Brain Smarts

About

Inspired by Brain Age, this Android game tests players' math knowledge by hand drawing numbers and quickly solving math problems. With a Play Games account, players can upload their high scores to the leaderboards.

A key feature of this app is that it uses on-device machine learning to interpret handwritten digits. This is done by feeding the MNIST dataset into a neural network created in TensorFlow Lite.

The Neural Network

The neural network was created using Keras and its design is based on the LeNet convolutional architecture for handwritten character recognition. The following describes each layer:

• Input: 1x28x28 (start with a 28x28 px grayscale drawing of the number)
• Conv1: 1x28x28 input, 32 3x3 filters → 32x26x26 output (ReLU activation)
• Conv2: 32x26x26 input, 64 3x3 filters → 64x24x24 output (ReLU activation)
• MaxPool: 64x24x24 input, 2x2 filter, stride 2 → 64x12x12 output
• Dropout: 25%
• Flatten: 64x12x12 input → 9216 layers
• Dense: 9216 layers → 10 layers (softmax activation, classify the number from 0-9)

Because players can draw anywhere on the canvas, data augmentation was used to improve the accuracy of the model—namely rotation, width & height shift, shear, and zoom. The model was trained for 5 epochs and was compressed to a TFLite file using quantization.

How to Run

To run the app, clone this repo and run it in Android Studio.

There are 3 Python files that can be executed after installing the TensorFlow module:

• train_ml_model.py creates the initial MNIST neural network with 98% accuracy
• improve_accuracy.py utilizes data augmentation to improve the accuracy on mobile devices
• write_metadata.py generates info about the tflite file that can be viewed on code generators, such as Android Studio's ML Binding

write_metadata.py is executed with the following syntax:

python3 ./write_metadata.py \
    --model_file=./<path-to>/mnist.tflite \
    --label_file=./labels.txt \
    --export_directory=<output-directory>

---

^{Google Play and the Google Play logo are trademarks of Google LLC.}