Top-Down Connections and its Applications to 3D Object Recognition - - PowerPoint PPT Presentation
Topographic Class Grouping through Top-Down Connections and its Applications to 3D Object Recognition Matthew D. Luciw and Juyang Weng Embodied Intelligence Laboratory Department of Computer Science Michigan State University East Lansing MI
Matthew D. Luciw and Juyang Weng Embodied Intelligence Laboratory Department of Computer Science Michigan State University East Lansing MI 48824 USA Email: {luciwmat, weng}@cse.msu.edu
Michigan State University 2
Found some brain areas where neurons grouped
E.g., fusiform face area (FFA) and
Neurons in inferotemporal cortex: object
How do such abstract meaning carrying neurons
Michigan State University 3
Discriminating features
Invariance
Motor-initiated abstraction
Can apply to any pattern recognition application
This study: 3D object recognition of center-normalized, background controlled objects.
TCG enabled
Significant reduction of the recognition errors.
Increased neuronal purity
Decreased class-response scatter
Michigan State University 4
Human-level intelligence must be developed [1] Model cortical developmental principals
Not modality specific Not application specific
Architecture capable of brain-like development + Experience
This work: Invariance & Abstraction
When output is label: invariance But motor output does not have to be class label General framework for motor-biased abstraction via supervision
from downstream
[1]: Weng, McClelland, Pentland, Sporns, Stockman, Sur and Thelen, Autonomous mental development by robots and animals, Science, 2001
Biologically inspired developmental networks
Hebbian-based firing rules on every layer develops feature detectors (neurons)
Hierarchical: from sensors (e.g., cameras) to motors (e.g., class label)
Three types of projections:
Bottom-up
Lateral
Top-down
Here we examine the effect of the top-down connections
Advantages:
Lowest complexity
Deals with high dimensional raw input
Fast and parallelizable
No local minima problem (does not use backpropagation for training)
Weng et al. IJHR 2007
Michigan State University 6
SOM (Kohenen, 1997)
Idea of using expanded input with input & output vectors
HDR (Zhang et al, 2005)
Used expanded input
AGREL (Roelfsema & Van Ooyen 2005)
Top-down connections for attention
LISSOM (Sit & Mikkulainen 2006)
Top-down connections for correlations (no motors)
MILN (Weng & Luciw, 2006, Weng et al. 2007)
Top-down connections to generate invariance
MSU Dataset: 25 Rotating Objects (25 classes) 200 images from each class 4/5 for training Used grayscale
NORB (5 classes) [2] – Left: training objects, Right: testing objects variations of elevation, lighting, rotation
[2]: Y. LeCun, F.J. Huang, and L. Bottou. Learning methods for generic object recognition with invariance to pose and lighting. CVPR, 2004.
Michigan State University 8
Three types of projections: Bottom-up, Lateral, Top-down Layer-one arranged in 2D plane
local firing rates (e.g., neighbors)
Lateral excitation, Do For each winner neuron j:
Lateral inhibition
Lobe component i: the principal component
Partition the input space Input space is top-down boosted if Beta nonzero
Michigan State University 12
Lateral Excitation Top-Down Connections
Both lateral excitation and top-down connections are required!
The top-down connections boost the variations in the neuronal between class directions during the training phase, leading to class discriminating features.
Lateral excitation ``pulls’’ nearby neurons in the neural plane to represent nearby features in the input space, which is boosted by the top-down connections.
Michigan State University 16
Two parameters tested
Weight of top-down connections (Beta)
Zero or 0.3
Size of neuronal map on layer-one
25 Objects: 20 x 20, 30 x 30, 40 x 40 NORB: 40 x 40, 60 x 60
5 trials for each set of parameters
Average values reported Each trial trained for 50,000 image/label pairs Error: disjoint testing samples Other metrics: used training data
(a): No Top-Down Excitation (Beta = 0) (b): Used Top-Down Connections (Beta = 0.3) Shows % of neuronal updating history for highest class
Invariance and abstraction are analogous
From top-down connections
1.
Within-class variation is disregarded
2.
Between-class variation is kept
Abstract class neurons are manifested in the representation.
Develops meaning: the features are tuned & organized from top-down to solve the imposed problem (here: classification, but could be something else)
No abstraction With abstraction
Michigan State University 19
Michigan State University 20
Michigan State University 21
Neuron purity: measured statistically by the average entropy of the
neurons’ development.
Use probability that neuron updated for each class Purer neurons are more “abstract,” -- characterizing class-specific (or
motor-specific) input information, resulting in better classification rates.
If purity is one, neuron i developed using samples from a single class.
Michigan State University 22
Entropy is (1-purity) (a): Without Top-Down (b): Using Top-Down
Michigan State University 23
Neurons that respond to the same class become relatively nearer. Measured statistically by a smaller within-class scatter of responses when
the neuronal plane has a fixed size.
c: # classes, n: # neurons, A: neuron probability matrix, N: neuron position
(2D), M: mean class firing positions
Michigan State University 24
(a): Eight images from a class (b): Bottom-up weight without TCG (c): Top-responding neuron for
each image without TCG
High class-response scatter
(d): Bottom-up weight with TCG (e): Top-responding neuron for
each image with TCG
Low class response scatter
Michigan State University 25
Developmental purity Class- response scatter Connectedness*
See paper for details.
Michigan State University 26
Purity Scatter Connected
Michigan State University 27
New Topographic Class Grouping mechanism Top-down connections in MILN develop class-
Led to lower error rates, more pure neurons, and
Towards a better understanding of how intelligent
Development of abstract, meaning carrying neurons General developmental principle for emergence-