Is AI Really Intelligent? Exploring Machine Learning, Neural Networks & Deep Learning

Machine Learning Overview

Machine learning can be viewed as an exhaustive search for the best parameters that fit a given dataset. It is the foundation for more complex techniques such as neural networks and deep learning.

Linear Regression Example

Linear regression assumes a linear relationship Y = C·X between an input X (e.g., square meters) and an output Y (e.g., mu). By collecting sample pairs (X, Y) we can estimate the coefficient C. The following figures illustrate a dataset, an initial guess C = 0.6, and the resulting prediction error.

Iteratively reducing C (e.g., to 0.05, then 0.0018) dramatically lowers the error, demonstrating the core training loop of adjusting parameters to minimise error.

Neural Networks

An artificial neuron sums weighted inputs and passes the result through an activation function, mimicking the behaviour of biological neurons that fire only when a threshold is exceeded.

The basic neuron diagram:

Common activation functions include the sigmoid (S‑function):

A three‑layer fully‑connected network consists of an input layer, one hidden layer, and an output layer, each with three neurons. All neurons in adjacent layers are connected with learnable weights.

Training steps:

Randomly initialise weight matrices.

Feed forward sample inputs.

Compute the error between predicted and true outputs.

Adjust the weights iteratively (e.g., via gradient descent) until the error is acceptable.

Concrete example:

Input vector I = [0.9, 0.1, 0.8] is multiplied by the first‑layer weight matrix W_ih to produce hidden‑layer activations.

First‑layer weight matrix W_ih:

Second‑layer weight matrix W_ho:

Final output after the second multiplication and sigmoid activation is 0.726:

Detailed calculation for the first hidden neuron:

I = [0.9, 0.1, 0.8]
W_ih (first row) = [0.9, 0.3, 0.4]
x = 0.9*0.9 + 0.1*0.3 + 0.8*0.4 = 1.16
y = sigmoid(1.16) ≈ 0.761

The hidden‑layer outputs become the input to the output layer, where the same matrix‑multiplication‑activation process repeats.

Python Implementation

import numpy
import scipy.special

class neuralNetwork:
    def __init__(self, inputnodes, hiddennodes, outputnodes, learningrate):
        self.inodes = inputnodes
        self.hnodes = hiddennodes
        self.onodes = outputnodes
        # Weight matrices initialised with a normal distribution
        self.wih = numpy.random.normal(0.0, pow(self.inodes, -0.5), (self.hnodes, self.inodes))
        self.who = numpy.random.normal(0.0, pow(self.hnodes, -0.5), (self.onodes, self.hnodes))
        self.lr = learningrate
        # Sigmoid activation function
        self.activation_function = lambda x: scipy.special.expit(x)

    def query(self, inputs):
        hidden_inputs = numpy.dot(inputs, self.wih)
        hidden_outputs = self.activation_function(hidden_inputs)
        final_inputs = numpy.dot(hidden_outputs, self.who)
        final_outputs = self.activation_function(final_inputs)
        return final_outputs

Deep Learning

Deep learning refers to neural networks with many hidden layers (typically three or more). Its practical success is largely confined to image‑recognition tasks, where inputs are naturally pixel matrices and convolution operations can extract useful features.

Convolution reduces irrelevant background pixels and emphasizes important structures (edges), which explains why deep learning excels at image classification but struggles with data lacking a regular grid structure.

Conclusion

Current AI systems execute strictly defined programs without consciousness or self‑awareness. They are sophisticated forms of automation; while parameter tuning can approximate intelligent behaviour, genuine independent thought is not achieved.