Skip to content. | Skip to navigation

Personal tools
Intrinsically Motivated
Cumulative Learning
Versatile Robots
Document Actions

USFD: Focussed Workshop on Action Learning

— filed under:

This workshop involved people from Sheffield (USFD) and Rome (CNR-ISTC-LOCEN) and focused on experimental and computational aspects of action representation, cumulative learning, and intrinsic motivation; and implementation-level discussions of the models both groups are using.

When nov 18, 2010 12:00 to
nov 19, 2010 12:00
Where Sheffield, United Kingdom
Contact Name
Contact Phone 0795 6592649
Attendees USFD
CNR-ISTC-LOCEN
Add event to calendar vCal
iCal

 

This workshop has been completed. Those who can access protected parts of this website (i.e., those who have a username and password) can access pdfs of the slides here.

 

 

 

 

 

Overall goals of workshop

 

Day 1 (Thursday 18 Nov)

General goals: describe and discuss experimental and computational views regarding the following questions:

  1. How are actions represented? Any difference between "hierarchical" and "primitive" actions?
  2. How are actions learned? (e.g., Redgrave et al.'s ideas with intrinsic motivation)
  3. How can we research cumulative learning with experimental methods?
  4. How are intrinsic motivations used? (Comment from Ashvin: I don't think that we should devote a separate "section" to intrinsic motivations. It can be brought up in context of action representation, action learning, and cumulative learning.)

Specific points:

  1. better understanding of experimental work being done at Sheffield
  2. what issues are being or can be addressed by models
  3. specifically, how are these related? what can experimental and modelling work add to each other?


Day 2 (Friday 19 Nov)

General goals: Implementation-level discussion of models. 

  1. Better understanding of each others' models. Make sure we're not "reinventing the wheel."
  2. What exactly is each model trying to do? What scientific question is it addressing?
  3. How easy would it be to integrate two models? 
  4. Plan to make demo with AU. What can we do to develop CLEVER-B2 that is i) scientifically interesting, and ii) will deliver a successful demo in June 2011?

Specific points:

  1. What does a single BG-cortical loop do? How do different loops interact? (e.g., accumbens-PFC loop and dorsolateral striatum-PMC/MI loop?)
  2. Come up with an "operational plan" for point 4 above.

 

Agenda

On 18 and 19 November we have access to the room from  09:00 to 17:00

 

(Please note that, because of a website malfunction, the agenda written here is a bit vague. However, slides from the talks can be viewed (by IM-CLeVeR members) from this page.)

 

Day 1

Short presentations and discussions.

Morning: Action discovery and action representation.

Afternoon: Cumulative learning.

 

Day 2

Short presentations and discussions.

Morning: Understand each others models.

Afternoon: CLEVER-B2

 

 

Snippet from GIanluca's October 2010 summary

(This is literally a copy and paste from page 14 of the 17 page pdf document)

 

Workshop 1
Intrinsic motivations, learning of action repertoires, and the hierarchical organisation of action. Modelling the Board Experiment in CLEVER-B2 and the issue of cumulative learning.

Participants: USFD, CNR-ISTC-LOCEN
Where: Sheffield
When: November/December

This workshop will have these goals and will have this general organisation:

1) Day one. This will involve the modelists of CNR and the experimentalists of USFD (Redgrave, Stafford, …). Of course, the modelists of USFD are welcome to participate if they like/can (Gurney et al.). The goal here would be to understand the reciprocal work of experimentalists and modelists at CNR, and see how the models of CNR can meet and collaborate with respects to the experiments of USFD on hierarchical organisation of actions, in particular intrinsically-based learning of action repertoires. We will also have a second goal: discussing the issue raised by the reviewers for which WP3 experiments, and their models, are not tackling cumulative learning (and abstraction). We will also make plans in relation to the concise document on this issue requested by the PO and due in December (this document will be written up in collaboration with UCBM).


2) Day two: this will involve mainly the modelists of both USFD (Gurney et al.) and CNR. Of course the experimentalists will be welcome if they like (Redgrave et al.). The meeting will be directed to get into the details of the models developed by the two Teams to envisage possible collaborations. We will also have a second goal: discussing the details of the “bio-constrained, hierarchical action learning/selection components” of CLEVER-B2, and see how to collaborate to implement and arrive to the demo due in June 2011 together with AU