From Brain Operating Principles to Computer Technology

1
From Brain Operating Principles to Computer
Technology

• Michael A. ArbibComputer Science, Neuroscience
  and the USC Brain ProjectUniversity of Southern
  CaliforniaLos Angeles, CA 90089-2520arbib_at_pollux
  .usc.edu

2
Rounds of Neural Computing

• Round 1
• Hebbian adaptation and Perceptrons adaptation
  and self-organization in neural networks
• Reinvigorated in the 1980s as work on
  reinforcement learning and backpropagation
  extended earlier insights.
• Round 2
• Compartmental modeling of the neuron
• Mead (1989) built on his earlier work on digital
  VLSI (Mead and Conway) to show how to exploit
  neuromorphic function in highly parallel analog
  VLSI

3
A Prospectus for Round 3

• Analyzing the architecture of the primate brain
• to extract neural information processing
  principles and
• translate them into biologically-inspired
  operating systems and computer architectures
• interplay between feedforward and feedback
  pathways
• sharing of neural resources between perception
  and action
• the role of plasticity in sensory, motor and
  central processing
• A Contrast
• Rounds 1 and 2 are based on brain capabilities
  that come from components of single neurons
  (synaptic plasticity and dendritic tree
  complexity, respectively)
• Work on Round 3 will develop the theme that many
  of the most interesting capabilities of the brain
  result not just from the individual component
  mechanisms, but from large scale organization as
  well

4
A Case Study Cerebellum
5
The Cerebellar Module The Microcomplex
The Microcomplex a patch of cerebellar cortex
and the nuclear cells it inhibits modulating the
activity of one MPG

• Schematic of our cerebellar model
• Inputs arrive via mossy fibers (MF)
• nuclear cells (NUC) generate output
• training signals are carried by climbing fibers
  (CF) from the inferior olive (IO) which depress
  the strength of PF?PC synapses
• (plus further subtleties!!)

6
The Role of the CerebellumModulation and
coordination of MPGs is also critical for motor
skill learning

• Hypothesis The cerebellum adjust the parameters
  of MPGs
• To tune MPGs adjusting metrics within a
  movement
• To coordinate MPGs grading the coordination
  between motor components

Plasticity within this system provides subtle
parameter adjustment dependent on an immense
wealth of context. In most cases, the tuning
often depends crucially on the uniquely rich
combinatorics of mossy fibers and granule cells,
and so cannot be replaced by processing in other
regions.
7
Lessons from the Cerebellum

• This brief review of the cerebellum shows three
  things
• The special type of learning involved ? learning
  how to reduce errors by adjusting the inhibitory
  sculpting to apply in different contexts
• The immense subtlety of individual neurons,
• The way these details are all embedded within a
  high-level architecture.
• Points (1), (2) and (3) correspond to what I call
  Rounds 1, 2 and 3 of neural computing.

8
A Few Brain Operating Principles

• Winner-Take-All
• Extensively used in several models
• Dynamic re-mapping
• Double saccade
• Path integration for locomotion
• Reinforcement learning
• Actor-critic model
• Hierarchical Reinforcement learning
• Competitive Queuing
• Attention control the combination of a
  saliency map with an inhibition of return
  mechanism forms the basic mechanism for
  controlling attention deployment in contemporary
  computational models of focal visual attention
• Parallel recall from long term memory

9
What is a Database in the Brain?

• Classic Database Style
• Maintain (perhaps in a federation of databases)
  a coherent set of correct up-to-date master data
• Provide a set of Views customized to
  different users
• Passive data with external inference engines
• The Brains Database Style is Cooperative
  Computation
• Maintain different views of the data as
  separate entities each separately updatable by
  experience
• Coordinate views (more or less) as they are
  dynamically integrated for action in novel
  situations
• Active schemas which integrate data and the
  processes for deploying them

