LIFE INTELLIGENCE

1
LIFE INTELLIGENCE

• Three-fold definition of intelligence
• 1) Purpose is to solve problems to achieve a goal
• 2) Ability to store situational patterns
  categorized as good (helps to achieve goal),
  neutral or bad
• 3) An analytic mechanism to extrapolate past
  patterns to unknown, incomplete patterns in order
  to categorize those unknowns as good, neutral or
  bad as well

2
METAPHYSICS

• The metaphysics of life lie in what or who has
  determined lifes goal when solving problems

3
METAPHYSICS

• The goal of life is to perpetuate itself
  endlessly
• Hopefully, this goal will pop out of a final
  cybernetic theory of life naturally and
  unexpectedly, much as electron spin arose from
  Diracs equation

4
METAPHYSICS

• Or the speed of light out of Maxwells equations

5
Intelligence has many faces
6
MAIN POINTS OF THE BOOK

• Biological intelligence is the product of network
  architecture using energy-relaxation computation
  in thermodynamic landscapes
• Life is merely the implementation of
  intelligence at the molecular level
• Biological intelligence is scale-invariant ---
  the same principles apply to intelligence at the
  molecular level, the organismal level, the
  societal level and the global level the nervous
  system and immune systems, an ant hill, the
  cytoplasm and the biosphere all use comparable
  architecture
• Evolution is the act of network learning using
  natural selection as a learning rule
• The world has been indeed crafted by Intelligent
  Design, but the Designer is life itself
• Because of feedback manipulation of
  selection/mutational mechanisms, evolution is a
  nonlinear process which has itself evolved

7
Origin of Life

• Networks, indeed any computational device,
  require nonlinear, quasi-stable components with
  at least two states
• The first computers used relays, then vacuum
  tubes, then transistors
• The nervous system uses neurons, the immune
  system lymphocytes, which both have quasi-stable
  states of activation/non-activation
• Molecular computation requires chemical reactions
  with transistor-like properties
• A few inorganic reactions have such properties
• However, enzyme-catalyzed reactions can have the
  bistable behavior of transistor switches

8
Bistable reactions

• Feedback control can induce switch behavior
• Some enzyme systems may allow switch behavior
  without feedback
• See Craciun et al PNAS 1038697-702 2006

9
Bistable reactions

• Reactions that admit two steady states for the
  same enzyme concentration/reaction kinetics
• Once thought to require feedback
• Now, even simple non-feedback enzyme reactions
  may be bistable

10
Bistable reactions

• PNAS 2006 paper shows that classes of single
  reactions, without feedback can have two steady
  states switchable by minor variations of reactant
  flux provided they have certain (severe)
  restrictions on reaction parameters
• Authors show that dihydrofolate reductase system
  meets the restrictions

11
(No Transcript)
12
Origin of Life

• Thus, the first molecular organic computers arose
  with the advent of enzymes
• Enzyme catalysis creates the equilibrium
  multi-stability that enables chemical systems to
  perform calculations
• Nucleic acids are simply the storage devices,
  computation is an enzymatic process and could
  arise historically before the advent of genetic
  machinery

13
Brain Chauvinism

• We have a natural tendency to consider
  intelligence a neurological process dependent
  upon electrical networks
• This also biases us to consider intelligence as
  the act of manipulating symbols and numbers, and
  to equate genius with the creation of complex
  macroscopic patterns of objects, sounds or visual
  patterns (automobiles, houses, symphonies, plays,
  sonnets, mathematical equations)

14
Brain Chauvinism

• The immune systems crafting of a precise
  stereochemical mold of an antigen is not
  considered an act of genius

15
Brain Chauvinism

• Nor is the fact that an ant colony can learn
  where food supplies are and which surrounding
  colonies are most aggressive, data learned over
  months and years (even though each ant lives only
  a few weeks)

16
Brain Chauvinism

• Nor is the fact that bacteria have met every
  challenge we throw at them, or the fact that
  cancer cells defeat every method we can devise to
  kill them

17
Brain Chauvinism

• We consider the amoeba to be a lowly
  unintelligent beast, but, in fact, it is
  genetically and chemically as sophisticated, if
  not more so, than the ova that make geniuses
  like ourselves

18
Brain Chauvinism

• We must not forget that intelligence is not a
  divine entity but a survival tool
• As such, the measure of intelligence is survival
• Until we have been around as long as the fly, we
  cant claim our species is more intelligent!

