An Overview of Artificial Intelligence, Machine Learning, and Neural Networks

As an amateur AI enthusiast, the author offers a modest introduction to neural networks, acknowledging possible inaccuracies.

Artificial Intelligence

AI is widely used in everyday life and platforms like Douyin; it is defined as the creation and use of algorithms to simulate human intelligence, with machine learning being a sub‑field of AI.

function ArtificialIntelligenceAdd(a, b) { return a + b }

AI aims to enable computers to think and act like humans, and machine learning currently excels at repetitive tasks.

Machine Learning

Machine learning extracts patterns from data to make predictions, relying on data and learning methods; deep learning is a subset of machine learning.

Four learning types exist: supervised, unsupervised, semi‑supervised, and reinforcement learning.

Supervised Learning

Models learn a function from labeled input‑output pairs; an example table shows inputs x1‑x3 and target y.

Unsupervised Learning

Models discover structure without labeled targets, e.g., clustering animal data.

Semi‑Supervised Learning

Combines a small amount of labeled data with abundant unlabeled data to reduce labeling effort.

Reinforcement Learning

Agents adjust behavior based on rewards and penalties from the environment, illustrated by a wolf‑sheep chase example.

Neural Networks

Artificial Neural Networks (ANN) are the most popular learning method, driving the rise of deep learning.

History

1943 – McCulloch & Pitts propose a threshold‑logic model.

1975 – Paul Werbos invents back‑propagation.

1980s – Convolutional Neural Networks introduced.

1993 – NVIDIA founded.

1999 – NVIDIA releases GeForce 256 GPU.

2006 – CUDA framework appears.

2010 onward – CNNs explode with GPU acceleration.

2014 – Residual networks enable deeper models, birthing deep learning.

Working Principle

An example predicts weight from weekly milk‑tea consumption using a simple linear model, introducing weights (w) and bias (b).

function simpleLossFun(prediction, groundtruth) { return Math.abs(prediction - groundtruth); }

The loss function measures prediction error; training aims to minimize this loss.

Back‑Propagation

Gradients of each weight are computed to adjust parameters, using the chain rule for composite functions.

Parameter Update

Weights are updated by gradient × learning‑rate; learning‑rate decay controls step size over epochs.

Overall Training Process

Forward propagation: model predicts outputs.

Loss computation: compare predictions with ground truth.

Back‑propagation: compute gradients.

Parameter correction: update weights using gradients, learning rate, and decay.

Repeating these steps yields a trained model.

Recruitment Notice

The ByteDance Interactive R&D department seeks front‑end engineers for products like Douyin, offering opportunities in Shanghai, Beijing, and Shenzhen. Links to job postings and a referral code are provided.