M Systems Presented at CASYS

1
M SystemsPresented at CASYS 09

• Harry R. Erwin, PhD
• Hybrid Intelligent Systems Group
• Faculty of Applied Sciences
• University of Sunderland

2
The MiCRAM Project

• Midbrain Computational and Robotic Auditory
  Model for focussed hearing
• This is a collaborative study of the Inferior
  Colliculus by the Universities of Sunderland and
  Newcastle.
• Sunderland
• Dr. Harry Erwin (PI)
• Professor Stefan Wermter (co-PI)
• Dr. Mark Elshaw
• Dr. Jindong Liu
• Newcastle
• Dr. Adrian Rees (PI)
• Dr. David Perez-Gonzalez
• Funded by the UK Engineering and Physical
  Sciences Research Council.

3
M Systems

• The concept of an M-system has its origin in Don
  Griffins ideas about animal consciousness.
• An M-system is a system that maintains a model of
  its environment and uses that model to assess the
  current value of actions leading to future
  rewards and penalties.
• The model of the environment need only have
  sufficient detail to support action assessment.
• There are a number of possible ways that
  behaviour can be produced by an M-system.

4
The People Involved

• Adrian Rees (Newcastle PI)
• Jindong Liu (Sunderland Researcher)
• David Perez-Gonzalez (Newcastle Researcher)
• Stefan Wermter (Sunderland co-PI)

5
Realistic Modelling of Neural Systems

• Robots can be designed to emulate natural
  intelligence abstractly or in detail.
• In the MiCRAM project, we are simulating the
  detailed processing of the inferior colliculus
  (IC), a large neural module at the top of the
  auditory brainstem.
• This module seems to play a major role in
  localizing and classifying sound sources.

6
Categories of Behaviour

• Stimulus-response or reflexive
• May be learned or innate
• Habitual
• Action values are cached based on experience.
• Modelled well by actor-critic systems
• Goal-directed or goal-oriented
• The animal plans ahead to rewards and back
  propagates predicted rewardsas they changeto
  current action values. Much faster than
  real-time.
• If the reward value changes, behaviour may
  change.
• Seen in bats and rats.

7
Well, How Do Bats Capture Targets?

• Figure from Webster and Brazier, Experimental
  Studies on Target Detection, Evaluation and
  Interception by Echo-locating Bats, 1965.
• A bat (Myotis lucifugus) capturing a moth in
  foliage.
• 100 millisecond intervals.
• The bat had first detected the tree about 500
  milliseconds before the first image.
• Data available to the bata few biosonar
  snapshots in the dark.

8
A Simple Task Performed by Bats

• Handle non-stationary target acceleration,
  velocity, and position accurately enough to be
  able to approach a moving target within 5-10
  centimetres.
• Address asynchronous echo return timing (with
  inter-cry intervals ranging over 2-3 orders of
  magnitude).
• Predict forward over a variable time interval
  ranging up to a second.
• Observed to abandon target capture as late as 30
  msec prior to contact when the target is
  inedible.
• Strong evidence for goal-oriented behaviour in a
  small lissencephalic mammal (ca. 10 gr).

9
Modeling Results

• The performance seen in bats cannot be matched by
  optimized predictor-corrector algorithms.
• Algorithms that approach bat performance use
  target location collected over time to fit target
  motion models.

10
Where is this Time-Space Representation in the
Brain?

• New result Lubenov EV, Siapas AG (2009)
  Hippocampal theta oscillations are travelling
  waves. Nature 459 (7246)534-539, 28 May 2009.
• Our results demonstrate that theta oscillations
  pattern hippocampal activity not only in time,
  but also across anatomical space. The presence of
  travelling waves indicates that the instantaneous
  output of the hippocampus is topographically
  organized and represents a segment, rather than a
  point, of physical space.
• The hippocampus is already contains place cells.
• So at least one area of the brain contains a
  time-space representation of the animals
  environment.

11
What is the Mechanism?

• Spike Frequency Adaption (SFA) is believed to
  underlie theta wave generation (R. D Traub, et
  al., 1991 R. D Traub, et al., 1994 X-J Wang,
  2002).
• The CA3 neurones of the hippocampus seem to play
  an important role in this. These contain a number
  of specialised Ca channel types in their
  dendrites and soma that interact with
  Ca-activated K channels to produce SFA.
• SFA then produces neurone activation at specific
  phases of the theta wave.

12
Where else are these channels found?

• Inter alia
• Thalamic relay cells
• Inferior colliculus (IC) rebound cells

13
And What is the Inferior Colliculus (IC)?

• Largest auditory structure of the brainstem on
  the roof of the midbrain. A tectal structure
  behind the superior colliculus (SC).
• Primary point of convergence in the auditory
  brainstem. Sounds arrive here 2-5 msec after the
  inner hair cells are activated.
• Bidirectional connectivity with the auditory
  cortex (AC). This is fast enough to support
  cortically-controlled analysis of current sound
  afference.

14
Spiking Patterns seen in the IC

• Sustained-regular cellsseem to detect continuous
  sounds.
• Onset cellsseem to detect the beginning of
  sounds. Appear important in eliminating echoes.
• Pause-build cellsseem to play a role in delay
  sensitivity.
• Rebound cellsseem to play a role in echo-delay
  measurement, possibly in match-mismatch
  processing, and now possibly in generating a
  time-space representation of the auditory
  environment.
• (Classification by Sivaramakrishnan S, Oliver DL
  (2001) Distinct K Currents Result in
  Physiologically Distinct Cell Types in the
  Inferior Colliculus of Rat. Journal of
  Neuroscience 212861-2877.)

15
Sustained-Regular Pattern
SO
MiCRAM modelling
16
Onset
SO
MiCRAM modelling
17
Pause-Build
SO
MiCRAM modelling
18
Rebound
SO
MiCRAM modelling
19
Speculation

• The Inferior Colliculus contains a time-space
  map. This allows the cortex to do
  pattern-matching over
• Time
• Space
• Frequency

20
Implications

• Wherever the hippocampus does, the IC seems to do
  as well.
• In particular, it supports the monitoring of
  moving sound sources.
• This may be the reason the bat can react
  effectively to object target motion during the
  last few tens of msec prior to contact.

21
Open Questions

• Given how time-space is represented in the
  hippocampus (and potentially the inferior
  colliculus), how are plans represented?
• How are plans controlled over time?
• How are plans generated?

22
Acknowledgements

• Cynthia F. Moss, Ph.D., Department of Psychology,
  Program in Neuroscience and Cognitive Science,
  University of Maryland.
• John Murray, Stefan Wermter, Jindong Liu and the
  other members of the Hybrid Intelligent Systems
  Group, School of Computing and Technology,
  University of Sunderland.
• Adrian Rees and David Perez-Gonzalez, Institute
  of Neuroscience, The Medical School, Newcastle
  University
• The MiCRAM project is a collaboration between the
  Universities of Sunderland and Newcastle,
  supported by the EPSRC (ref EP/D055466/1)