Signal Perceptron: On the Identifiability of Boolean Function Spaces and Beyond

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Theory

Background:

• The perceptron, a foundational model in artificial intelligence, is limited to learning linearly separable functions.
• Minsky and Papert demonstrated that single-layer perceptrons cannot learn non-linear functions like XOR.
• Existing parallel machine implementations face capacity limitations in learning complex functions.

Purpose of the Study:

• To propose a novel, more powerful implementation of parallel machines.
• To overcome the limitations of traditional perceptrons in learning non-linear Boolean functions.
• To introduce a new mathematical tool utilizing analytic sinusoids for function representation.

Main Methods:

• Developed a new parallel machine architecture based on analytic sinusoids.
• Formulated an analytic signal representation for functions.
• Utilized a single-layer mechanism for learning non-linear k-ary Boolean functions.

Main Results:

• The proposed mechanism can learn any non-linear k-ary Boolean function with a single layer.
• Demonstrated superior performance compared to single hidden layer multilayer perceptrons.
• Achieved faster learning and required fewer parameters in Boolean function learning and image classification.

Conclusions:

• The new analytic sinusoid-based parallel machine offers enhanced learning capacity.
• This approach represents a significant advancement over traditional perceptron models.
• The method shows practical advantages in computational efficiency and performance for complex learning tasks.