Project

Project#

Academic year: 2024-2025

Project Description: Build a Deep Learning Pipeline with TensorFlow#

Objective
Develop a complete deep learning pipeline using TensorFlow. Choose one domain—text, audio, or images—select a suitable dataset, and define a specific subject for your project. Develop a complete pipeline, including data preprocessing, model design, training, evaluation, and deployment. Enhance your project by integrating Symbolic AI components for added functionality or interpretability.

Steps to Complete the Project

Choose Your Domain and Subject
- Select a domain: text, audio, or images.
- Define a clear and specific subject related to your chosen domain. Examples include:
  - Text: emotion analysis, paraphrasing, question answering.
  - Images: image classification, object detection.
  - Audio: speech recognition, emotion detection, music vs. speech classification.
Select a Dataset
- Pick a dataset from the provided sources or propose your own.
- Ensure the dataset is relevant to your chosen domain and task.
Preprocess the Data
- Clean, transform, and augment the data as needed.
- Use TensorFlow tools like tf.data or tf.keras.preprocessing for efficient pipelines.
- For text, consider tokenization or embedding; for images, apply normalization or augmentation; for audio, extract features like spectrograms or MFCCs.
Design and Train Your Model
- Build a model suitable for your task:
  - For text: use RNNs, LSTMs, or Transformers.
  - For images: use CNNs or pre-trained architectures like ResNet.
  - For audio: combine feature extraction layers with RNNs or CNNs.
- Experiment with hyperparameters, activation functions, and layers.
- Train your model using TensorFlow and evaluate its performance on a validation set.
Integrate Symbolic AI (Optional Bonus)
- Combine your model with rule-based or logic-driven systems to improve interpretability or accuracy.
- For example:
  - Use knowledge graphs in text analysis.
  - Add reasoning components for emotion recognition in audio.
  - Implement rule-based constraints for object detection in images.
Evaluate and Deploy
- Assess your model using metrics appropriate to your task (e.g., accuracy, precision, recall).
- Deploy your model as an interactive application or notebook.

Deliverables

A complete TensorFlow implementation of your pipeline.
A detailed report covering:
- The chosen subject, problem statement, and objectives.
- The dataset used.
- Preprocessing methods.
- Model architecture and training process.
- Evaluation results and potential improvements.
A deployed demo or app.

Example Notebooks#

Project domains#

Text
Images
Audio

Datasets#

Existing catalogues#

Domains#

Text:
- Start with datasets like IMDB reviews, SQuAD, or CoNLL-2003.
- Use pre-trained embeddings like GloVe, Word2Vec, or BERT.
Images:
- Use datasets such as CIFAR-10, ImageNet, or Oxford Flowers.
- Try transfer learning with TensorFlow’s pre-trained models.
Audio:
- Choose datasets like LibriSpeech or UrbanSound8K.
- Preprocess with audio-specific techniques like spectrograms.

Possible topics#

Text
- Language identification
- Speaker identification
- Question answering
  - yes or no answering
  - answers to questions related to multiline paragraphs
  - mathematical question answering
- Analysis of citations
- Analysis of reviews
- Paraphrasing
- Common knowledge facts
- Common sense explanation
- Analysis of emotions
Images
- Object detection
- Image classification
Audio
- Detection of music genre
- Analysis of musical notes
  - pitch, timbre, envelope, etc.
- Analysis of sentiments
- Speech recognition
  - Single speaker
  - Multiple speakers
  - Accents
- Emotion recognition
- Distinction between speech and music
- Speech commands
- Transcription