【G検定 聞き流し 練習問題】③機械学習の概要|練習問題52問・解答解説付き ~2026年試験対応~ #ワイズデータ — Transcript
Full transcript
- 0:02In this video, I will present 52
- 0:04practice questions for Chapter 3:
- 0:06Overview of Machine Learning for the
- 0:08G-Certification.
- 0:10The questions are original, based on the
- 0:13G-Certification Syllabus 2024 Version
- 0:161.4, revised on May 11, 2026, by the
- 0:19Japan Deep Learning Association.
- 0:22We will cover four areas: supervised
- 0:25learning, unsupervised learning,
- 0:26reinforcement learning, and model
- 0:28selection and evaluation. After reading
- 0:31the question and the four options, I
- 0:32will provide 8 seconds of thinking time
- 0:34before presenting the correct answer and
- 0:36an explanation.
- 0:38You can learn by just listening or by
- 0:39watching the screen.
- 0:41Let's begin with Question 1.
- 0:45Question 1: Which of the following is
- 0:47the most appropriate description of
- 0:49training data in supervised learning?
- 0:52A: Provide only input data, with no
- 0:54correct labels given at all.
- 0:58B: Provide input data paired with the
- 1:00corresponding correct answer information
- 1:02.
- 1:04C: Provide only rewards, without
- 1:06providing any input data.
- 1:08D: Do not let the model memorize answers
- 1:10, but require human correction at every
- 1:12step.
- 1:22The correct answer to Question 1 is B:
- 1:24Provide input data paired with the
- 1:25corresponding correct answer information
- 1:27.
- 1:29In supervised learning, the model is
- 1:31trained using pairs of input features
- 1:33and the correct labels or target
- 1:35variables to be predicted.
- 1:37This is the fundamental concept for both
- 1:39classification and regression. A is
- 1:40similar to unsupervised learning, and C
- 1:42brings to mind the rewards used in
- 1:44reinforcement learning. The sequential
- 1:46human correction mentioned in D is not a
- 1:48required condition.
- 1:50Question 2: Which of the following
- 1:52supervised learning tasks generally
- 1:54qualifies as a regression problem?
- 1:58A: Predicting the price of a house based
- 2:01on its area, etc. B: Categorizing emails
- 2:04as spam or not spam.
- 2:06C: Sorting images of animals into dogs,
- 2:09cats, or birds. D: Determining whether
- 2:11or not a customer will cancel their
- 2:13subscription.
- 2:22The correct answer to Question 2 is A:
- 2:24Predicting the sale price based on
- 2:26housing area, etc. Regression is a task
- 2:28that predicts continuous numerical
- 2:30values. Typical examples include
- 2:32predicting housing prices, sales revenue
- 2:35, or temperature. B, C, and D are all
- 2:37classification problems that predict
- 2:39discrete categories.
- 2:42Question 3: Which is the most
- 2:44appropriate difference between simple
- 2:45linear regression and multiple linear
- 2:47regression?
- 2:49A: Simple regression performs
- 2:51classification, while multiple
- 2:53regression performs clustering. B: In
- 2:55simple regression, there is no target
- 2:57variable.
- 2:59C: Simple regression uses one
- 3:01explanatory variable, while multiple
- 3:03regression uses multiple.
- 3:05D: Simple regression is a linear model,
- 3:07while multiple regression is always a
- 3:09non-linear model.
- 3:19The correct answer for question 3 is C:
- 3:21simple regression uses one explanatory
- 3:23variable, while multiple regression uses
- 3:26several.
- 3:27Simple regression analysis uses one
- 3:28explanatory variable, while multiple
- 3:30regression analysis uses multiple
- 3:32variables to analyze the relationship
- 3:34with an objective variable. Both can be
- 3:37treated as linear regression.
- 3:39A misunderstands the analysis task.
- 3:41B denies the need for an objective
- 3:43variable in regression.
- 3:45D is incorrect in stating that multiple
- 3:47regression must be non-linear.
- 3:50Question 4: Which is the most
- 3:51appropriate application for logistic
- 3:53regression?
- 3:56A: Divide videos into clusters without
- 3:58training data.
- 4:00B: Maximize the discounted reward for
- 4:02the next action.
- 4:04C: Apply minimum information to
- 4:06continuous objective variables.
- 4:09D: Estimate the probability of a
- 4:11customer churning and classify whether
- 4:13they will churn or not.
