1. Consciousness and the Frame Problem
Murray Shanahan
Dept. of Computing
Imperial College London

2. Overview Global workspace theory Coalitions and the core
The frame problem
Coalition combinatorics
Coalition dynamics

3. Global workspace theory

4. Global Workspace Architecture
Parallel
Specialist Processes
Global Workspace
Multiple parallel specialist processes compete and co-operate for access to a global workspace (Baars, 1988)
When they win control of the global workspace, processes get to broadcast to the entire set of specialists, and exercise widespread influence

5. The Tenets of the Theory
The global workspace architecture harnesses massively parallel computation
The global workspace itself exhibits a serial procession of states
Yet each state-to-state transition is the result of filtering and integrating the contributions of huge numbers of parallel computations
According to global workspace theory (Baars, 1988)
The human brain instantiates such an architecture
Activity confined to the specialists does not contribute to consciousness
The conscious condition is associated with broadcast

6. From Fibre Tracts to Networks
How might the brain instantiate a global workspace architecture?
The most likely substrate for a global neuronal workspace is the cortical white matter (Dehaene, et al., PNAS 1998).
Network structure
(Hagmann, et al., PLoS Biol )
Cortical white matter

7. Small Worlds and Modules
A non-modular
small-world network
A modular small-world network
without connector hubs

8. Connective Cores The connective core (connector hubs and their
inter-connections).
This is often a “rich club” — a set of densely intra-connected nodes that “own” a lot of connectivity
A modular small-world network
with connector hubs

9. Cortical Networks “Rich club” nodes (van den Heuvel &
Sporns, J.Neuro. 2011)
Modules
(Hagmann, et al., 2008)
Hub nodes
(Hagmann, et al., 2008)

10. The Avian Connective Core
We find a connective core in the brains of birds as well as mammals
Here is the connectome of the pigeon forebrain
Connections funnel in to and fan out from five hub nodes
(Joint work with Vern Bingman, Onur Güntürkün, Toru Shimizu & Martin Wild)
The pigeon connectome
(Shanahan, et al., 2013)

11. The Connective Core as a Global Workspace=
Global
work-
-space
The connective core as locus of broadcast and competitive arena (Shanahan, 2010; 2012)
Modules, connector hubs,
and the connective core

12. Computer Models Global workspace Diagram adapted from
The scope of
Shanahan, Consc. Cog
The scope of Dehaene, et al. PNAS 2003
Diagram adapted from
Dehaene, et al. Trends Cog. Sci. 2006
Conscious
Preconscious
Subliminal

13. A Computer Model
This spiking neuron model includes a connective core and generates a sequence of reverberating broadcast states (Shanahan, Consc. Cog. 2008)
Global neuronal
workspace W1 W2 W3 W4 W5 C1 C2 C3
Cortical
modules
Workspace
nodes (hubs)

14. Coalitions and the core

15. Coalitions
Behaviour is generated by coalitions of coupled brain processes (sensory, motor, memory, affective)
Input Processes
Output Processes
Coupling
Coalitions are metastable. As time passes, coalitions form, then break up, then new coalitions form, and so on
It is also competitive. Only one coalition can eventually take charge of a shared resource (eg: hand position, gaze direction)