10
The Claim Brain Operating Principles have much
to offer for future Computer Technology

• But theres a paradox
• millions or billions
• 15 to 25
• 8 plus 2
• six billion

11
Point and Counterpoint

• I argue
• for the promise for computer science of
  developing an explicit formulation of the brains
  approach to reusable computing by adding
  evolutionary refinements to augment available
  circuitry to handle new tasks
• that what is known about the organization and
  architecture of these capabilities is also
  critical to the development of a new approach to
  computer architecture and operating systems.
• However, new architectural developments will
  include, but not be restricted by, biological
  principles
• Example the inclusion of a non-biological
  reflection technology will allow the re-use of
  biological computing strategies in a way that in
  biology is available only on an evolutionary time
  scale.

12
Programmable, Reflective Self-Organization

• The goal Extending brain-style computing by
  augmenting self-organization with
  wrappings-based programming to mobilize
  resources for each new problem
• Seeking to exploit an understanding of how the
  brain marshals the specialized capabilities of
  different subsystems such as
• multiple levels of sensory analysis and
  integration
• declarative and episodic memory
• planning and motor control
• emotion and social interaction
• language and other communication interfaces
• Issue How can we have the wrappings/high-level
  specifications (the essence of a reflective
  architecture) keep track of the distributed
  self-organization of successful systems so that
  emergent resources can be recognized as providing
  approximate solutions to subproblems?
• Aim To have novel problems programmed by
  negotiating assemblages of resources

13
Focusing on the Mirror

• We now consider a dramatic pattern of "re-use" in
  a neural architecture, focused on a whole
  progression of neural systems concerned with
  behaviors ranging from
• the visual control of grasping to
• the mirror system action recognition and even
• the mirror system hypothesis human language
• Prior and continuing research
• modeling the primate mirror system
• extending the Mirror System Hypothesis
• Round 3 of Neural Computation
• Studying how sensory, planning and executive
  stages of neural processing converge in a
  flexible manner to yield a very powerful
  integrated system
• Building on this to translate high-level neural
  computation principles into new computer systems
  and architectures.

14
Visual Control of Grasping
A recurring theme of mammalian brain
design parietal affordances are coupled to
frontal motor schemas
AIP - grasp affordances in parietal cortex Hideo
Sakata
F5 - grasp commands in premotor cortex Giacomo
Rizzolatti
15
A Mirror Neuron
Rizzolatti, Fadiga, Gallese, and Fogassi, 1995
Premotor cortex and the recognition of motor
actions This neuron is active for both execution
and observation of a precision pinch
16
A New Approach to the Evolution of Human Language
Homology
Monkey Not to scale Human

• Monkey F5 is homologous to human Brocas area
• Rizzolatti, G, and Arbib, M.A., 1998, Language
  Within Our Grasp, Trends in Neuroscience,
  21(5)188-194.
• Thus suggest an evolutionary basis for language
  parity
• rooting speech in communication based on manual
  gesture.

17
The Room-Brain is Sometimes Like a Colony of
Reassignable Brains
18
Tracking

• Bottom-Up Attention ? Target acquisition ?
  saccades
• Top-Down Attention ? Locating a designated target
• Fovea versus Periphery
• Retinal coordinates ? Other reference frames
• Sensor Fusion Cues from different sensor sets
• Smooth Pursuit
• Social Invoking extra cameras for different
  views
• Generalizing the BOPs
• Going from 2 eyes to n cameras
•  
• What happens when you need to keep track of more
  objects and agents than you have cameras?

19
A Room with a ViewAllocentric and Egocentric
Coordinates
Room Strategy
Body is fixed, its circum global
Reach for what the eye is looking at
Sensors may (a) move (b) form transientcoalitions
. BOP Sensor fusion Sensor data must be linked
to room coordinates and/or effector coordinates
Deploy locomotion as needed.
20
An Invitation

• To learn more about this subject, take
• CS564 Brain Theory and Artificial Intelligence
• next Fall.
• . and read the book!!