- 4:23The correct answer for question 4 is D:
- 4:25Estimate the probability of a customer
- 4:27churning and classify whether they will
- 4:29churn or not. Logistic regression is a
- 4:31method that models the probability of an
- 4:33event occurring, especially for binary
- 4:35classification. Although the name
- 4:36contains "regression," it is widely used
- 4:38for classification. A is unsupervised
- 4:40learning, B is reinforcement learning,
- 4:42and C is a description close to standard
- 4:45linear regression.
- 4:47Question 5: Which is the most
- 4:48appropriate explanation for the basic
- 4:50mechanism of decision trees?
- 4:53A: Always project data into a random
- 4:55low-dimensional space.
- 4:58B: Branch data based on conditions for
- 5:00features to perform classification or
- 5:02prediction.
- 5:03C: Refers only to neural networks that
- 5:06propagate errors to subsequent layers. D
- 5:08: Select actions based solely on
- 5:10state-value functions.
- 5:20The correct answer for question 5 is B:
- 5:22Branch data based on conditions for
- 5:24features to perform classification or
- 5:26prediction.
- 5:28Decision trees are models that
- 5:29repeatedly branch data based on feature
- 5:31conditions to predict classes or
- 5:33numerical values.
- 5:34A is close to the explanation for
- 5:36dimensionality reduction, C is about
- 5:38neural network learning, and D is a
- 5:40topic of reinforcement learning.
- 5:44Question 6: Which is the most
- 5:46appropriate explanation of bootstrap
- 5:48sampling used in bagging?
- 5:52A: Create multiple training sets by
- 5:54sampling with replacement from the
- 5:56original data.
- 5:58B: Sort the original data only once and
- 6:01provide it to all models in the same
- 6:03order.
- 6:04C: Always remove teacher labels and
- 6:06change it to unsupervised learning.
- 6:10D: Exclude all samples one by one and
- 6:12learn only one model.
- 6:23The correct answer for Question 6 is A:
- 6:25Create multiple training sets by
- 6:27sampling with replacement from the
- 6:29original data.
- 6:31Bootstrap sampling performs sampling
- 6:33with replacement from the original
- 6:35dataset.
- 6:37In bagging, multiple models are trained
- 6:39on different samples and their
- 6:40predictions are aggregated.
- 6:42B does not mention sampling with
- 6:44replacement. C does not require the
- 6:46deletion of teacher labels. D differs in
- 6:48that it is limited to only one model.
- 6:51Question 7: Which is the most
- 6:53appropriate characteristic of random
- 6:55forests?
- 6:57A: Always train multiple linear
- 6:59regression models in sequence.
- 7:03B: Create only one decision tree and fix
- 7:05the features.
- 7:08C: Train multiple decision trees and
- 7:10aggregate their predictions.
- 7:13D: Always train by stacking deep neural
- 7:15networks.
- 7:25The correct answer for Question 7 is C:
- 7:27Train multiple decision trees and
- 7:29aggregate their predictions.
- 7:32Random forest is an ensemble method that
- 7:34uses multiple decision trees.
- 7:36It randomizes the data sampling and
- 7:38features used for branching to aggregate
- 7:40predictions from diverse trees.
- 7:43A does not necessarily use linear
- 7:44regression.
- 7:46B differs in that it uses only one tree
- 7:48, and D is not deep learning.
- 7:52Question 8: Which is the most
- 7:53appropriate comparison between bagging
- 7:55and boosting?
- 7:58A: Both cannot use teacher labels.
- 8:02B: Bagging is always one model, boosting
- 8:04always has an infinite number of models
- 8:07; C: Bagging is always a neural network
- 8:09.
- 8:11Boosting always uses decision trees.
- 8:14D: Bagging trains models in parallel for
- 8:16easy aggregation, while boosting often
- 8:18trains sequentially to compensate for
- 8:20previous weaknesses.
- 8:31The correct answer for Question 8 is D:
- 8:33Bagging trains models in parallel for
- 8:35easy aggregation, while boosting often
- 8:37trains sequentially to compensate for
- 8:39previous weaknesses.
- 8:41Bagging is a representative ensemble
- 8:43method that trains multiple models
- 8:45independently and aggregates their
- 8:46predictions.
- 8:48Boosting sequentially adds models to
- 8:49compensate for the errors of the
- 8:51previous models.
- 8:53A can be used in supervised learning.
- 8:55B misunderstands the number of models. C
- 8:58unnecessarily limits the models to be
- 9:00used.
- 9:02Question 9: What is the most appropriate
- 9:04basic concept of gradient boosting?
- 9:09A: Sequentially add weak learners to
- 9:11improve the loss function.
- 9:14B: Use the initially selected single
- 9:16predictor without modification.
