AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

生成AIパスポート · LESSON 1

Chapter 1: AI Fundamentals and History13%（当サイト設定） · 4 topics

Learn how AI basically works, the types of machine learning, neural networks and the history of AI.

1.1How AI basically works (rule-based and machine learning)

Concept

Rule-based AI runs on “if X, then do Y” rules written by people. Its behavior is easy to explain, but it cannot handle situations outside its rules, and writing and maintaining rules takes effort. Machine learning learns rules (patterns) automatically from data. With lots of data it can make complex judgments, but it depends on the training data and its reasons can be hard to see.

How it works / steps

  1. If the rules of the problem are clear and stable, consider rule-based
  2. Use machine learning for complex judgments where rules can't be fully written (images, language)
  3. For machine learning, prepare data of suitable quality and quantity for the goal

Workplace example

Checking expense rules (send back if over the limit) suits rule-based; reading amounts from receipt images suits machine learning.

Common confusions and exam tips

Mini quiz

Q1. What is an AI mechanism that judges according to rules written by people (if X then Y) called?

  1. Deep learning
  2. Machine learning
  3. Reinforcement learning
  4. Rule-based
Show answer and explanation

Answer: D Rule-based

Rule-based systems have rules written in by people. They are weak against unexpected input and become harder to manage as rules grow. This limit led to machine learning, which learns rules from data.

Q2. Which best describes a limitation of rule-based systems?

  1. They struggle with unexpected inputs, and the effort to add and maintain rules keeps growing
  2. They cannot create a single rule without large amounts of training data and high computing power
  3. Results change every run even for the same input, so behavior is not reproducible
  4. Because they often cannot explain the reason for a judgment or trace which rule produced a result
Show answer and explanation

Answer: A They struggle with unexpected inputs, and the effort to add and maintain rules keeps growing

Rule-based systems are easy to explain and reproducible within the range their rules cover, but writing rules for every varied real-world input is difficult.

Reference links

1.2Types of machine learning

Machine learning (supervised, unsupervised, reinforcement and semi-supervised learning; no free lunch theorem)

Concept

Supervised learning learns from data with correct answers (labels) to do classification or regression (predicting numbers). Unsupervised learning finds structure in unlabeled data, used for clustering and the like. Reinforcement learning learns actions that maximize reward through trial and error. Semi-supervised learning combines a little labeled data with lots of unlabeled data. The no free lunch theorem says “no universal algorithm is best for every problem.”

How it works / steps

  1. Check whether correct labels exist (if so, supervised)
  2. No labels and you want grouping: unsupervised
  3. Actions earn rewards and trial and error is possible: reinforcement learning

Workplace example

Spam detection suits supervised learning, grouping customers by buying patterns suits unsupervised, and mastering games suits reinforcement learning.

Common confusions and exam tips

Mini quiz

Q1. What is the idea that “there is no universal algorithm that always beats the others on every problem” called?

  1. Scaling laws (power laws)
  2. Occam's razor
  3. No free lunch theorem
  4. Singularity hypothesis
Show answer and explanation

Answer: C No free lunch theorem

The no free lunch theorem shows that methods must be chosen to fit the nature of the problem. Scaling laws concern the relationship between model scale and performance; Moore's law is an empirical rule about semiconductor integration.

Q2. Which best describes the difference between discriminative AI and generative AI?

  1. Discriminative AI classifies or judges data; generative AI creates new data based on the features of the data it learned
  2. Discriminative AI runs on rule-based mechanisms; generative AI runs on neural networks
  3. They are distinguished by discriminative AI using only unsupervised learning and generative AI using only supervised learning
  4. Discriminative AI is a general term for models handling only images; generative AI handles only text
Show answer and explanation

Answer: A Discriminative AI classifies or judges data; generative AI creates new data based on the features of the data it learned

Discriminative AI judges which class an input belongs to; generative AI generates new data such as text, images and audio. They are not divided by learning method or modality.

