Neural Networks untuk Pemula - Perkuliahan Soft Computing #06

Neural Networks Core Concepts and Capabilities
📌 Neural Networks (NNs) are techniques shared by both Soft Computing and Machine Learning, situated within the broader field of Artificial Intelligence (AI).
📌 Key capabilities of NNs include Classification (outputting categories, e.g., spam vs. non-spam email), Prediction (outputting a numerical value, e.g., predicting house prices or temperature), and Clustering (grouping data based on similarity without pre-defined categories).
📌 The structure of an Artificial Neural Network (ANN) consists of an Input Layer, one or more Hidden Layers (the "bridge"), and an Output Layer.
📌 Biological neurons inspired the ANN, consisting of a cell body, axon, and dendrites, which transmit signals via synapses; this biological function is modeled mathematically in ANNs.

Classification Examples and Structure
📌 Classification examples include identifying spam emails, classifying COVID-19 test results (positive/negative), facial recognition (identifying person A vs. person B), and speech recognition (identifying phonemes or words).
📌 In classification tasks, the output layer neurons represent categories; if there are $K$ categories, the output layer typically has $K$ neurons, often utilizing One-Hot Encoding (e.g., if output neuron $Y_i = 1$, all others are $0$).
📌 For digit recognition (0-9), an image of $13 \times 10$ pixels (130 pixels total) serves as input, where each pixel value (0 or 1) becomes an input feature ($X_1$ to $X_{130}$).

Prediction Implementation Details
📌 Prediction tasks, like forecasting gold prices or house values, always result in a single numerical output, meaning the output layer typically contains only one neuron.
📌 The raw output value from the single prediction neuron, which ranges from $0$ to $1$, must undergo denormalization to convert it back into a meaningful real-world value (e.g., currency or temperature).
📌 For instance, predicting house prices uses inputs like land size and number of bedrooms; the output $Y_1$ is calculated via the weighted sum of hidden layer outputs, followed by activation using the sigmoid function, $Y_1 = 1 / (1 + e^{-x})$.

Learning and Network Architectures
📌 The crucial element in NNs is the Weight associated with each Connection; these weights determine the strength of the connection and are not assigned arbitrarily.
📌 Weights are determined through a Learning Process (training), which involves forward propagation to calculate an output, comparing it to the target (error calculation), and then using Backpropagation to iteratively adjust weights backward through the network until the error is minimized.
📌 Initial weights are set randomly, and the network converges toward optimal weights through repeated cycles of forward calculation and backward weight adjustment.
📌 Basic architectures include the Perceptron (input and output layers only), the Multi-Layer Perceptron (MLP) (including one or more hidden layers), and Deep Neural Networks (having more than two hidden layers).
📌 Recurrent Neural Networks (RNNs) feature feedback loops where outputs from a layer are fed back as inputs to a previous layer, serving as a form of memory to retain information from sequential data.

Key Points & Insights
➡️ NNs perform Classification (outputting categories) and Prediction (outputting numerical values); the key distinction lies in the output structure.
➡️ The computational value inside any neuron is calculated as the sum of (Input $\times$ Weight) from the preceding layer, which must then be passed through an activation function (like sigmoid) to produce the final neuron output.
➡️ Determining the correct weights is the core of NN training, achieved by minimizing the difference between the calculated output and the desired target output through backpropagation.
➡️ MLPs are widely used due to their good performance and relative structural simplicity compared to deep networks or RNNs.

📸 Video summarized with SummaryTube.com on Feb 06, 2026, 15:38 UTC