Loading...
Industrial & Product Design

Robotic Parenting

Robotic Parenting
best value 3d printer
Image by jurvetson
A 20 year line-up of ASIMO humanoids… The early models look like a variety of Star Wars droids.

In this video clip video from the Honda labs, ASIMO looks like a child reaching out for a toy.

From Cognitive Computing ’07 in Berkeley today:

“Cognitive Computing is about engineering the mind by reverse engineering the brain.”

I ended my talk with a quote from Danny Hillis in The Pattern on the Stone:
“We will not engineer an artificial intelligence; rather we will set up the right conditions under which an intelligence can emerge. The greatest achievement of our technology may well be creation of tools that allow us to go beyond engineering – that allow us to create more than we can understand.”

Quotes from the Honda Research Institute talk, my favorite of the morning:
• for Honda, intelligence is a technology
• the essence of brain-like intelligence lies in the global organisation and self-referential control of processing
• following the analysis by synthesis principle, we verify our large scale hypotheses on our demonstrators in direct interaction with their environment
• in our strategy we approach the problem on several different levels of system organisation: macroscopic, mesoscopic, microscopic, microscopic & developmental
• first results confirm our approach to brain-like intelligent systems
• Open question: what is the role of the substrate? How close must a successful interpretation of the brain (in a technical sense) be to its underlying bio-chemical processes

Intelligence is a technology and a strategy for
• robust and flexible problem solving
• under resource limitations (time, energy)
• in complex environments (natural and artificial)

• the brain is the only intelligent system that we know of
• robots with rich environmental interaction provide us for the 1st time with the
means to study and verify large-scale hypotheses on brain-like intelligence
• our approach is to build the brain to understand the brain – the analysis by synthesis principle
• the brain is the most complex structure ever investigated by science
• it is not suitable to the most successful scientific analysis by decomposition
• the brain exhibits structural, chemical, plastic and dynamical complexity all intertwinned on different levels
• all processes in the brain are a result of information processing in a bio-chemical environment
• understanding the brain means unravelling the meaning of ourselves(related to cosmology)

brain = control system for organizing behavior

1) animals without cortex: autonomous systems (reflex automatons)
• genetically encoded reflex hierarchy with the limbic system at the top
• value system = genetically encoded mapping of sensory trigger features to behavioral prototypes

2) animals with cortex: flexible autonomous systems (learning systems)
reflex automaton +
• general memory architecture for storing experience
• genetically encoded self-referential control architecture

The stack [like OSI stack]:

A)Evo/Devo
Function:
• task embedded controlled cellular growth
• evolvable structures of spiking neural systems
• evolution of learning
• extract principles of simple brain evolution

Principles:
• co-evolution of genetic control and information expression
• evolutionary situated design
• selection driven interaction between evolution and learning
• major structural transitions of the co-evolution of early nervous systems and morphology

B) Microscopic Control Level
Function:
• elementary cortical processor
• rapid forward processing
• mixing prediction into afferent stream
• epochs of clocked, within asynchronous processing

Principles:
• spiking neurons
• cortical columnar architecture
• relative latency encoding
• rhythmic control of spike processing

Cortical development
• System architecture develops top-down.
• The basic control structure of the final system is present from the beginning.
• Development is marked by increasing sensory resolution and specialization of analysis, representation and control.

Self-referential Control Architecture
Minicolumn as elementary cortical processor
• mediates mixing of experience into afferent stream
• generates and synchronises rhythmic control for self-referential decomposition and learning
• relative spike latency encoding to control association width

The interplay between cortex and hippocampus increases memory capacity.
How does the cortex learn with:
• high memory capacity,
• fast retrieval speed, and
• high noise tolerance?
1. Store association A→B with HC (low memory capacity)
2. HC replays A→B to induce structural plasticity in cortex
3. Association A→B is stored in high-capacity cortical connections.
⇒ Structural plasticity leads to
– 10-20x memory capacity
– faster recall
– sparse connectivity
Short term memory is photographic — limited and inefficient — for a limited number of objects. Transferred to long term with more efficient and robust encoding.

C) Mesoscopic Control Level
Exploring, Learning and Understanding Visual Scenes
Function:
• active vision: fixation, saccading, tracking
• robust recognition and autonomous learning
• working memory and internal simulation
• self-organization of knowledge representation

Principles:
• columnar organization of multi-layered networks
• integration of different sensory analysis pathways
• stacked associative memories
• flexible selection of best-performing modular processing architecture (prediction, system monitoring)
• knowledge representation in task-related metric

Active Vision
• Decompose the sensory input into features & objects
• Use motion to distinguish foreground and background
• Compose a description of a scene
• Fixation by bottom-up & top-down attention
• Scan path & tracking
• Segmentation & prediction from movement
• Dynamic scene memory

D) Macroscopic Control Level
Function:
• self-development of practical intelligence
• autonomous interaction with environment
• a system that evolves itself from few innate abilities towards an autonomous and socially compliant partner

Principles:
• macroscopic architecture of the human brain
• child-like developmental strategy of learning
• integration of system components in a growing architecture
• self-referential control of learning
• a priori value system shaped by experience

Developing Intelligence
Child-like Acquisition of Representation and Language

Crossing the Levels
A-B) evolution of spiking neural systems
B-C) mixing of top-down prediction into afferent signal stream and active sensing and
online learning
A-D) evolutionary optimisation of functional modules

This research team in Frankfurt: 36 full time scientists + 52 students and interns

Q&A:
Q: How about building in a heart, or the machines will destroy us?
A: With emotion: we show our internal state
Value system. Map unknown input to output. Interaction with environment

Q from Stanford Prof. about vision:
A: We take several views of a 2D representation instead of building a 3D model

Q from Lloyd Watts: Do you use a spiking neuron model?
A: No. Open question: spiking neuron model, is it important? We are limited by computational resources.

Q from IBM Almaden: Can’t Asimo can use better arithmetic engines than the human brain
A: Hmmm…. We have not thought about teaching Asimo arithmetic. Good question. I will keep it in mind and pose the question to the robot.

Honda’s History of Humanoids provides a slider linking to great photos of their 20 year developmental effort.