G検定 StudyG Kentei (JDLA Deep Learning for GENERAL)

Learning goals

• Explain the four levels of AI and the AI effect with examples
• Distinguish issues such as the frame problem, the symbol grounding problem and the Chinese room
• Explain the technologies of the three AI booms (search and inference, knowledge representation, machine learning and deep learning) with representative examples

Key points

• AI levels are explained in four stages: “simple control programs,” “classical AI (using search, inference and knowledge),” “machine learning” and “deep learning (which also learns features).”
• AI effect: once how something works is understood, it is regarded as “not intelligence.”
• The frame problem is the difficulty of picking out only what is relevant to the task at hand; the symbol grounding problem is being unable to connect symbols to real-world meaning.
• The Turing test checks whether a machine can be told apart from a person through conversation. The Chinese room is a thought experiment countering it: manipulating symbols by rules does not mean understanding meaning (the strong vs. weak AI debate).
• The first boom (1950s–60s) used search and inference and solved toy problems but not real ones; the second (1980s) used knowledge representation and expert systems and hit the knowledge acquisition bottleneck; the third (2000s onward) is machine learning and deep learning.
• The Mini-Max method evaluates assuming the opponent plays the best move; alpha-beta pruning cuts unneeded branches to reduce search. The Monte Carlo method evaluates with random playouts.
• At ILSVRC 2012, AlexNet (deep learning) won by a wide margin, and the third boom took off. In 2016, AlphaGo beat a professional Go player.

Workplace uses

• PlanningWhen considering automating inquiry handling, separate the scope rules (rule-based) can cover from the scope to learn from data.

Exam objectives (from the exam guide)

• 人工知能の定義（AI のレベル分け、AI 効果、人工知能とロボットの違い）
• 人工知能分野で議論される問題（チューリングテスト、中国語の部屋、フレーム問題、シンボルグラウンディング問題、シンギュラリティ）
• 探索・推論（探索木、幅優先・深さ優先探索、Mini-Max 法、αβ 法、モンテカルロ法、STRIPS、SHRDLU）
• 知識表現とエキスパートシステム（意味ネットワーク、オントロジー、MYCIN、DENDRAL、Cyc、ワトソン）
• 機械学習（ビッグデータ、次元の呪い、統計的自然言語処理）
• ディープラーニング（ILSVRC、AlexNet、AlphaGo、生成 AI）

Learning resources

• G Kentei exam page (syllabus download) ↗
• G検定 exam guide / study guide ↗

Learning goals

• Choose supervised, unsupervised or reinforcement learning and representative methods to fit the problem
• Calculate accuracy, precision, recall and F1 from a confusion matrix and choose the metric that fits the goal
• Explain the holdout method, k-fold cross-validation and how to spot overfitting

Key points

• Supervised learning learns from pairs of features and correct answers (training labels). Predicting numbers is regression; predicting categories is classification. Despite its name, logistic regression is a classification method.
• Ensemble learning: bagging (training in parallel on bootstrap samples, e.g., random forest) and boosting (each model makes up for the previous one's errors, e.g., gradient boosting).
• SVMs set the boundary by maximizing the margin, and the kernel trick lets them handle nonlinear separation.
• Unsupervised learning: k-means splits data into k clusters (k is set in advance); Ward's method is hierarchical clustering drawn as a dendrogram. PCA reduces dimensions along directions of large variance; t-SNE is nonlinear dimensionality reduction for visualization.
• Reinforcement learning learns a policy from actions and rewards. There are value-based methods (Q-learning, SARSA), policy-based methods (REINFORCE, policy gradient) and Actor-Critic combining both. ε-greedy explores randomly with probability ε.
• Precision = TP/(TP+FP): the share of correct ones among those predicted positive. Recall = TP/(TP+FN): the share of actual positives found. Stress recall to avoid misses and precision to avoid false alarms.
• Overfitting means small error on training data but large error on validation data. With little data, estimate generalization with k-fold cross-validation.

Workplace uses

• SalesIn churn prediction (classification), you want fewer missed churners, so stress recall and check the balance with precision via F1.
• MarketingGroup customers by purchase history with k-means and vary measures by group.