- 9:19C: Delete all training data and perform
- 9:21clustering.
- 9:23D: Convert multiple features into
- 9:25principal components.
- 9:35The correct answer to Question 9 is A:
- 9:37Sequentially add weak learners to
- 9:39improve the loss function.
- 9:42Gradient boosting is an ensemble method
- 9:44that sequentially adds weak learners to
- 9:46improve predictions, using clues like
- 9:48the negative gradient of the loss
- 9:50function.
- 9:51B does not involve sequential model
- 9:53addition. C is unsupervised learning. D
- 9:55is a description close to principal
- 9:57component analysis.
- 10:00Question 10: Which is the most
- 10:01appropriate description of the
- 10:03fundamental concept of Support Vector
- 10:05Machines (SVM)?
- 10:09A: The sole purpose is to randomize the
- 10:11order of training data.
- 10:13B: Emphasize the distance, or margin,
- 10:15between the boundary separating classes
- 10:17and the nearest data points.
- 10:20C: Learn the discount rate for rewards
- 10:22from training data.
- 10:24D: Always classify all data into the
- 10:26same class.
- 10:35The correct answer to Question 10 is B:
- 10:37Emphasize the distance, or margin,
- 10:39between the boundary separating classes
- 10:41and the nearest data points. SVM is a
- 10:43classification method that uses the
- 10:45concept of maximizing the margin between
- 10:47the classification boundary and the data
- 10:49. Using soft margins allows for a
- 10:51certain degree of misclassification.
- 10:53A is not the objective of learning.
- 10:56C is a term from reinforcement learning
- 10:58, and D is inappropriate as a classifier
- 11:00.
- 11:01Question 11: Which is the most
- 11:03appropriate description of the kernel
- 11:05trick used in SVMs and other methods?
- 11:09A: Calculate by deleting all features.
- 11:13B: Guarantee learning without using
- 11:15training labels.
- 11:17C: Calculate inner products in a
- 11:19high-dimensional space without explicit
- 11:22mapping.
- 11:23D: Ensure that all data can always be
- 11:25separated by a straight line.
- 11:36The correct answer to Question 11 is C:
- 11:38Calculate inner products in a
- 11:39high-dimensional space without explicit
- 11:42mapping.
- 11:44Kernel functions allow for the direct
- 11:46calculation of inner products in a
- 11:47high-dimensional feature space.
- 11:50While useful for handling non-linear
- 11:52boundaries, it does not guarantee that
- 11:53all data can be perfectly separated.
- 11:56A does not delete features.
- 11:59B does not prohibit the use of labels.
- 12:01E does not guarantee separability.
- 12:05Question 12: Which of the following
- 12:07tasks best corresponds to multi-class
- 12:09classification?
- 12:11A Predict a customer's purchase amount
- 12:13for the next month.
- 12:15B Group customers by similarity using
- 12:17only their purchase history.
- 12:20C Learn a delivery route based on
- 12:22rewards.
- 12:23D Classify a handwritten digit image
- 12:26into one of ten classes, 0 through 9.
- 12:36The correct answer to question 12 is D,
- 12:39classify a handwritten digit image into
- 12:41one of ten classes, 0 through 9.
- 12:44Multi-class classification is a task of
- 12:46choosing one from three or more discrete
- 12:48classes. Handwritten digit 10-class
- 12:50classification is a typical example. A
- 12:52is regression, B is unsupervised
- 12:54clustering, and C is an application of
- 12:56reinforcement learning.
- 12:59Question 13: Which is the most
- 13:01appropriate description of the
- 13:03autoregressive (AR) model and vector
- 13:06autoregressive (VAR) model in
- 13:07time-series analysis?
- 13:11A. AR handles past values of one series
- 13:14, while VAR handles past values of
- 13:16multiple series.
- 13:19B AR is exclusively for image
- 13:21recognition and VAR is for speech
- 13:22synthesis.
- 13:25C AR is limited to unsupervised learning
- 13:28and VAR is limited to reinforcement
- 13:30learning.
- 13:32D Neither uses past values of time
- 13:34series.
- 13:44The correct answer to question 13 is A.
- 13:46AR handles past values of one series,
- 13:48while VAR handles past values of
- 13:50multiple series.
- 13:52The AR model explains the current value
- 13:54using past values of a single time
- 13:56series.
- 13:58The VAR model aggregates multiple time
- 14:00series and models them using the past
- 14:02values of each.
- 14:04B incorrectly limits the usage.
- 14:06C makes an inappropriate limitation
- 14:08regarding the learning method.
