Large-scale projects to build artificial brains: ABACCUS.

1
Large-scale projects to build artificial brains
ABACCUS.

• Wlodzislaw Duch (Google Duch)
• Department of Informatics,
• Nicolaus Copernicus University, Torun, Poland
• School of Computer Engineering,
• Nanyang Technological University (NTU),
• Singapore

Building Artificial Brain workshop after ICANN
2005, Sept 15, 2005
2
Attention-Based Artificial Cognitive Control
Understanding System (ABACCUS)

• Large EU integrated project (gt150 pp), with 9
  participants
• Kings College London (John G. Taylor,
  coordinator), UK
• Centre for Brain Cognitive Development,
  Berkbeck College, University of London, UK
• Cognition and Brain Sciences Unit, Medical
  Research Council, UK
• Robotics and Embedded Systems, Technical
  University of Munich, G
• Institute of Neurophysiology and Pathophysiology,
  Universitätsklinikum Hamburg-Eppendorf, G
• Institute of Computer Science, Foundation for
  Research and Technology Hellas, Heraklion,
  Crete, GR
• National Center for Scientific Research
  Demokritos, Athens, GR
• Dipartimento di Informatica, Sistemistica,
  Telematica, Universita di Genova, I
• Dep. of Informatics, Nicholaus Copernicus
  University, Torun, PL
• Now changed to Brain as Complex System (BRACS)

3
ABACCUS Goals

• Assumption gross neuroanatomical brain structure
  is critical for its function, therefore it should
  be preserved.
• To demonstrate how fusion of the appropriate
  brain-based models, guided by the overall
  architecture of the brain, and by its
  developmental learning stages, can help attain
  high-level cognitive processing capabilities.
• Show basic language understanding and reasoning
  abilities for direct human-machine communication,
  at the level of a pre-school child, mimicking
  solutions used by the human brain.
• Develop an attention control systems for focusing
  in sensory surveillance tasks, and for image
  searching.
• Development of control structures for autonomous
  machines.
• Create its own goals in an autonomous fashion.
• Founded on neuro-scientific understanding of
  attention and the sensory and motor systems it
  controls, development in children, simplified
  modeling, computer power.

4
Sketch of the ABACCUS system

• Rough sketch of the ABACCUS system, based on
  simplified spiking neurons.

5
Primary objective 1

• To develop linguistic powers of ABACCUS system.
• Use training of single words by associating their
  representations to internal representations of
  objects and actions.
• Use pair-wise associations to learn word pairs
  (like kick ball), extend syntactically and
  functionally the use of function words.
• Working memory modules, with associated
  phonological coding, will be created and fused in
  the language component, both for speech
  understanding and generation.
• Extension for abstract concepts by tagging the
  associated words to clusters of concrete
  object/action representations.

6
Primary objective 2

• To create a high-level cognitive system, able to
  solve problems requiring reasoning, thinking,
  imagination and creativity.
• Based on the basic control concept of a forward
  model, acting as a predictor and working under
  attention control.
• Forward models include various semantic and goal
  networks.
• Sequences of activations of representations will
  be learnt and thereby used to achieve goals.
• Various of these forward models will be present
  and branch into each other during the running
  process by means of lateral connections.
• New routes will occur allowing complex goals to
  be achieved, or even new, previously unrecognized
  goals, to be arrived at (required for creativity).

7
Example of action internal models

• Rough sketch of the ACTION subnetwork.
• Sensory information enters the supplementary
  motor area (SMA)
• (1) cortico-thalamo-cortical (the short loop)
• (2) cortico-striatal-GPi-thalamo-cortical (the
  long loop)
• (3) cortico-(STN?GPe)-GPi-thalamo-cortical
  (first indirect loop)
• (4) cortico-striatal-(STN?GPe)-GPi-thalamo-cortica
  l (the second indirect loop).

