This is the presentation I gave at AGI-12 (also called the Winter Intelligence 2012 conferece) in Oxford, UK, on Dec.11, 2012. There is an AGI-12 proceedings paper that accompanies this talk. I will make that available on my publications page at http://randalkoene.com and I will put both together on the http://carboncopies.org page about this event. The video (recorded by Adam Ford) should also appear soon.
Abstract. Feasible and practical routes to Artificial General Intelligence involve short-cuts tailored to environments and challenges. A prime example of a system with built-in short-cuts is the human brain. Deriving from the brain the functioning system that implements intelligence and generality at the level of neurophysiology is interesting for many reasons, but also poses a set of specific challenges. Representations and models demand that we pick a constrained set of signals and behaviors of interest. The systematic and iterative process of model building involves what is known as System Identification, which is made feasible by decomposing the overall problem into a collection of smaller System Identification problems. There is a roadmap to tackle that includes structural scanning (a way to obtain the “connectome”) as well as new tools for functional recording. We examine the scale of the endeavor, and the many challenges that remain, as we consider specific approaches to System Identification in neural circuitry.
Toward Tractable AGI: Challenges for System Identification in Neural Circuitry
1. Toward Tractable AGI: Challenges for
System Identification in Neural Circuitry
Randal A. Koene
Carboncopies.org & NeuraLink Co.
Interfaces Reconstruction
prostheses project
Special Specific System
tools Identification Problems
AGI-12, Winter Intelligence Conference 2012, Oxford
2. Tractable AGI through System Identification in
Neural Circuitry
Representations and Models
Behavior of interest; Signals of interest; Discovering the transfer
function
Mental Processes and Neural Circuitry: Brain
Emulation
System identification in neural circuitry
Simplification of an Intractable System into a
Collection of System Identification Problems
Tools for structural decomposition; Data from structure;
Parameter tuning among connected systems; Tools for
characteristic reference recordings
Challenges
Signals and predicting spikes; Validation, reconstruction and
plasticity; Interference during measurement; Data quantities;
Proof of concept
3. Tractable AGI
Challenges in our
environment
Theoretically sound
AGI
Practical feasibility
Short-cuts
Brain-like AGI
Legg, Hutter
Optimal boundedlength
spacetime embedded
agent Orseau
Abstraction level special case:
Our reverse interests: Taking a niche system
Neuronal circuitry/physiology
and making it more adaptable, more general
(100 years of grounding)
4. Tractable AGI through System Identification in
Neural Circuitry
Representations and Models
Behavior of interest; Signals of interest; Discovering the transfer
function
Mental Processes and Neural Circuitry: Brain
Emulation
System identification in neural circuitry
Simplification of an Intractable System into a
Collection of System Identification Problems
Tools for structural decomposition; Data from structure;
Parameter tuning among connected systems; Tools for
characteristic reference recordings
Challenges
Signals and predicting spikes; Validation, reconstruction and
plasticity; Interference during measurement; Data quantities;
Proof of concept
5. Representations and Models
Modern science:
Observed effects
described as
Model (testing)
improve
Understanding
Pieces of natural
environment
Not independent!
Signals, information, P(x)
May seem obvious to comp.neurophys. modelers... but consider whole problem not only typical solutions
6. Behavior of Interest
Lots of pieces
Systematic modeling
Keep it simple
Interesting effect
Focus, constrain
scope/model
Neuroscience:
Effect = Behavior
(e.g. object recognition)
7. Signals of Interest
How do pieces
communicate?
Signals
Physics: 4 interactions
(gravity, electromagnetism, weak
& strong nuclear force)
Constrain
Neurons: current,
temperature, pressure, EM, etc…
Priority of interest:
empirical (noise,
predictive value)
8. Discovering the Transfer
Function
SI in Control Theory:
Black/gray box
State, input, output
Find: Transfer
Function
Formal methods
E.g. Volterra series
expansion
kernels & history of input
9. Tractable AGI through System Identification in
Neural Circuitry
Representations and Models
Behavior of interest; Signals of interest; Discovering the transfer
function
Mental Processes and Neural Circuitry: Brain
Emulation
System identification in neural circuitry
Simplification of an Intractable System into a
Collection of System Identification Problems
Tools for structural decomposition; Data from structure;
Parameter tuning among connected systems; Tools for
characteristic reference recordings
Challenges
Signals and predicting spikes; Validation, reconstruction and
plasticity; Interference during measurement; Data quantities;
Proof of concept
10. Mental Processes and Neural
Circuitry: Brain Emulation
Effects =
Experiences
Perception, learning &
memory, goal directed
decision making, emotional
responses, consciousness,
language, motor
Observable / internal
Involves ensembles of
neurons in a circuit
layout
12. System Identification in Neural
Circuitry
Signals of interest
Chip – “bits”
Initial assumptions,
reliable neural
communication
Sensory, muscle, learning
– “spikes”
Example methods:
Berger chip Aurel A. Lazar: Channel Identification Machines
(Volterra exp.)