19
Networks

• Processing units with nonlinear I/O behavior
• Flexible, scaled connections
• Learning rule
• Thermodynamic form of learning/recall

20
Networks
21
Networks

• With the advent of enzyme catalysts, many
  organic reaction could have bi-stable behavior
  capable of organizing into cognitive networks

22
Networks

• We enter into the realm of fluid phase networks,
  where the connections are not hardwired but,
  instead, consist of interaction cross-sections
  of mobile units that are thermodynamically mixed
• I refer to these as party networks in analogy
  to cocktail parties, where mobile party-goers mix
  and connect according to mutual affinities
  (interest in sports, political affiliations, etc)

23
Networks

• Examples of connection weights
• Protein-protein affinities
• Shared pool of reactants
• In ecological networks, shared resources or a
  predator/prey affinity
• In social networks, common interests

24
Scale-free networks

• Scale-free --- the network obeys a power-law
  distribution of connectivity, usually with a
  power of 2-3 (as opposed to Poisson distribution
  in random networks)
• Network architecture is independent of scale
• Depends on a small number of hubs
• Describes most organic networks, including the
  Internet, cytoplasmic webs, protein folding,
  scientific citation networks, phone networks,
  food webs and power grids

25
Scale-free networks
26
Scale-free networks

• Simplest evolutionary model assumes linear growth
  and a linear preferential attachment rule,
  i.e., as each node is added, the probability that
  a given node will add a connection is
  proportional to its current connectivity
• Will yield SF network with power 3
• More complex models have been developed

27
Scale-free networks

• From whence the preferential attachment
  rule?????????????

28
Scale-free networks
29
Scale-free networks

• Topological models dont yet adequately
  incorporate the most important aspect of
  networks learning, i.e., dynamic weighting of
  connections
• The assumption is that a) all connections are
  either 1 or 0 and b) the connections are fixed in
  time, i.e., what is a hub now will be a hub
  forever

30
Learning

• Biological networks learn by modifying
  connectivity
• Such learning can be acute (modification of
  synapses or enzymatic switches in seconds or
  minutes), subacute (alteration of predator-prey
  balances over years or centuries) or chronic
  (modification of protein structure via Darwinian
  evolution over thousands of years)

31
Which is better?????
32
Learning rules

• Learning rules alter the connectivity of
  ensembles in response to external training
• Is Darwinian selection a learning rule????

33
Learning rules

• Hebbian learning
• Example for given learning rate n, change in
  weight is proportional to output y with a y2
  decay component to prevent unbound growth

34
Relaxation computation

• Learning crafts an energy landscape
• If the system is initialized near an energy
  minima, system relaxes into solution state
• Thermodynamic representation of pattern completion

35
Annealing

• Annealing heats, or shakes the network to
  bounce it out of local minima which reflect
  imperfect solutions
• A global temperature T can be considered a source
  of system noise that can be raised or lowered to
  optimize network behavior

36
Noise

• In biological networks, some form of noise is
  essential to allow learning and permit annealing
  of established networks
• Moreover, the noise level should be under system
  control
• Cells can regulate their mutation rates by
  altering rate of DNA repair
• Brains can vary their signal to noise ratio

37
Noise

• The feedback control of mutation rate, however,
  implies that evolution has become a nonlinear
  process
• Evolution has itself evolved

38
Evolutionary computation

• 1) Genetic Algorithms
• 2) Ant Colony Optimization
• 3) Swarm intelligence
• 4) Bacteriologic Computation models

39
Evolutionary computation

• GA --- uses fitness landscape, mutations and
  natural selection (prone to local minima and
  mutation rate limits dealing with dynamic data)
• Swarm intelligence --- more local than global
  optimization, less prone to local minima
• ACO --- good with dynamic data

40
Bio-inspired computing

• 1) Use of smaller, repetitive subunits linked in
  fluid or hardwired ensembles
• 2) Pattern-recognition dominates
• 3) Rapid good solutions generally better than
  slow, perfect solutions
• 4) Systems must be robust and expandable
• 5) Learning is typically self-organizing or
  unsupervised

41
Evolutionary computation

• Take home message--- biological ensembles are
  massive, complex, likely use a variety of
  computational schemes and are far more
  intelligent, even at the molecular level, than we
  give them credit for

42
WHICH IS SMARTER???
43
WHICH IS THE ACT OF GENIUS??
44
Noise

• There are some interesting ideas about the
  regulation of noise and mental illness
• Is a schizophrenic brain annealing itself into
  oblivion?
• Is dopamine the regulator of neural network
  temperature?

45
Future directions

• Large antigen pattern completion
• Learning rules in molecular networks
• Enzyme kinetics --- which systems are bistable
• Supercomputer modeling of Darwinian eco-networks,
  including Gaian models that incorporated
  inorganic nodes
• More theoretical work in fluid-phase networks
• Quantitative analysis of social insect colonies
• Signal/noise behavior in normal and diseased
  brains

46
(No Transcript)