Papers I reimplemented and things I built to understand how they work.
A backend-agnostic quantum ML framework over PennyLane and Lightning. It keeps the math separate from the training loop, and runs Monte Carlo barren-plateau detection to tell you whether a circuit is trainable before you spend hours finding out. Still in progress.
The xLSTMTime architecture for long-horizon multivariate forecasting, built from sLSTM and mLSTM blocks with exponential gating. On electricity and traffic data it beats LSTM baselines by 5 to 10% on both MSE and MAE.
Differentiable Tversky similarity as a network layer, from the Tversky Neural Networks paper. It lets a model learn asymmetric similarity instead of assuming a distance metric.
VGG16 features feeding an attention LSTM decoder, trained on Flickr8k. BLEU-1 through BLEU-4 land at 0.485, 0.303, 0.208 and 0.101.
Seq2seq translation with an attention encoder-decoder, trained on parallel corpora.
A DCGAN that generates anime faces. Tuning the learning rates and regularization is what kept the adversarial training from collapsing.
Pulls product details and answers in a sentence instead of a page of results.