Reference links

1.3Neural networks

Neural networks (how they learn, overfitting and countermeasures, transfer learning, CNN, RNN, LSTM, VAE, GAN, discriminative vs. generative AI)

Concept

Neural networks stack layers modeled on brain neurons and learn by adjusting weights with backpropagation. Deep stacks are deep learning. Overfitting (fitting training data too closely and doing poorly on new data) is handled by adding data, dropout, regularization and so on. Transfer learning reuses a trained model for another task. CNNs are used for images, RNNs and LSTMs for time series and text, and VAEs and GANs for generating images and more. Discriminative AI classifies and judges; generative AI creates new data.

How it works / steps

  1. For image tasks use CNNs; for sequence data, RNNs/LSTMs (Transformers in recent years)
  2. With little data, use a trained model via transfer learning
  3. Check for overfitting by comparing training and validation error

Workplace example

In product visual inspection, transfer learning from a trained image classifier achieves high accuracy even with few defect images.

Common confusions and exam tips

Mini quiz

Q1. In neural network training, which method propagates the difference between output and correct answer from the output layer toward the input layer and adjusts each layer's weights?

  1. Regularization (weight decay)
  2. Backpropagation
  3. Stochastic gradient descent (mini-batch learning)
  4. Forward propagation
Show answer and explanation

Answer: B Backpropagation

Backpropagation computes the gradient of the loss function in order from the output side, and gradient descent updates the weights. Dropout and early stopping are countermeasures against overfitting.

Q2. In a neural network, what is the function that transforms the sum of a node's inputs and passes it to the next layer called?

  1. Activation function
  2. Loss function
  3. Reward function
  4. Target variable
Show answer and explanation

Answer: A Activation function

Activation functions (sigmoid, ReLU, etc.) add non-linear transformations so complex relationships can be represented. A loss function measures the difference between prediction and correct answer.

Reference links

1.4History of AI

History of AI (from the Dartmouth workshop to the third AI boom, generative AI, the singularity)

Concept

The term “artificial intelligence” was used at the 1956 Dartmouth workshop, and AI developed through the first boom (search and inference), the second (knowledge and expert systems) and the third (machine learning, deep learning and big data). After deep learning's success in the 2010s, the 2017 Transformer and 2022 ChatGPT spread generative AI widely. The singularity (technological singularity) is a concept for the point where AI surpasses human intelligence and society changes unpredictably.

How it works / steps

  1. Sort out each boom's core technology and limits (why winters came)
  2. Know the drivers of the third boom (data, computing resources, algorithms)
  3. Know generative AI's turning points (Transformer, ChatGPT)

Workplace example

In training, explain “why AI became usable now” with three points: more data, better GPUs and deep learning.

Common confusions and exam tips

Mini quiz

Q1. Which conference, held in 1956, is said to be where the term “artificial intelligence” was first used?

  1. Asilomar Conference
  2. NeurIPS
  3. G7 Hiroshima Summit
  4. Dartmouth Conference
Show answer and explanation

Answer: D Dartmouth Conference

The Dartmouth Conference is considered the start of AI research and led to the first AI boom.

Q2. Which best describes the main reason the second AI boom faded?

  1. It is thought that there was no way to collect large numbers of images, so deep learning could not be trained
  2. Generative AI output plausible errors and drew heavy public criticism
  3. Having people write specialists' knowledge in as rules took great effort, and handling exceptions was also difficult
  4. Semiconductors were expensive, so search and inference programs could not run at all
Show answer and explanation

Answer: C Having people write specialists' knowledge in as rules took great effort, and handling exceptions was also difficult

For expert systems, acquiring and maintaining knowledge became the bottleneck (the knowledge acquisition bottleneck).

Reference links

Lessons are written by this site based on public information and do not guarantee exam content. Product specifications change, so check the latest information at the reference links.

Notes (unofficial, original questions, no pass guarantee)