Exam objectives (from the exam guide)

• 教師あり学習（回帰・分類、線形回帰、ロジスティック回帰、決定木、ランダムフォレスト、勾配ブースティング、SVM、AR モデル）
• 教師なし学習（k-means 法、ウォード法、主成分分析、t-SNE、協調フィルタリング、トピックモデル）
• 強化学習（マルコフ決定過程、割引率、ε-greedy、Q 学習、SARSA、方策勾配法、Actor-Critic）
• モデルの選択・評価（交差検証、混同行列、適合率・再現率・F 値、ROC 曲線・AUC、MSE・RMSE・MAE、AIC・BIC、過学習）

Learning resources

• G Kentei exam page (syllabus download) ↗
• G検定 exam guide / study guide ↗

Learning goals

• Choose activation and loss functions to fit the task
• Explain the relationship between backpropagation and the vanishing gradient problem, and countermeasures
• Explain the issues with gradient descent (local minima, saddle points, learning rate), representative optimizers and hyperparameter tuning

Key points

• A simple perceptron can solve only linearly separable problems (it cannot solve XOR). A multilayer perceptron with hidden layers can handle nonlinear problems.
• GPUs excel at large amounts of simple parallel computation (matrix operations); TPUs are processors specialized for tensor computation, developed by Google.
• The sigmoid's derivative peaks at 0.25, so gradients shrink as layers stack (vanishing gradients). ReLU's derivative is 1 in the positive region, so gradients vanish less easily. Output layer: identity for regression, sigmoid for binary classification, softmax for multiclass classification.
• Loss functions: mean squared error for regression, cross-entropy for classification. KL divergence measures differences between distributions. Contrastive loss and triplet loss are used for metric learning.
• L1 regularization (lasso) tends to set unneeded weights to 0, acting as feature selection; L2 regularization (ridge) shrinks weights overall. Dropout randomly disables neurons during training.
• Backpropagation computes gradients efficiently from the output side using the chain rule.
• Stochastic gradient descent updates using part of the data (mini-batches). Momentum dampens oscillation with inertia; AdaGrad and RMSprop adjust the learning rate per parameter; Adam combines both ideas.
• Hyperparameters (learning rate, number of layers, etc.) are searched by grid search, random search or Bayesian optimization. Early stopping halts training when validation error starts to worsen.

Workplace uses

• ITMove training of an image inspection model from in-house PCs (CPU) to cloud GPU instances to shorten training time.

Exam objectives (from the exam guide)

• ニューラルネットワークとディープラーニング（単純パーセプトロン、多層パーセプトロン、CPU・GPU・TPU）
• 活性化関数（シグモイド、tanh、ReLU、Leaky ReLU、ソフトマックス、勾配消失問題）
• 誤差関数（平均二乗誤差関数、交差エントロピー、KL 情報量、Contrastive Loss、Triplet Loss）
• 正則化（L1・L2 正則化、ラッソ回帰、ドロップアウト）
• 誤差逆伝播法（連鎖律、勾配消失・勾配爆発、信用割当問題）
• 最適化手法（勾配降下法、確率的勾配降下法、モーメンタム、AdaGrad、RMSprop、Adam、学習率、早期終了、二重降下現象、ハイパーパラメータ探索）

Learning resources

• G Kentei exam page (syllabus download) ↗
• G検定 exam guide / study guide ↗

Learning goals

• Explain how to think about convolution layer output size and parameter count, and how they differ from fully connected layers
• Explain the differences and roles of RNN, LSTM, GRU, attention and the Transformer
• Explain the purposes of autoencoders and VAEs, normalization layers and data augmentation, and when to use each

Key points

• A convolution layer slides filters (kernels) across the image to extract local features, and shares weights, so it has fewer parameters than a fully connected layer. Output size = (input − filter + 2 × padding) / stride + 1.
• Dilated (atrous) convolution widens the receptive field by leaving gaps; depthwise separable convolution splits into per-channel convolution and 1×1 convolution to cut computation (MobileNet).
• Batch normalization normalizes per channel within a mini-batch; layer normalization normalizes over all features within one sample (used in Transformers). Instance normalization works per sample and per channel (style transfer).
• Max pooling gives invariance to small positional shifts and has no parameters to learn. GAP takes the average of each feature map and can replace fully connected layers.
• Skip connections (ResNet) add the input past skipped layers, so gradients flow easily even in deep networks.
• RNNs are trained with BPTT and suffer vanishing gradients on long sequences. LSTMs keep long-term dependencies with input, forget and output gates and a cell (CEC); GRUs simplify this. Teacher forcing feeds the previous step's correct answer as input during training.
• The Transformer uses no RNN and computes relationships among all tokens in parallel with self-attention. Word order is given by positional encoding. Multi-head attention computes attention from multiple viewpoints.
• An autoencoder learns to compress input into a latent representation and reconstruct it (dimensionality reduction, anomaly detection). A VAE represents latent variables as probability distributions and can generate new data.
• Data augmentation transforms training data to increase it and improve generalization. Cutout and Random Erasing hide parts, Mixup blends two images and CutMix cuts and pastes parts. For text, paraphrasing and similar methods.

