Bio-mimetic Control Research Center, RIKEN Guided learning from images using an uncertain granular model and bio-mimicry of the human fovea Jonathan Rossiter.

理化学研究所Bio-mimetic Control Research Center, RIKEN

Guided learning from images using an uncertain granular model and

bio-mimicry of the human fovea

Jonathan Rossiter

Toshiharu Mukai

Institute of Physical and Chemical Research

RIKEN, Japan

+

University of Bristol

理化学研究所2 Bio-mimetic Control

Research Center, RIKEN

Bio-mimetic AI

Studying and copying human intelligence and behaviours in artificially intelligent systems

• Perceptions and sensing• Representations• Reasoning• Learning • Adapting and updating



Motivation: human-like robotics

Rescue Robot• Hazardous• Real-time/on-site training• Remote control • Autonomous Intelligence

Guided operation• Dumb

Guided learning• Dumb, but at least it’s learning…



Consider only image domain Learning from image data (goal is high level model)

Crisp image data• Conventional features

• Crisp values

But what is uncertain image data?• High level concepts encroaching on low level data

• Degrees of applicability/relevance across larger scale features

So need to combine both crisp image data and uncertain image data

crisp image data → induction → uncertainty model

uncertain image data → induction → uncertainty model



Learning with granules

Size(P) = { small : 0.3, medium : 0.7}

Cost(P) = { reasonable : 0.2, cheap : 0.8}

GP = { small^reasonable: 0.3*0.2, small^cheap: 0.3*0.8, medium^reasonable: 0.7*0.2, medium^ cheap: 0.7*0.8 }



Learning with granules A granule is thus a discrete fuzzy set G over the universe

of cross-product labels L:G = {l L : m [0, 1]}

where:

L =×(Ki | i = 1, . . . , n) and Ki is a single fuzzy set label (e.g. small, medium, etc) In this paper the aggregation operation used to turn

training instances Gj into the model GM is simply :

GM = Norm(j Gj)

And with applicability values this becomes:

GM = Norm(j Gj × aj)



Human visual system – from light to electricity



Light sensors in the retina



Vision and active learning





Fovea- based region focus

Applicability

Fovea scaling

Relative Absolute





Unit applicability

34.6%Relative scale Gaussian-type

applicability ( = 0.3)

49.5%

Absolute scale Gaussian-type applicability ( = 0.4 over 5%)

48.9%



Trapezoidal applicabilityx = 0

83.9%

Relative scale Gaussian-typeapplicability ( = 0.2)

83.9%

Unit applicability

82.1%



Conclusions Fovea-like applicability functions better

• Natural• Incorporates into linguistic inductive learning

Not clear whether relative or absolute functions are better

• But, with relative applicability we need not worry about absolute scale. Easier.

Further research • Optimize the choice of applicability function • Incorporating such a system into tools to aid medical diagnosis

and into vision systems for rescue robots operating in hazardous environments.



Thank you



Image feature scale



High level + low level information

Low level• Sensor based• Data rich• Crisp/precise

High level • Taxononomical• Conceptual• Linguistic• Uncertain

Fusion in training• High + low best of both worlds



Updating robot vision

Human guidance of robot • Varied environments



Human-like perception

Goldberg and terminus of perception• Image features in abstract to terminus• Kind of high-level from low level

Also have high level information• Modifies/constrains our views of image data

• Examples

Humans combine both high and low level image information

• Good place to look for inspiration • Bio-mimetic high level approaches to reasoning with

information and sensor data

Bio-mimetic Control Research Center, RIKEN Guided learning from images using an uncertain granular model and bio-mimicry of the human fovea Jonathan Rossiter.

Documents

biomimetic ai

chemical research riken

j slide

riken guided learning

electricity slide

retina slide

active learning slide

image feature scale