataset + NIST misclassifications	83.30%

Reasoning - After training the model on NIST misclassifications, improvement in accuracy

Iteration 3:

Model details	Dataset	Accuracy achieved
Model trained on	Existing dataset + NIST misclassifications	93.90%

Reasoning - After training the model on NIST misclassifications as per previous checkpoint, improvement in accuracy

Iteration 4:

Model details	Dataset	Accuracy achieved
Model trained on	Existing dataset + manually collected dataset from sheets (~1K samples)	93.90%

Reasoning - Training the model upon previous checkpoint and adding manually collected data, improvement in accuracy

Handwritten Digits model

Iteration 1:

Model details	Dataset	Accuracy achieved
Model trained and inference on	Existing dataset	99.90%
Model inference on	NIST dataset	60.00%
Model inference on	Obtained production data (~50 samples)	97.70%
Average across all datasets	-	85.80%

Reasoning - As model is trained on the existing dataset, doesn't perform good on NIST dataset

Iteration 2:

Model details	Dataset	Accuracy achieved
Model trained on	Existing dataset + NIST misclasifications	96.40%

Reasoning - After training the model on NIST misclassifications, improvement in accuracy

Iteration 3:

Model details	Dataset	Accuracy achieved
Model trained on	Existing dataset + NIST misclasifications + production dataset (~50 samples)	99.70%

Reasoning: After training the model on NIST misclassifications and production dataset, improvement in accuracy

Iteration 4:

Model details	Dataset	Accuracy achieved
Model trained on	Existing dataset + NIST misclasifications + manually collected dataset from sheets (~8.6k samples)	98.30%

Reasoning: As averaging upon a large production dataset, the accuracy slightly dips as compared to iteration 3

Sample dataset images

Existing dataset

Handwritten alphanumeric

Handwritten digits

NIST dataset

Handwritten alphanumeric

Handwritten digits

Manually collected

Handwritten alphanumeric

Handwritten digits

Some unhandled misclassifications

Reasoning: Generally occurs if the digits are written in corners of the cell