8
Primary objective 3

• Extract a simplified architecture for the
  attention control system.
• Should involve both sensory and motor control,
  especially joint sensory-motor control, systems
  whose creation is guided by neuroscientific
  knowledge about the brain.
• Approach based on infant development to train
  ABACCUS incrementally on the computer platform.
• Use neuroscientific data to guide the
  architecture of a large-scale neural simulation
  of the relevant components of the human brain.
• Use feed-forward/feedback training simulations
  based on the simplified brain architecture
  following developmental processes, using the
  sensor and response signals of the robotic
  embodiment.

9
Primary objective 4

• Developing the ability to learn novel objects,
  both as stimuli and as associated reward values.
  
• Use reward/penalty feedback training with
  attention control, with associated value maps
  constructed to learn to encode the values of
  stimuli and responses to them in the environment.
  
• Temporal sequence or schemata (as object/action
  sequences as well as the associated rewards
  involved) will be constructed to function as
  predictors and so support reasoning, along with
  automatic response learning.
• Neural codes of visual, auditory and tactile
  concepts will be learnt in a feedforward manner,
  without attention, as will codes for motor
  responses controlling the embodied robot.

10
Primary objective 5

• Develop a robotic embodiment this involves
• Specification of robots available for ABACCUS
  control.
• Development of sensor and response systems in the
  robot.
• Sensory data mappings to the central
  computational platform.
• An environmental task domain in which ABACCUS can
  develop its concepts and cognitive powers.
• Simulation platform for the developmental
  training (on a Beowulf cluster or similar
  powerful computer system).
• Specifications for integration of various models
  on the central computational platform.
• Compatibility of computational languages used by
  partners.

11
Attention Control

• Attention control systems will be achieved by use
  of the early flow of activity from low-level
  cortex rapidly to prefrontal sites.
• This initially uncoded activity will be used to
  bias feedback activity sent to parietal and/or
  temporal lobe, so as to bias the attention
  feedback from there down to the activated inputs
  encoded in the object representations in the
  temporal lobe.
• This feedback loop will be trained by causal
  Hebbian learning, with the attention feedback
  represented by contrast gain amplification, so by
  sigma-pi network quadratic nodes, onto neurons in
  hierarchically lower modules the goal
  representations will be learnt as part of this
  processing.
• The manner in which a dual-control stereo camera
  can be used most effectively will be explored by
  modeling the manner in which eye and attention
  control are separable.

12
Drives, Value Maps Emotions

• Drives and reflexes will be enumerated and
  modeled as separate systems to be used as a
  launching pad for more controlled developmental
  sensory representations and responses.
• The rewards system will also be developed as a
  separate component, based on the learning of
  reward/error prediction.
• ABACCUS will be trained by reward error to
  develop value maps for objects and responses as
  goals, to bias learning of feedback attention to
  filter out responses in complex environments. The
  effect of the reward signal will be checked to
  determine if there is any modification of the
  earlier goal representations learnt under
  attention control.
• Value maps associated to rewards attached to
  sequences of stimuli (schemata) will give
  emotional bias in the cognitive decision-making
  it can perform.

13
Short-term working memory

• Two sorts of working memory are presently known
  buffer or slave working memories, of a fixed and
  limited persistence, and longer lasting
  prefrontal activities of duration determined by
  the importance and difficulty of the goal they
  represent.
• Buffer memory is fed by the semantic
  representations, whose activity they extend, and
  thereby allow the development of attended state
  estimators in the various modalities for more
  general use across other cortical regions,
  especially goal sites.
• Buffer sites will be modeled by suitable
  recurrent circuits coupled to the semantic maps
  from which they receive input.
• Rehearsal of the working memory will be extended
  to various transformations that are also known to
  occur in goal sites, such as deletion, comparison
  and related functions.

