Probabilistic Graphical Models (PGM) 2024

Wednesday September 11

Speaker: Arto Klami (Helsinki University)
Title: Prior distribution – how should we really choose them?

Abstract:
Learning in the context of probabilistic graphical models often builds on Bayesian inference, where prior assumptions on either the parameters or the conditional structure are updated based on the observed data. Much of the literature focuses on efficient inference algorithms or new model designs and largely avoids the question of the prior distribution, either assuming a simple conjugate prior with default parameters or asking a practitioner to encode their subjective knowledge as a valid prior distribution. In practice, however, it is often near impossible to faithfully encode the true prior knowledge and we need tools for assisting this process.

This talk focuses on the question of how to choose the prior distributions in varying scenarios, describing concepts and tools that help forming better priors with less cognitive and computational effort. I will introduce the concept of prior elicitation as means of transforming tacit human knowledge into valid priors and discuss the current state of practical prior elicitation tools. The talk also goes through some examples of modern prior specification methods, for both elicitation of subjective knowledge and for automating the prior choice without relying on repeated posterior inference.

Thursday September 12

Speaker: Sara Magliacane (University of Amsterdam)
Title: Causal representation learning in temporal settings

Abstract: Causal inference reasons about the effect of unseen interventions or external manipulations on a system. Similar to classic approaches to AI, it typically assumes that the causal variables of interest are given from the outset. However, real-world data often comprises high-dimensional, low-level observations (e.g., pixels in a video) and is thus usually not structured into such meaningful causal units. Causal representation learning aims at addressing this gap by learning high-level causal variables along with their causal relations directly from raw, unstructured data, e.g. images or videos.

In this talk I will focus on learning causal representations in temporal sequences, e.g. sequences of images. In particular I will present some of our recent work on causal representation learning in environments in which we can perform interventions or actions. I will start by presenting CITRIS, where we leveraged the knowledge of which variables are intervened in each timestep to learn a provably disentangled representation of the potentially multidimensional ground truth causal variables, as well as a dynamic bayesian network representing the causal relations between these variables. I will then show iCITRIS, an extension that allows for instantaneous effects between variables. Finally, I will focus on our most recent method, BISCUIT, which overcomes one of the biggest limitations of our previous methods: the need to know which variables are intervened. In BISCUIT we instead leverage actions with unknown effects on an environment. Assuming that each causal variable has exactly two distinct causal mechanisms, we prove that we can recover each ground truth variable from a sequence of images and actions up to permutation and element-wise transformations. This allows us to apply BISCUIT to realistic simulated environments for embodied AI, where we can learn a latent representation that allows us to identify and manipulate each causal variable, as well as a mapping between each high-level action and its effects on the latent causal variables.
 

The conference welcomes contributions on all aspects of graphical models, including but not limited to:

• Principles of Bayesian networks, chain graphs, sum-product networks, influence diagrams, probabilistic relational models, and other probabilistic graphical models (PGMs).
• Information processing in PGMs (exact and approximate inference).
• Learning and data mining in the context of PGMs: machine learning approaches, statistical testing and search methods, MCMC simulation.
• Construction and explanation of PGMs, exploiting results from related disciplines such as statistics, information theory, optimization, deep learning, and decision making under uncertainty.
• Software systems based on PGMs.
• Application of PGMs to real-world problems.

Both theoretical and applied contributions are welcome.

Submit your paper

Papers submitted for review should report on original, previously unpublished work. The deadline for abstract submissions is May 20th, 2024 and the deadline for paper submissions is May 27th, 2024.

Accepted papers will be published electronically in a volume of the Proceedings of Machine Learning Research. 

Conference papers must be submitted electronically through Microsoft CMT. Please download the Author Kit here. Submissions need to be prepared using this Author Kit and are limited to 12 pages excluding references. Each submission will go through a standard reviewing process. Papers will be accepted for either plenary or poster presentation, but, in the proceedings, no distinction is made between the two.

At least one author of each accepted paper is required to register for the conference.

There will be a workshop on PGMs in Health Care, co-located with the conference, just prior to PGM 2024 (on September 10, 2024).

IJAR Special Issue

Selected papers will be invited to be extended and then submitted to a special issue of the International Journal of Approximate Reasoning (IJAR).

Best student paper award

Our sponsor BayesFusion offers a best student paper award.

Travel support grants for students

The Donders Institute has sponsored a number of grants to support student travel. We are looking into additional funding for student travel grants. More information on how to apply will be provided once the registration period is open.

Deadlines

• May 20, 2024: Abstract submission deadline
• May 27, 2024: Paper submission deadline
• July 12, 2024: Author notification
• August 26, 2024: Camera-ready copy due
• September 11-13, 2024: Conference

Hotels

• Mercure Hotel - € 104
• Hotel Nimma - € 81
• Appart Hotel Oranjestaete - € 125
• La Boutique - € 116
• Bastion Hotel - € 91
• Van der Valk Lent - € 99
• Barbarossa hostel, City Center - € 74
• BB Noviomagnus, City Center - € 77
• Blue Nijmegen, City Center - € 135
• Guesthouse Vertoef, City Center - € 82
• Holt, City Center - € 99
• Hotel Courage Waalkade, City Center - € 103
• Hotel Credible, City Center - € 116
• Hotel Pauw, City Center - € 111
• MANNA - € 151
• Hotel de Prince - € 118
• Sanadome Hotel & Spa - € 214

B&B's

• Aan de Anna B&B 
• Opoe Sientje  - € 115
• B&B pension Bottendaal 
• B&B de Smederij - € 120
• B&B Het Blauwe Uur - € 121

The best way to get around the city of Nijmegen is by bike. Parking is expensive throughout the whole city (€ 5 per hour).

Nijmegen is excellently connected by public transport (bus). You can search Google Maps for your connection. The payments can be done via your bank card or your phone (when your bank card is imported).

From outside the Netherlands:

Flight to either Amsterdam Schiphol or Eindhoven airport, take the train to Nijmegen Central Station.

Nijmegen Central Station to the Conference location de Lindenberg:

• Bus: 58 / 300 / 331 / 5 / 6 / 14 / 85 / 2 / 8
• Bike: you can rent a bike at the station
• Walk: the Lindenberg is in walking distance (20 min.)

Nijmegen Central Station to Radboud University:

• Bus: 6 / 8 / 83 / 10 / 58 / 9 / 15
• Bike: Rent a bike at the station for a short period or elsewhere in the city for the time of your stay
• Walk: 50 min