search
githubEdit

Capture AI/ML Training Data

Steps how to capture field specific training data to improve accuracy of AI/ML predictions

Saral App internally uses AI/ML model to predict scanned layout ROIs(Region Of Interest). Current capabilities of this model is to detect handwritten digits and OMR answers. Internally model is trained with MNISTarrow-up-right database. But as the handwritten digits can be sometimes field specific , this feature is handy to collect training data from production and improve accuracy of the model predictions.

Step 1: For a selected school , make sure storeTrainingData flag to be returned as true from backend API GET /schools.

{
    "schools": [
        {
            "name": "Dummy school 1",
            "schoolId": "up001",
            "state": "up",
            "district": "delhi",
            "block": "haldwani",
            "hmName": "acbc",
            "noOfStudents": "100",
            "storeTrainingData": true,
            "createdAt": "2021-12-06T05:25:38.487Z",
            "updatedAt": "2021-12-06T05:25:38.487Z"
        }
      ]
  }

Step 2: Saral frontend application will capture layout ROIs(Region of Interest) images and convert them into Base64 images and send them as part of PUT /saveMarks API. Saral App also compares the actual preidictions with user overrides if any , and populate training data only if necessary. This training data can be extracted from backend to train the AI/ML model to improve accuracy on iterative fashion.

Sample request from Saral App to backend APIPUT /saveMarks

PreviousConfigurable BrandingNextSupport

Last updated

Was this helpful?