16. Neural Combinatorics
Sensory
Motor
Affective
Prefrontal
Hippocampal
``````
Connective
Core
The coalition formation problem is most challenging with a combinatorial repertoire of processes
The problem is easy when the situation calls for an established coalition (a reactive response)
The problem is harder when planning is needed
The problem is hardest when an innovative response is called for

17. The Connective Core Hypothesis
The blueprint for a cognitively capable brain includes a connective core (Shanahan, Phil. Trans. Roy. Soc. 2012)
The connective core is a limited capacity, highly connected communications infrastructure that provides
a locus of broadcast,
an arena for competition among coalitions of brain processes, and
a medium for coupling and communication between coalition members
As well as supporting a conscious / unconscious distinction, it promotes cognitive integration

18. The Value of Broadcast
Broadcast subserves open-ended, combinatorial coalition formation
A global communications infrastructure disseminates influence and information among otherwise independent processes
It allows each of a system’s components to influence and be influenced by the whole system
Such a system will exhibit dynamical complexity, a balance of integrated and segregated activity
As a consequence, it promotes a high level of information integration (Tononi, Biol. Bull. 2008)

19. The Value of a Bottleneck
The broadcast mechanism is a limited capacity “bottleneck”?
Only one coalition at a time can take charge of what the animal or robot does next
So coalition formation is competitive, and there can only be one winner at a time
The connective core enforces this winner-takes-all rule because only one coalition at a time can dominate it
This also enforces serial processing, which is vital for the sequential chaining of mental operations
The connective core is a bottleneck, in a good way

20. Unity from Multiplicity
Seriality and Unity
The connective core explains (the) two essential features of consciousness: seriality and unity
Information and influence funnel in to and fan out from the connective core (GNW), which acts as a limited bandwidth processing bottleneck, allowing for serial mental operations
Serial from Parallel
It also promotes integration across the brain, so that a coherent response to the ongoing situation can be orchestrated from its full resources
Unity from Multiplicity

21. The frame problem

22. A Brief History
McCarthy and Hayes discovered the frame problem in the 1960s when trying to describe the effects of actions in logic
After some teething troubles, ways were found to use non-monotonic formalisms to solve it in the 1990s
In the mean time, various philosophers saw a deeper issue in the frame problem
Dennett
Fodor
Dreyfus

23. The Original Frame Problem
How can the effects of actions be described in a logical formalism without having to write numerous formulae that describe the non-effects of actions
For example: scratching my head does not (typically) cause the walls to change colour
Solution
Assume things stay the same unless the logic dictates otherwise
Use a “non-monotonic” formulation to say this
For example: predicate completion (logic programming)

24. What the Philosophers Saw
Suppose we are building a robot that decides what to do on the basis of a representation of the world (classical AI)
How does a program maintain its model of the world, its set of beliefs, in the face of change? How does it work out what beliefs to revise without considering all of them? (Dennett)
More generally, how does a program determine the set of beliefs that are relevant to some ongoing cognitive function (eg: belief update, planning, analogical reasoning)? (Fodor)
Everything is potentially relevant to everything! (Dreyfus)

25. Innovative Behaviour
When cast as the problem of determining relevance, the philosopher’s frame problem isn’t just an artefact of classical AI / representation
There are a number of studies that show planning and / or apparent “insight” in corvids during tool-use
Here we see a crow who has bent a wire into a hook to retrieve a bucket from a tube (Weir, et al., 2002)
In such cases, how does the brain even entertain the necessary novel combination of actions?
Betty, celebrity corvid

26. Planning
Planning involves finding a sequence of actions to attain a goal
The sequence must be “novel”
Initial state
The sequence cannot be found by chance or trial-and-error
There must be multiple possibilities for action at each stage in the sequence
This entails combinatorial search
Goal state

27. Innovation and Insight
Insight: “the sudden production of a new adaptive response not arrived at by trial (and error) behaviour” (Thorpe, 1956)
“The spontaneous interconnection of two repertoires of behavior which [have] been established separately” (Epstein, 1985, p.627)
This entails a wide search tree, and the resulting combinatorics are daunting
Aha! 

28. Coalition combinatorics

29. Combinatorics and Relevance
“Everything is potentially relevant to everything”
Bending a piece of wire is relevant to obtaining food
How could a system search every combination of everything with everything, every possible coalition?
Of course, in classical AI we search through combinatorial spaces all the time — eg: SAT, constraint satisfation, planning
But in the most challenging form of the coalition formation problem, the constraints / goals are unknown
All the system can do is try out combinations (overtly or internally) and see what happens / see what reward it gets
Like using forward models rather than inverse models in kinematics

30. Two Stages Potential relevance Coalition formation
