Final 35% of the grade with a short oral defense · proposal around week 8 · due week 13
The final project takes a task end to end with any architecture family from the course. Work solo or in a pair: frame the task, build a suitable network, train and tune it, diagnose failures, and report results. The examples below span vision, sequences, and representation learning; or propose a variant instead.
Data: A small image dataset.
Idea: When does transfer learning help?
Methods: Compare from-scratch, frozen-features, and fine-tuning a pretrained backbone under a fixed budget.
Evaluation: Accuracy versus training cost across regimes; characterize when transfer wins.
Stretch goal: Vary the target dataset size and find the crossover point.
Data: A small text sentiment set (for example an IMDB or SST subset).
Idea: Sequence classification.
Methods: An embedding layer with an LSTM or GRU.
Evaluation: Accuracy and F1; compare LSTM versus GRU and long versus short inputs.
Stretch goal: Add pretrained word embeddings.
Data: A plain-text corpus (for example a public-domain book).
Idea: Generate text with a recurrent net.
Methods: A character-level RNN or LSTM.
Evaluation: Held-out perplexity plus qualitative samples; study the sampling temperature.
Stretch goal: Compare an RNN with an LSTM at matched size.
Data: MNIST or FashionMNIST with added noise.
Idea: Denoise images and inspect the latent space.
Methods: A convolutional autoencoder.
Evaluation: Reconstruction error and latent-space structure (clusters, interpolation).
Stretch goal: Use reconstruction error for anomaly detection.
Data: CIFAR-10 (labels withheld for pretraining).
Idea: Learn features without labels.
Methods: A SimCLR-style setup: augmentations, a contrastive loss, and a linear probe.
Evaluation: Linear-probe accuracy versus a supervised baseline at low label counts.
Stretch goal: Compare two augmentation policies and their effect on the representation.
Data: A small image-to-sequence toy dataset.
Idea: Combine vision and sequence models.
Methods: A CNN encoder feeding an RNN decoder.
Evaluation: A task-appropriate metric; ablate the encoder.
Stretch goal: Add an attention-free pooling variant and compare.
Data: A public time series (for example energy or weather).
Idea: Forecast future values.
Methods: An RNN or LSTM against a naive or linear baseline.
Evaluation: MAE and RMSE; study how error grows with the forecast horizon.
Stretch goal: Add exogenous features and measure the gain.
Data: A small audio set (for example an ESC-50 or UrbanSound subset).
Idea: Treat spectrograms as images.
Methods: A CNN, optionally a pretrained image backbone.
Evaluation: Classification accuracy; try time and frequency augmentation.
Stretch goal: Compare raw-waveform input against the spectrogram.
Data: A task of the student's choice.
Idea: A careful, controlled empirical study.
Methods: A fixed architecture with a rigorous protocol over optimizers and regularizers.
Evaluation: Curves with multiple seeds and visible variance, leading to clear conclusions.
Stretch goal: Add confidence intervals via bootstrapping.
Data: A Kaggle-style image or tabular problem.
Idea: Solve a real problem end to end.
Methods: An appropriate neural architecture with a sound train/validation/test protocol.
Evaluation: A leaderboard-style metric reported honestly on a held-out split.
Stretch goal: Compare against a strong non-neural baseline.