Workplace uses

• ManufacturingSince inspection images are scarce, train with augmentation by flipping, rotating and changing brightness to improve accuracy on photos of unseen products.

Exam objectives (from the exam guide)

• 全結合層・畳み込み層（カーネル、ストライド、パディング、特徴マップ、Dilated Convolution、Depthwise Separable Convolution）
• 正規化層（バッチ正規化、レイヤー正規化、インスタンス正規化、グループ正規化）
• プーリング層（最大値プーリング、グローバルアベレージプーリング）・スキップ結合（ResNet）
• 回帰結合層（RNN、BPTT、LSTM、GRU、双方向 RNN、教師強制、エルマンネットワーク）
• Attention（Seq2Seq、Self-Attention、Source-Target Attention、Multi-Head Attention、Transformer、位置エンコーディング、クエリ・キー・バリュー）
• オートエンコーダ（積層オートエンコーダ、VAE、VQ-VAE・β-VAE）
• データ拡張（Crop、Flip、Rotate、Cutout、Random Erasing、Mixup、CutMix、RandAugment、paraphrasing）

Learning resources

• G Kentei exam page (syllabus download) ↗
• G検定 exam guide / study guide ↗

Learning goals

• Match image recognition task types with representative models
• Explain feature representations and representative models in natural language and speech processing
• Choose representative methods for generative models, transfer learning, multimodality, interpretability and model compression to fit the goal

Key points

• Object recognition gives the class of the whole image; object detection gives positions (bounding boxes) and classes; semantic segmentation gives a class per pixel; instance segmentation gives each object's region; panoptic segmentation does both.
• Object detection: two-stage R-CNN → Fast R-CNN → Faster R-CNN (which also learns region proposals with an RPN), and one-stage YOLO and SSD (fast). Segmentation: FCN, U-Net (encoder-decoder with skip connections), Mask R-CNN (instances).
• word2vec turns words into distributed representations (vectors); CBOW predicts the center word from surrounding words, and skip-gram predicts surrounding words from the center word. fastText uses subwords. ELMo gives context-dependent representations.
• BERT is a bidirectional model using the Transformer encoder, pre-trained with masked word prediction (MLM) and next sentence prediction. GPT is an autoregressive decoder model that predicts the next word.
• Speech: A-D conversion (sampling, quantization) → spectrum via FFT → features such as mel-scale MFCCs. CTC learns alignment between inputs and outputs of different lengths. WaveNet generates speech waveforms directly.
• DQN approximates the Q-function with a CNN, stabilized with experience replay and a target network. RLHF builds a reward model from human ratings and tunes the language model with reinforcement learning. sim2real transfers learning from simulation to real hardware (domain randomization).
• GANs pit a generator against a discriminator. Pix2Pix converts paired images; CycleGAN converts without pairs. Diffusion models generate by learning the noise-removal process. NeRF represents 3D scenes from images taken from multiple viewpoints.
• Transfer learning reuses a trained model for another task; fine-tuning also further adjusts the weights. Catastrophic forgetting is losing earlier knowledge when learning a new task. CLIP learns to match images and text and can classify zero-shot.
• Grad-CAM uses gradients to show, as a heat map, which parts of the image drove the decision. LIME approximates locally with a simple model; SHAP shows feature contributions with Shapley values from game theory.
• Distillation trains a small model (student) using a large model's (teacher's) outputs. Pruning removes low-importance weights; quantization lowers numeric precision (bit width).

Workplace uses

• RetailFind empty shelves via object detection on store camera footage, running a model quantized and distilled for edge AI on in-store devices.

