Deep learning: CNN, RNN/LSTM, transfer learning

Objectives

• Train a dense neural network on tabular data and compare it with a gradient-boosting model, in order to decide, with figures to back it up, which one to keep.
• Build and train a convolutional network to classify images or scanned documents, with data augmentation and monitoring of the learning curves.
• Model a sequence (text or time series) with a recurrent or LSTM network and compare it with a classical model.
• Apply transfer learning from a pre-trained model to reach a target performance with a small dataset.
• Document a training run so that it can be reproduced: data, preprocessing, hyperparameters, curves, results, limits.

Audience

Engineers and data scientists who have completed DATA-201 or already practise machine learning; R&D engineers; technical teams in charge of image or document processing.

Prerequisites

Having completed DATA-201, or being able to train and evaluate a supervised model in Python with scikit-learn. Checked by a placement test. One laptop per participant; an online GPU environment with PyTorch and TensorFlow preinstalled, provided by Yolea, credits included, with no local installation; for remote delivery, a stable connection of at least 10 Mbit/s, microphone and camera. Training datasets provided (scanned documents, images of parts, operational series); the participant's own dataset is accepted if it is anonymised.

Programme

• Designing and training neural networks
• Computer vision applied to documents
• Sequence modelling
• Reusing pre-trained models
• Reproducibility and documentation

Format

In-company or open enrolment. 14 hours, 2 consecutive days of 7 hours; remotely, 4 sessions of 3 hours 30 minutes over two weeks. Can follow DATA-201 directly. On site, live remote or hybrid; 8 to 12 participants. 30% input, 70% notebook workshops on GPU: three architectures trained in session, one per half-day, then transfer learning on the last afternoon; a mandatory comparison with a classical model is the rule.

Assessment

Beforehand, a placement test. During the course, check quizzes and a practical challenge (reach a target performance on an imposed image set, with a training report assessed against a grid). At the end, a same-day feedback survey and a named certificate of course completion. Three weeks later, a follow-up check-in.