- 14:10D contradicts the concept of
- 14:11autoregression.
- 14:14Question 14: Which is the most
- 14:16appropriate benefit of ensemble learning
- 14:18, which combines multiple predictive
- 14:20models?
- 14:22A. It can guarantee a 100%accuracy rate
- 14:25for all training data.
- 14:28B. It can sometimes suppress the
- 14:30weaknesses or variance of individual
- 14:31models by combining their predictions.
- 14:35C. It always converts all models into
- 14:37random predictions.
- 14:39D. It can guarantee optimal predictions
- 14:41even without any training data.
- 14:52The correct answer to question 14 is B.
- 14:54It can sometimes suppress the weaknesses
- 14:56or variance of individual models by
- 14:58combining their predictions.
- 15:00Ensemble learning is a method of
- 15:01combining the predictions of multiple
- 15:03models.
- 15:04Bagging and boosting are representative
- 15:06examples, which can improve prediction
- 15:08accuracy and robustness depending on
- 15:10conditions. A and D are incorrect as
- 15:12they guarantee performance
- 15:13unconditionally. C is not a method that
- 15:16randomizes model predictions.
- 15:19Question 15.
- 15:20Which is the most appropriate
- 15:22description of unsupervised learning?
- 15:25A. It always minimizes the error against
- 15:27correct labels.
- 15:29B. It aims only to maximize rewards.
- 15:32C. It finds structures or regularities
- 15:35from data without correct labels.
- 15:37D. It predicts solely based on pre-set
- 15:40business rules.
- 15:49The correct answer to Question 15 is C.
- 15:51It finds structures or regularities from
- 15:53data without correct labels.
- 15:56Unsupervised learning typically learns
- 15:58cluster structures or low-dimensional
- 16:00representations from unlabeled data.
- 16:02Clustering and dimensionality reduction
- 16:04are typical examples.
- 16:06A describes supervised learning, B
- 16:08describes reinforcement learning, and D
- 16:10describes a rule-based system.
- 16:13Question 16.
- 16:15Which is the most appropriate
- 16:16description of the basic process of
- 16:18K-means?
- 16:20It alternately updates cluster centers
- 16:22and the assignment of each data point.
- 16:25B. It learns decision tree branching
- 16:27conditions based on correct labels.
- 16:30C. It selects actions by calculating
- 16:32future discounted rewards.
- 16:34D. It performs linear regression using
- 16:36only past values from a time series.
- 16:46The correct answer to Question 16 is
- 16:50alternately updating cluster centers and
- 16:52the assignment of each data point.
- 16:54In K-means, you set the number of
- 16:56clusters K, and repeat the process of
- 16:58assigning each data point to the nearest
- 17:00center and updating the centers. It is
- 17:02influenced by the choice of initial
- 17:03values and the number of clusters. B is
- 17:05about supervised learning, C is about
- 17:07reinforcement learning, and D is about
- 17:09autoregressive analysis.
- 17:12Question 17. Which of the following
- 17:14represents the results obtained from
- 17:15hierarchical clustering as a tree
- 17:17diagram along with the order of merging
- 17:19and distances?
- 17:23A. ROC curve, B. Confusion matrix, C.
- 17:26Decision boundary, D. Dendrogram.
- 17:38The correct answer to Question 17 is D.
- 17:40Dendrogram.
- 17:42A dendrogram is a diagram that
- 17:44visualizes the order and distance in
- 17:46which clusters are merged in
- 17:47hierarchical clustering.
- 17:49A represents evaluation of
- 17:51classification models, B represents
- 17:53aggregation of classification results,
- 17:54and C represents classifier boundaries;
- 17:56none are dendrograms for hierarchical
- 17:58clustering.
- 18:00Question 18. Which is the most
- 18:02appropriate description of Ward's method
- 18:04in hierarchical clustering?
- 18:08A. It always merges based solely on the
- 18:10maximum distance between clusters. B. It
- 18:12chooses the pair that results in the
- 18:14smallest increase in within-cluster sum
- 18:16of squares when merged. C. It maximizes
- 18:19the classification accuracy against
- 18:21external labels.
- 18:23D. It can only be used if principal
- 18:24components have been determined first.
- 18:35The correct answer to question 18 is B.
- 18:37Ward's method, which selects pairs that
- 18:39result in the smallest increase in
- 18:41within-cluster sum of squares, is a
- 18:43representative hierarchical clustering
- 18:45technique that proceeds based on the
- 18:47increase in within-cluster variance when
- 18:49two clusters are merged.
