jaishreeRAM

What it does

It is an Ml-based calculator. You just need to input an image of your expression into the program and it tells you what the answer is most likely. It can also be used as a validator for your answer.

How we built it

We built it using TensorFlow and Keras. The architecture is a CNN(convolutional network) consisting of four hidden convolutional layers with ReLU activation and three max-pooling layers followed by a softmax output layer. The network is trained over a dataset of digits and the four arithmetic symbols containing around 4800 training images.

Challenges we ran into

1. We ran out of RAM during training very frequently.
2. The dataset we were using was not that comprehensive, so we augmented data using Augly, a data augmentation library by facebook.
3. We had to decide the best model and at times there were issues like overfitting or getting struck in local minima.

Accomplishments that we're proud of

1. We built a deep understanding of neural networks and other unsupervised models that we tried, but failed to get high accuracies. We understood the importance of max-pooling and how to fine-tune the parameters to get the best accuracy within the given memory limit.
2. We got to know how do libraries like NumPy and OpenCV actually function. We also went through the documentation of Augly which helped us understand elaborately about tha library.

What's next for jaishreeRAM

The possibilities for jaishreeRAM are limitless. We plan to build a mathematical expression evaluator which can solve a wider variety of complex questions including algebra and callculus.

Built With

• augly
• python
• tensorflow