14
Concept Understanding and Abstraction

• Highlevel concepts will be created by use of a
  hierarchy of networks, each being more specific
  in its features to which it is responsive than
  the next lower one.
• The understanding based on such ontology allows
  to understand various levels of concepts, by
  the associated super- or sub-ordinate concept
  levels activated by trained connectivity at the
  levels of specificity allowed by the number of
  hierarchical layers.
• Abstraction will arise through creation of
  specific connections of a new word
  representation, describing the concept, to its
  various more concrete word and object/action
  representatives.
• Word representations will be linked to emotion
  state descriptors, that are connected to the
  associated reward value map representations of
  the associated system state. For example the word
  anger would be associated with the underlying
  state arising from blockage, by the environment,
  of the gaining of a goal.

15
Basic tasks

• 2D domain, enclosed within four walls with
  obstacles on the floor and 2D objects on the
  walls, of varying shapes (square, triangle,
  circle) and colors (red, green, blue), emitting
  various tones. Objects are potential goals, to be
  learnt and responded to.
• Perceptions/Concepts To develop internal
  representations of the 2D objects, based on
  shape/color feature maps created by learning,
  after relevant clustering on the feature map
  activations, to learn a representation for the
  objects.
• Actions/Goals learn to touch/manipulate the 2D
  objects based on stimulus-action rewards and
  fused representation as a goal to attentionally
  bias lower order perceptual representations just
  learnt (using dopamine-based reward learning).

16
Intermediate tasks

• 2D domain with walls, with 3D objects as
  potential goals objects with simple shapes
  (sphere, box, ring), colors, emitting various
  sounds, some in motion (to help develop
  tracking).
• Perception/Concepts/Goals To develop internal
  representations of the 3D objects, based on a set
  of their 2D views, clustered over time (by use of
  binocular representations of depth). Certain
  examples of moving stimuli will be created so as
  to be learnt.
• Actions/rewards Motor responses to move specific
  objects to certain places, such as to stack an
  object on top of another, or place one object
  inside another (object fusion observed developing
  in both chimps and human infants).
• Language Word maps for actions and objects
  concepts will be created, using associations of
  the word maps to the relevant object maps
  attention will be important in this learning
  process.

17
Advanced task domain

• Perception/Concepts/Goals complex 3D objects
  present with various features in various
  modalities, including animal shapes emitting
  sounds and smells. The internal concept
  representations will be extended to hierarchies
  of concepts, so leading to abstraction processes
  (where more abstract, basic level concepts are
  activated, such as that for animal by specific
  animal inputs, so leading to spread of activity
  to related concepts and to associated reasoning
  potentialities).
• Actions with added complex tasks requiring
  reasoning to achieve fast/accurate responses
  manipulating objects, putting objects into
  slots/matching shapes, Tower of Hanoi (involving
  virtual movements of objects translations and
  stacking at a fixed set of different sites),
  Incident Detection, etc. AI-based approaches to
  such tasks exist (cf. SOAR), but ABACCUS will be
  based on the flexible posterior representation
  transformation, goal creating and resolving
  powers of a neural system.

18
More on advanced task domain

• Language powers will be expanded so that ABACCUS
  can be talked to, using limited length word
  strings (initially of length two or three). These
  strings will be understood in terms of the
  associated object/action maps and responses made
  by the system, either as actions or action
  sequences or suitably meaningful speech
  responses. Reasoning and creation of new types of
  responses will be done by use of lateral
  spreading in memory to achieve new trajectories
  in the schemata space.
• The Incident Detection Tracking Scenario
  involves a robot moving in the internal
  environment, although with novel stimuli and
  obstacle layout. This includes ramps and
  stationary or slow moving obstacles. Candidate
  events of interest in this paradigm are
• a) Moving entity detection and classification
  (humans, animals).
• b) Static environment variation detection
  (unexpected objects).
• c) Abnormal auditory events.
• d) Abnormal visual events.

19
Work packages

• Summary of ABACCUS workpackages.

20
Summary

• ABACCUS is the most ambitious project formulated
  so far, an attempt to simulate functions of most
  brain structures.
• High order functions (including conscious-like
  behavior?) should result from elementary
  interactions in a system with proper brain-like
  architecture.
• Neuroscience and developmental science integrated
  in one computational model, useful for behavioral
  experiments and computational psychiatry.
• Wish us good luck ...