Illustrative Example
Who? UBC4006-01
From? Underwood International College, Yonsei University
When? Spring Semester 2026
Problem Statement: The Fruit Sorter
The Scenario A produce dealer needs a machine to automatically sort
mixed fruit on a conveyor belt. Three primitive sensors
measure specific properties:
Shape: 1 (round), −1 (elliptical)
Texture: 1 (smooth), −1 (rough)
Weight: 1 (> 1 lb), −1 (< 1 lb)
Goal: Classify the fruit based on these three binary inputs.
Problem Statement: System Overview
Figure: Conveyor belt and sensor system.
Problem Statement: Data Representation
The input vector p represents the three sensor readings:
p1 shape
p = p2 = texture
p3 weight
Target Prototype Vectors:
Orange: porange = [1, −1, −1]T
Apple: papple = [1, 1, −1]T
The Perceptron Network
We first use a single-layer perceptron.
It utilizes the hardlims (Symmetrical Hard Limit)
transfer function.
a = hardlims(Wp + b)
The Two-Input Case: Linear Separation
Perceptrons classify inputs into two distinct regions.
For a 2D case where W = [−1, 1] and b = −1:
n = [−1, 1]p − 1 = 0 =⇒ −p1 + p2 − 1 = 0
The decision boundary is where the net input n = 0.
Key Properties of Perceptrons
The decision boundary is always orthogonal to the
weight matrix W.
The bias b shifts the position of the boundary away
from the origin.
Limitation: Perceptrons can only classify patterns that
are linearly separable.
Wp + b = 0
Example: Separating Apples and Oranges
We aim for: Orange → −1 and Apple → 1.
Looking at the prototypes, texture (p2 ) is the
distinguishing factor.
We choose W = [0, 1, 0] and b = 0.
Testing the Perceptron
Perfect Apple: hardlims([0, 1, 0][1, 1, −1]T + 0) = 1 ✓
Perfect Orange:
hardlims([0, 1, 0][1, −1, −1]T + 0) = −1 ✓
Elliptical Orange: Input p = [−1, −1, −1]T :
a = hardlims(−1) = −1 (Classified as Orange)
Result: Any vector closer to the prototype in the feature
space is correctly classified.
The Hamming Network
Designed for binary pattern recognition.
Consists of two layers:
1 Feedforward Layer: Calculates correlation with prototypes.
2 Recurrent Layer: A competitive layer (MaxNet) to find the
”winner.”
Hamming Layer 1: Feedforward
Weights are set to the prototype patterns:
T
1 porange 1 −1 −1
W = =
pTapple 1 1 −1
Bias b1 is set to R (number of elements, here 3):
1 1 3
a =W p+
3
The output is maximum for the prototype with the
smallest Hamming Distance.
Hamming Layer 2: Recurrent (MaxNet)
Neurons compete; only the one with the highest initial
value survives.
Dynamics:
a2 (0) = a1 , a2 (t + 1) = poslin(W2 a2 (t))
Lateral Inhibition: W2 has 1 on the diagonal and −ϵ
elsewhere.
2 1 −ϵ 1
W = , ϵ<
−ϵ 1 S −1
The Hopfield Network
A recurrent network that acts as an associative memory.
Unlike Hamming (which outputs a class index), the
Hopfield network outputs the reconstructed prototype
pattern itself.
Hopfield Operation
Update Rule: a(t + 1) = satlins(Wa(t) + b)
Example Weight/Bias for a specific problem:
0.2 0 0 0.9
W = 0 1.2
0 , b = 0
0 0 0.2 −0.9
Behavior: Values saturate at 1 or −1. a2 acts as a
decision maker because its weight (1.2) is > 1.
Summary of Pattern Recognition Networks
Network Type Function
Perceptron Feedforward Linear decision boundary
Hamming FF + Recurrent Finds closest Hamming distan
Hopfield Recurrent Pattern reconstruction (Assoc