Exam objectives (from the exam guide)

• 画像認識（物体識別・物体検出・セグメンテーション・姿勢推定；AlexNet、VGG、GoogLeNet、ResNet、EfficientNet、R-CNN 系、YOLO、SSD、U-Net、Mask R-CNN、Vision Transformer）
• 自然言語処理（BoW、TF-IDF、word2vec、fastText、ELMo、BERT、GPT、LLM、GLUE）
• 音声処理（A-D 変換、PCM、FFT、MFCC、隠れマルコフモデル、WaveNet、CTC、話者識別）
• 深層強化学習（DQN、ダブル DQN、Rainbow、A3C、PPO、AlphaGo・AlphaStar、RLHF、sim2real、オフライン強化学習）
• データ生成（GAN、DCGAN、Pix2Pix、CycleGAN、Diffusion Model、NeRF）
• 転移学習・ファインチューニング（Few-shot、破滅的忘却、自己教師あり学習）・マルチモーダル（CLIP、DALL-E、基盤モデル、Zero-shot）
• モデルの解釈性（XAI、CAM、Grad-CAM、LIME、SHAP、Permutation Importance）・モデルの軽量化（蒸留、プルーニング、量子化、宝くじ仮説、エッジ AI）

Learning resources

• G Kentei exam page (syllabus download) ↗
• G検定 exam guide / study guide ↗

Learning goals

• Explain the CRISP-DM phases and the roles of PoC and MLOps
• Explain cautions for data collection and annotation, and pitfalls such as leakage
• Calculate and choose basic statistics, probability distributions and distances

Key points

• CRISP-DM: business understanding → data understanding → data preparation → modeling → evaluation → deployment.
• A PoC (proof of concept) checks effects and feasibility on a small scale before full adoption. Repeating PoCs without reaching production is sometimes called “PoC poverty.”
• MLOps runs model development and operation as one, monitoring accuracy, retraining and managing versions. When data distributions change in operation, accuracy drops (data drift, concept drift).
• Annotation is the work of attaching correct labels; document the criteria and check for variation among annotators.
• Data leakage is when information that should not be available (future information or the answer itself) gets mixed into training, making evaluation unjustly good.
• The median is less affected by outliers. Correlation does not necessarily mean causation; apparent correlation due to a third factor is called spurious correlation.
• Cosine similarity measures closeness of vector direction (e.g., document similarity); Mahalanobis distance accounts for the data's distribution (covariance) and is used in anomaly detection.

Workplace uses

• Corporate PlanningTrial a demand forecasting AI in a three-month PoC, setting evaluation metrics and pass criteria in advance before deciding on production adoption.

Exam objectives (from the exam guide)

• AI プロジェクトの進め方（CRISP-DM、CRISP-ML、PoC、MLOps、BPR、アジャイル、Web API、クラウド、Docker、Jupyter Notebook）
• データの収集・加工・分析・学習（オープンデータセット、アノテーション、コーパス、データリーケージ）
• AI に必要な数理・統計知識（平均・中央値・最頻値、分散・標準偏差、相関係数・疑似相関、正規分布・二項分布・ポアソン分布、条件付き確率、最尤法、最小二乗法、仮説検定、ユークリッド距離・マハラノビス距離・コサイン類似度）

Learning resources

• G Kentei exam page (syllabus download) ↗
• G検定 exam guide / study guide ↗

Learning goals

• Distinguish and apply personal information, personal data, special care-required personal information, anonymously processed information and pseudonymously processed information
• Explain the intent of Article 30-4 and the thinking on copyrightability of AI output and copyright infringement
• Explain the three requirements for trade secrets, the stages of AI development contracts, and the difference between quasi-mandate and contracts for work

