WORKSHOP
“Neural networks and deep learning: architecture and physical intuition”
We bring together leading academic scientists and researchers!
September 11th, 2026 1pm - 5pm
Scope of the event
Neural networks underpin the majority of recent breakthroughs in scientific machine learning, yet their inner workings are often treated as opaque. This workshop provides a conceptual and practical introduction to feedforward neural networks, moving from the single perceptron to deep architectures, with emphasis on building physical intuition for what a neural network is doing at each layer — function composition, representation learning, and non-linear basis expansion.
The session develops backpropagation and automatic differentiation as computational tools, and examines the practical choices that determine network performance: activation functions, batch normalisation, dropout, and learning rate schedules. Convolutional neural networks for structured data and recurrent architectures are surveyed briefly to establish vocabulary. Critically, the workshop introduces the concept of symmetry and equivariance — the principle that physical systems possess rotational, translational, and permutational symmetries that generic architectures do not respect — setting the stage for the equivariant architectures developed in depth in Workshops 5, 6, and 7.
Participants will leave with a working understanding of how neural networks learn representations from data, how architectural choices encode assumptions about the problem structure, and why physical systems demand specialised network designs that generic deep learning does not provide.
Extended Key Takeaways & Outputs
· Physical intuition for neural network function: composition, representation, and non-linear basis expansion
· Working knowledge of backpropagation, automatic differentiation, and practical training choices
· Understanding of convolutional and recurrent architectures as building blocks
· Introduction to symmetry and equivariance as design principles for scientific architectures
· Preparation for the specialised architectures addressed in Module II
Agenda & Speakers
13:00–13:05 — Introduction & Objectives
Organizer
· From classical ML to deep learning: why neural networks matter in science
· Linking architecture design to physical intuition and problem structure
· Workshop objectives: understanding representation learning and model design
13:05–13:35 — Foundations of Neural Networks: From Perceptrons to Deep Architectures
Speaker (Academia): N.N.
· Single-layer perceptron and non-linear function approximation
· Deep networks as compositions of functions and feature transformations
13:35–14:05 — Training Neural Networks: Backpropagation and Optimisation
Speaker (Academia): N.N.
· Backpropagation and automatic differentiation
· Optimisation strategies, learning rates, and convergence behaviour
14:05–14:35 — Architectural Choices and Regularisation Techniques
Speaker (Academia): N.N.
· Activation functions, batch normalisation, and dropout
· Preventing overfitting and improving generalisation
14:35–14:45 — Break
14:45–15:15 — Neural Networks in Scientific Applications: Opportunities and Limitations
Speaker (Industry): N.N.
· Applications in physics, chemistry, and materials science
· Challenges in interpretability, robustness, and data efficiency
15:15–15:45 — Structured Architectures: Convolutional and Recurrent Networks
Speaker (Industry): N.N.
· Convolutional neural networks for spatial and structured data
· Recurrent architectures for sequential and time-dependent systems
15:45–16:15 — Symmetry and Equivariance in Scientific Machine Learning
Speaker (Policy/Applied Research): N.N.
· Role of symmetry in physical systems (translation, rotation, permutation)
· Limitations of generic architectures and need for physics-informed design
16:15–16:55 — Panel Discussion & Strategic Alignment
All Speakers + Moderator
· How much complexity is necessary in scientific neural networks?
· Balancing performance, interpretability, and physical consistency
· Preparing for specialised architectures (equivariant and physics-informed models)
· Skills and tools required for researchers entering deep learning
Concluding remarks
Registration
Deadlines:
Early-Bird: until July 15th, 2026
Registration : until September 5th, 2026
The registration takes place online by filling out the form.
The registration and access to the virtual room for workshop is effective only if the fees payment has been confirmed.
Please note that INVIRTA MEMBERS can purchase all our workshops and services with 10% discount . If you desire to get more information on our membership please press here .
Prices* :
Industry: 150 € (early bird**: 125 €)
Academy: 125 € (early bird**: 100 €)
Students***: 90 € (early bird**: 75 €)
* Note that prices are subject to VAT (for European countries, except Germany, provide the VAT number of your organisation to be within the VAT reverse framework payment).
** Until November 15th, 2020
*** Copy of the registration certificate (Ph.D, undergraduates)
Special price offers for Workshop packages
Our Workshop packages are valid for any workshops offered by our academy and can be used anytime. If you want to use one of our packages press here.
