Important For CANSOFCOM judges
Since the challenge document was made for an old hackathon with an online submission portal, we took some liberties in changing how we showed off our model. It is built on a Jupyter Notebook, with the input string read from a input.txt file. We made Y_test vectors for quick testing, but the string outputs are printed at the end of the program in console. Please let us know if this is appropriate for model evaluation! :)
Challenge Statement
Given a set of data points sorted into and labeled with three different classes, design a variational quantum classifier that can classify test data from the same distribution with an accuracy of > 95%.
The Hybrid Model
Layer 1 -> Classical 3 Neuron Layer with ReLU activations
Layer 2 -> 3 Qubit Quantum Circuit, packaged into a QuNode object which can be used as a layer
Layer 3 -> Classical 3 Neuron Layer with Softmax activations, for multi-class classification
The Quantum Circuit
Built with 3 Qubits or Quantum Bits, which use superposition (0 and 1 at the same time!) properties to perform multi state calculations, then record measurements. Features are first encoded into each Qubit given size of input, then sublayers are added with tuneable parameters. These sublayers apply “rotations” to Qubits, using parameters for tuning values. Finally, the final sublayer output is measured using a PauliZ matrix. The entire QuNode object is packaged into a Keras layer. Thanks to PennyLane, we can integrate the layer with our model pretty easily, and it works with various Tensorflow optimization and loss functions!
The Results
The model will always train towards 100% test accuracy on two epochs of training! Try it yourself using the notebook in the Github link below.
Challenges and Learning!
We didn't know anything about quantum computing and had limited knowledge of machine learning going into this weekend. It was a blast to learn about the cool applications of Quantum Bits in models! We had issues with long computation times with our quantum circuit, and had to go through some iterations and ideas before settling on a separate circuit.
Log in or sign up for Devpost to join the conversation.