Key points

• Personal information is information about a living individual that can identify a specific person. Individual identification codes such as face authentication data and My Number are personal information on their own. Race, creed, medical history and the like are special care-required personal information, which in principle requires the person's consent to obtain.
• Third-party provision in principle requires the person's consent. However, outsourcing, business succession and joint use do not count as “third parties.”
• Anonymously processed information is processed so the person cannot be identified or restored, and can be provided to third parties without consent under certain conditions. Pseudonymously processed information is processed so it cannot identify the person unless matched with other information; it can be used for internal analysis but in principle cannot be provided to third parties.
• GDPR is the EU's personal data protection regulation and can apply to businesses outside the EU that offer goods or services to individuals in the EU.
• Article 30-4 of the Copyright Act: uses not aimed at enjoying the thoughts or feelings expressed (such as information analysis for AI training) are in principle allowed without permission, except where they unreasonably harm the copyright holder's interests.
• AI output can be a copyrighted work when a person has creative intent and makes a creative contribution. If output resembles an existing work and relies on it, it can be infringement.
• Patent requirements include novelty and inventive step. The three requirements for trade secrets are secrecy management, usefulness and non-public status. Shared data with limited access is data managed with IDs, passwords and the like and provided to specific parties.
• METI's Contract Guidelines on Utilization of AI and Data propose exploratory, staged development: assessment → PoC → development → additional training. Quasi-mandate contracts are considered suitable for stages with uncertain outcomes.

Workplace uses

• LegalWhen co-developing a demand forecasting model with a partner, set by contract the contract form for each stage, ownership of rights to the trained model and data, and terms of use.

Exam objectives (from the exam guide)

• 個人情報保護法（個人情報・個人識別符号・要配慮個人情報、利用目的、第三者提供、委託、匿名加工情報・仮名加工情報、GDPR）
• 著作権法（著作物・創作性、著作権法 30 条の 4、AI 生成物、利用規約）
• 特許法（発明・新規性・進歩性、職務発明）
• 不正競争防止法（営業秘密の三要件、限定提供データ）・独占禁止法
• AI 開発委託契約（AI・データの利用に関する契約ガイドライン、探索的段階型、NDA、準委任契約・請負契約）
• AI サービス提供契約（SaaS、利用規約、データ利用権、精度保証）

Learning resources

• G Kentei exam page (syllabus download) ↗
• G検定 exam guide / study guide ↗

Learning goals

• Explain hard law, soft law and the risk-based approach, with their pros and cons
• Apply the causes of and countermeasures for privacy, fairness, safety and transparency issues to cases
• Explain AI governance structures and methods (ethics assessments, diverse members, audits)

Key points

• Hard law is legally binding law; soft law is non-binding norms such as guidelines. Soft law adapts flexibly to change but has weak enforcement.
• Risk-based approach: the strength of regulation varies with the size of the risk (the EU AI Act classifies AI as prohibited, high risk, limited risk and minimal risk).
• Privacy by design builds privacy protection in from the planning and design stage. When using camera images, provide notices and signs, prevent use beyond the purpose, limit retention periods and so on.
• Fairness: biased training data (sampling bias) or reflected past discrimination produce results unfavorable to certain attributes. Even when sensitive attributes are removed, discrimination can arise via correlated proxy variables.
• Adversarial examples are attacks that induce AI misrecognition with small noise imperceptible to people. Data poisoning mixes malicious data into training data. Model extraction copies a model through massive API queries.
• A filter bubble is being surrounded by only the information you like because of recommendations; an echo chamber is opinions amplifying among people who share them.
• Power consumption and CO2 emissions from model training are environmental issues. Changes in employment due to AI, loss of skills and collaboration with AI are labor issues.
• AI governance combines setting an AI policy, ethics assessments, review by diverse members (diversity), external audits and ensuring traceability.

Workplace uses

• HRBefore using AI in document screening for hiring, check bias in training data and differences in results by attribute, and decide on a structure where people make the final decision and how to explain it.

Exam objectives (from the exam guide)

• 国内外のガイドライン（AI 原則、ハードロー・ソフトロー、リスクベースアプローチ、EU AI Act、AI 事業者ガイドライン）
• プライバシー（プライバシー・バイ・デザイン、カメラ画像利活用ガイドブック）
• 公平性（アルゴリズムバイアス、サンプリングバイアス、センシティブ属性）
• 安全性とセキュリティ（Adversarial Examples、データ汚染、モデル窃取、セキュリティ・バイ・デザイン）・悪用（ディープフェイク、フェイクニュース）
• 透明性（説明可能性、データの来歴、ブラックボックス）・民主主義（フィルターバブル、エコーチェンバー）
• 環境保護・労働政策・その他の価値（インクルージョン、軍事利用）
• AI ガバナンス（AI ポリシー、倫理アセスメント、ステークホルダーの関与、監査、トレーサビリティ）

Learning resources

• G Kentei exam page (syllabus download) ↗
• G検定 exam guide / study guide ↗