(CNS2012 workshop on SI)
13. Tractable AGI through System Identification in
Neural Circuitry
Representations and Models
Behavior of interest; Signals of interest; Discovering the transfer
function
Mental Processes and Neural Circuitry: Brain
Emulation
System identification in neural circuitry
Simplification of an Intractable System into a
Collection of System Identification Problems
Tools for structural decomposition; Data from structure;
Parameter tuning among connected systems; Tools for
characteristic reference recordings
Challenges
Signals and predicting spikes; Validation, reconstruction and
plasticity; Interference during measurement; Data quantities;
Proof of concept
14. Simplification of an Intractable System into a
Collection of System Identification Problems
SI of observable + internal =
intractable if black box is brain
Many communicating black boxes
with accessible I/O
Communication = note locations,
trace connectivity (“Connectomics”)
E.g. compartmental modeling
Briggman et al & Bock et al, Nature 2011
Characteristic responses
ADP, AHP, AMPA, s/fNMDA
15. Whole Brain Emulation: A Roadmap to data
acquisition & representation
Characterize the parts
Break into parts.
How can the parts communicate? Platform suiting
representation
Iterating
Improving the
Other pillars
(See earlier presentations, carboncopies.org.
More about roadmap & projects – leading up
To Global Future 2045 Congress NYC, June 15-16.)
16. Tools for Structural
Decomposition
Voxel geometric
decomposition (e.g. MRI)
Cell body locations &
functional connectivity
Zador RNA/DNA tags
Stacks of EM images
(Denk, Hayworth, Lichtman)
17. Data from Structure
SI for compartments
Electric circuit analogy
3D shape
Conductance, class, etc.
“Invisible” parameters?
Measurement reliability?
18. Parameter Tuning among Connected Systems:
Reference Points
Parameters –
sensible collective
behavior
Reference points:
constrain & validate
Resolution of
reference points –
combinatorial size of
SI problem
# and duration of
(purposely abstract: measurements
- resolutions reference/SI decomposition
- not one path (e.g. Briggman et al.
= problem specific criteria, not method specific – compare & collaborate projects)
19. Tools for Characteristic
Reference Recordings
Large arrays of
recording electrodes
+ optogenetic
selectivity
Microscopic wireless
probes
Molecular “ticker-
tape” by DNA
amplification
20. Tractable AGI through System Identification in
Neural Circuitry
Representations and Models
Behavior of interest; Signals of interest; Discovering the transfer
function
Mental Processes and Neural Circuitry: Brain
Emulation
System identification in neural circuitry
Simplification of an Intractable System into a
Collection of System Identification Problems
Tools for structural decomposition; Data from structure;
Parameter tuning among connected systems; Tools for
characteristic reference recordings
Challenges
Signals and predicting spikes; Validation, reconstruction and
plasticity; Interference during measurement; Data quantities;
Proof of concept
21. Challenges
General SI problems
Particular to neurons &
neural models
Unique to pieces of
neural tissue & large
neuronal circuits
Exclusive to whole brain
circuit reconstruction
Integration of data from
structure & function
acquisition tools
22. Care about signals
Contributions outside
spiking domain?
Other cells?
Neuron-neuron effects
without spiking?
Evidence of sensations
retained?
Assume: spikes = currency
of sensations
Not epiphenomenal! (test?)
23. Predicting spikes
Observe / deduce
spike times original
system
Additional
information aids
prediction
What information
do the tools give
us?
Izhikevich
25. 3D reconstructions at 5nm
General classification (e.g.
pyramidal vs interneuron)
Detailed morphology,
segmented into
compartments
E.g. radius – resistance,
capacitance
Depends on neuron type
Measurement reliability,
cumulative
27. Ticker-Tape Data
Many neurons, several tapes
per neuron
Time stamps + spike /
membrane potential samples
Recovery of DNA snippets
Not combinable with EM
Interference with cell
mechanisms
Spatial registration:
Which part of ultrastructure
did it come from?
29. Microscopic wireless
Power & data volumes
compete with continuous
sampling
When enough sporadic
data?
Long term dynamics
Demands frequent spatial
registration
EM registration in tissue
Ongoing collaboration
(MIT, Harvard)
31. Learning from virtual systems
NETMORPH
Acquire structure data
Acquire functional data
Test algorithms &
iteratively improving
constraints
Calculate abstract
boundary conditions?
Netmorph.org
32. Proof of Concept: Starting Small
Test process in small system
C.Elegans (Dalrymple)
Retina (Briggman)
Hippocampal neuroprosthetic
(Berger)
Cerebellar neuroprosthetic
(Bamford)
Memory from piece of neural
tissue (Seung)
33. Discussion
Good gage of Tools = problem 1,
problems – proof of turning data into
concept! model = problem 2
SI is not new! True effort
Many fields can underway –
contribute seeking input from
SI experts!
34. Thanks
Ed Boyden (MIT)
Yael Maguire (MIT)
Konrad Kording
(NW)
Ken Hayworth (JF)
Many others in the
WBE group! Carboncopies.org
2045.com