Processes that are potentially relevant to the current situation become active
All such processes are reached thanks to a broadcast mechanism
Coalition formation
Coalitions of jointly relevant processes form and compete for dominance
This is where the combinatorics kicks in, and where the right dynamics is needed

31. Peripheral Processes (Modules)
Potential Relevance 1
Fodor seems to have a strictly serial architecture in mind when they characterise the frame problem
Peripheral Processes (Modules) B C B C B C A D A D A D
“Is C relevant?
Yes!”
“Is A relevant?”
“Is B relevant?”
Time
Central Processes
This certainly looks computationally demanding

32. Potential Relevance 2
But global workspace architecture offers a parallel alternative
Parallel Specialists
“Am I relevant?
Yes!”
“Am I relevant?” B C
“Am I relevant?” B C
“Am I relevant?” A D D A
Time
Global Workspace
In the context of an appropriate parallel architecture, the frame problem looks more manageable (Shanahan & Baars, 2005)

33. Coalition Formation 1
State-to-state transitions result from parallel competitive attractor dynamics
Broadcast
Broadcast
Serial procession of broadcast states
punctuated by competition

34. Coalition Formation 2
So the difficulty of assessing potential overall relevance is diminished by broadcast
But how can previously unrelated processes come together in the same coalition?
How is joint relevance determined, enabling candidate processes to mutually enhance their overall relevance
We need a system with the right dynamics

35. Exploratory Dynamics
Suppose you could try (almost) every combination of everything with everything else
Perhaps this is possible in a dynamical system that is sufficiently “holistic”, where every process exerts a little tentative influence on every other process
Processes can then coerce each other into alliances
Process alliances would need to compete, jostling for dominance
Larger process alliances could take shape until one overall coalition dominates
What sort of dynamical system would support this?

36. Coalition dynamics

37. Synchronised Coalitions
According to the communication through coherence hypothesis (Fries, Trends Cog. Sci. 2005) synchronous activity allows the opening and closing of channels of communication between oscillating neuronal populations
Quiescent
Internally
desynchronised
synchronised, but
not synchronised
with any other
population
COALITION:
synchronised
internally and with
other populations

38. Metastability & Productivity
It’s possible for a system of coupled oscillators to produce a large repertoire of coalitions (Shanahan, Chaos 2010)
These are called metastable chimera states

39. Integration and Competition
In this system, any oscillator (brain process) can be recruited into a coalition with any other oscillator (brain process)
The system’s parts are influenced by the system as a whole, and the system as a whole aggregates the influence of all its parts (integration as well as segregation)
This is facilitated by the modular, small-world topology of the network
If there is also a connective core it acts as a limited capacity medium of coupling, so competition is ensured
So only one coalition can emerge as globally (albeit temporarily) dominant, the result of an implicit exploration of the combinatorial space of possible coalitions

40. Satisfiability Problems
In the oscillator system, every combination is just as likely to arise as every other. The system is not directed
What sort of dynamical system could explore a combinatorial space in a more directed fashion?
Let’s consider 3-variable SAT problems, the quitessence of combinatorial problem solving
Given a conjunctive normal form formula, such as
(A   B   C)  (B  C  D)
the task is to find an assignment of true or false to each variable such that the whole formula is true

41. Dynamical Systems for SAT
Ercsey-Ravasz & Toroczkai (2011) defined a system of differential equations that solves satisfiability problems
Each discrete variable in the SAT problem is mapped to a continuous variable in the dynamical system
Their system carries out a form of combinatorial optimisation through energy minimisation
Their approach is completely unfamiliar for computer scientists, but resembles the dynamics of the brain
It carries out a chaotic search with no backtracking, no stack, and no memory, using purely local computation

42. Combinatorial “Insight”
Here is such a system solving a 50-variable 3-SAT problem
(This is not exactly the system of Ercsey-Ravasz & Toroczkai, but a variation)

43. Conclusion
In the brain, the same architectural features that support the conscious / unconscious distinction also mitigate the frame problem
The brain uses a specific combination of
parallelism
connectivity
dynamics
Perhaps AI should replicate this architecture

44. References
Shanahan, M. & Baars, B,J. (2005). Applying global workspace to the frame problem. Cognition 98 (2), 157–176.
Shanahan, M. (2010). Embodiment and the Inner Life: Cognition and Consciousness in the Space of Possible Minds. Oxford University Press, 2010.
Shanahan, M. (2012). The brain’s connective core and its role in animal cognition. Phil. Trans. Roy. Soc. B 367,
Shanahan, M., Bingman, V.P., Shimizu, T., Wild, M. & Güntürkün, O. (2013). Large-scale network organisation in the avian forebrain: a connectivity matrix and theoretical analysis. Fronts. Comp. Neuro. 7, article 89.

45. More References
Shanahan, M. (2008). A spiking neuron model of cortical broadcast and competition. Consc. Cog. 17, 288–303.
Shanahan, M. (2010). Metastable chimera states in community-structured oscillator networks. Chaos 20,