Jump to: 

WAI meeting summaries

Meetings take place every Wednesday 10am - 11am online: https://aalto.zoom.us/j/581826846?pwd=N00vdEJ0OHJka1Q0N3BlWTAybHU2Zz09

State of the Art

Previous projects aimed to provide efficient likelihood free inference, cognitive modelling and user knowledge elicitation. In particular, two works encourage further researches:

1. Interactive knowledge elicitation (a): the user evaluates the relevance of covariates in a linear regression (Daee et al. 2017; Micallef et al. 2017).
2. Approximate Bayesian Computation (b): (ABC) used to condition the parameters of the model to data and prior knowledge (Kangasrääsiö et al. 2017).

Project Summary

With White-boxed Artificial Intelligence (WAI), we mean a novel type of interaction between AI and the user, where AI interactively learns from a white-boxed model of the user (i.e. the expert).

Problem: Scientific modelling is a key methodology for understanding the world, yet its potential is being held back by both of the necessary steps being difficult for humans:

1. Formulating the models,
2. Fitting them to data. 

Techniques are lacking for fitting simulator-based models, which most fields use, to data. Equally importantly, few methods are available for supporting users in formulating advanced models. A clear bottleneck in the modeling process is that a modelling expert is required for manually completing all steps. We argue that modelling can be made considerably more efficient by interactive machine learning: an artificial intelligence (AI) which interactively supports the modeller.

Main objective: The main objective of WAI is principles and methods for novel interactive AI that increases the capacity of domain experts to acquire accurate and understandable (“white-box,” in contrast to black-box machine learning) models based on data. In short, we develop a novel type of interactive AI that interactively learns a white-box model of the expert, in order to then improve her ability to construct and understand models that combine her prior knowledge and what can be learned from data.

Technical objectives: We approach this with two ambitious sub-objectives:

1. A novel approach to the theory of mind problem that improves computers' ability to interpret a user's actions. Our sub-goal here is an inverse modelling method that infers a simulator model of a user from interaction data that -for the first time - disentangles her goals, interests, prior knowledge, and capabilities as explanatory factors of her observable actions. 
2. Interactive modelling with the user model, where these simulator models allow the computer to better anticipate how the user will react, and hence allow designing its actions (visualisations, requests for feedback, adaptations) to both elicit maximal amount of user knowledge and be convenient to use and interpretable to the user. This serves as the principle of interactive facilitation and support with the expert in the loop.

Method development: Simulator-based modelling of both the world and the user requires new (M1) likelihood-free inference techniques. Approximate Bayesian Computation (ABC) is an emerging solution which needs to be made scalable both for high-dimensional problems and fast computation required for interactivity. (M2) Computational rationality -based cognitive simulators will be developed for the user models, for interacting with visualisations of the models and data, and (M3) interactive modelling will be formulated as experimental design, where an experiment is an interaction with the user through visualisations. Given an advanced user model, the design requires new and fast enough approximative solutions.

Feasibility: This project is challenging to the extent that it would not be feasible if this consortium would not already have preliminary results on M1, M2 and M3 (Figures 1 and 2). The hardest challenge is to get all the complicated parts to work together on the little data available from the user. The contingency plan, which is completely feasible given this consortium, is to develop groundbreaking new results in each of M1, M2 and M3.

Method Development is a focal point of WAI. It can be summarised by considering three aspects:

1. M1: Likelihood-free Inference Techniques

   When working with ABC, computational efficiency of the algorithm is key. The approach by Guttaman & Corander 2016 has worked really well for models with small to moderate number of parameters. As the dimension of the parametric space increases, scalability and accuracy of the Gaussian Process becomes challenging, requiring further improvement of the ABC procedure (see elfi project). 

2. M2: Computational Rationality based on Cognitive Simulators

   Learning a cognitive simulator from the data is a thrilling challenge. As the cognitive simulators, we use existing models of computational rationality (Kangasrääsiö et al. 2017) which use reinforcement learning in which the discrete state spaces are inherently very large. Increasing the speed of the simulator as well as dealing with the small amount of data are key aspects to address.

3. M3: Interaction knowledge elicitation

   In this part the goal is to design the interaction with the user given a (learning) cognitive simulator and world data. The interaction will be made by either showing to the user a selected part of the data or model, or by showing all data (or full model) and highlighting chosen aspects for user’s feedback or manipulation. With the help of the cognitive simulator of the user, the AI can factor the user’s actions to alternative underlying reasons and in the world modelling focus on the factors attributable to user’s prior knowledge. 

The project is a perfect match to the main content of the programme of conducting basic research and seeking new initiatives for next-generation AI. We address 5 call themes. The unique novel angle is to (1) model cognitive functions, to develop a “theory of mind” to AI for efficient interaction with humans, for (2) interactive machine learning. To achieve this, (3) development of novel machine learning and (4) mathematical statistical methods associated with AI are required. The synergistic operation of the user and the advanced cognitive user model results in a (5) rapidly converging, novel solution for supporting modelling that utilises different approaches.

If we succeed, the improvement over the nowadays state of the art can be summarised as follow:

The complete material presented during the kick-off is available here:

Meetings (see also Meeting notes)

Weekly meetings on Wednesdays 9:45 am - 11am in Otaniemi, Aalto T-building, room A346

• On February 28 the meeting will take place in room A237

Admin

• Current admin: Umberto Simola - umberto.simola@helsinki.fi
• Mailing list: wai-project@helsinki.fi

Main team and Contacts

Recent space activity

• 
  Pierre-Alexandre Murena
  • WAI meeting summaries updated about 4 hours ago view change
• 
  Umberto Simola
  • WAI meeting summaries updated Sep 16, 2020 view change
• 
  Anonymous
  • WAI meeting summaries updated Sep 02, 2020 view change
• 
  Sebastiaan De Peuter
  • WAI meeting summaries updated Aug 19, 2020 view change
• 
  Pierre-Alexandre Murena
  • White-Boxed Artificial Intelligence (WAI) Project Home updated Aug 19, 2020 view change

Show More

Useful links

Modelling Made Easy (Bayesian workflow after 26)

Benchmarks for models of short-term and working memory, research info Bristol

Frontiers | Reinforcement Learning With Parsimonious Computation and a Forgetting Process

Deep Active Inference as Variational Policy Gradients