- 18:51A is an explanation close to the
- 18:53complete-linkage method. C is a
- 18:54supervised evaluation metric. D is
- 18:57incorrect because it requires Principal
- 18:59Component Analysis.
- 19:01Question 19: Which is the most
- 19:03appropriate main purpose of Principal
- 19:06Component Analysis (PCA)?
- 19:09A. Determine the action that maximizes
- 19:11rewards.
- 19:13B. Automatically add correct labels to
- 19:15observed data.
- 19:17C. Find directions that preserve as much
- 19:19data variance as possible to represent
- 19:21it in low dimensions.
- 19:23D. Always assign a specific theme name
- 19:26to unknown text.
- 19:35The correct answer to question 19 is C,
- 19:37finding directions that preserve as much
- 19:39data variance as possible to represent
- 19:41it in low dimensions.
- 19:43PCA is a linear dimensionality reduction
- 19:45method that finds directions with large
- 19:47data variance as principal components
- 19:49and aggregates information into fewer
- 19:51axes. A refers to reinforcement learning
- 19:53. B refers to labeling. D is related to
- 19:56text classification and is not the
- 19:58primary purpose of PCA.
- 20:00Question 20: Which is the most
- 20:02appropriate explanation for Singular
- 20:05Value Decomposition (SVD)?
- 20:07A. A method that decomposes a matrix
- 20:09into singular values and corresponding
- 20:11orthogonal matrices, also used for
- 20:13dimensionality reduction.
- 20:16B. A method to calculate rewards for
- 20:18multiple agents.
- 20:20C. An evaluation metric that separates
- 20:22classification results into true
- 20:23positives and false positives. D. A
- 20:25decision tree that always converts
- 20:27categories into a single value.
- 20:37The correct answer to question 20 is A.
- 20:39A method that decomposes a matrix into
- 20:42singular values and corresponding
- 20:43orthogonal matrices, also used for
- 20:45dimensionality reduction. SVD is a
- 20:47method that decomposes a matrix into a
- 20:49product of orthogonal matrices and a
- 20:51matrix containing singular values. It is
- 20:53used for dimensionality reduction and
- 20:55recommendations through low-rank
- 20:56approximation and other techniques. B
- 20:58relates to reinforcement learning, C to
- 21:00a confusion matrix, and D to a decision
- 21:02tree.
- 21:04Question 21: Which is the most
- 21:06appropriate purpose for using t-SNE?
- 21:10A. Directly calculate the accuracy of a
- 21:12classification model.
- 21:14B. Update action values in reinforcement
- 21:17learning.
- 21:18C. Predict sales based solely on past
- 21:21time-series values. D. Visualize
- 21:23high-dimensional data in 2D or 3D by
- 21:25emphasizing neighborhood relationships.
- 21:35The correct answer for Question 21 is D
- 21:37: Visualize in 2D or 3D while
- 21:39prioritizing the neighborhood
- 21:41relationships of high-dimensional data.
- 21:43t-SNE is a non-linear dimensionality
- 21:45reduction method used for visualization
- 21:47by representing local neighborhood
- 21:49relationships of high-dimensional data
- 21:50in a low-dimensional space. Care should
- 21:52be taken when interpreting the distance
- 21:54between distant clusters. A is about
- 21:56classification evaluation, B is
- 21:58reinforcement learning, and C is about
- 22:00prediction by autoregression.
- 22:04Question 22: Which is the most
- 22:06appropriate description of
- 22:07Multidimensional Scaling (MDS)?
- 22:11A: Create decision trees equal to the
- 22:13number of data types. B: Arrange in a
- 22:15low-dimensional space to reproduce the
- 22:17relationships of dissimilarity or
- 22:19distance between objects as much as
- 22:21possible. C: Always assign correct
- 22:23labels to data that lacks teacher labels
- 22:25.
- 22:26B: Determine the next action using an
- 22:28epsilon-greedy policy.
- 22:38The correct answer for Question 22 is B
- 22:40: Arrange in a low-dimensional space to
- 22:42reproduce the relationships of
- 22:43dissimilarity or distance between
- 22:45objects as much as possible. MDS is a
- 22:47method for finding a low-dimensional
- 22:48arrangement that preserves the
- 22:50relationships between objects based on
- 22:51their distances or dissimilarities as
- 22:53much as possible. It differs from PCA in
- 22:55terms of input and the criteria to be
- 22:57optimized.
- 22:58A does not describe an ensemble either.
- 23:00C is labeling, and D is action selection
- 23:03in reinforcement learning.
- 23:06Question 23: Which is the most
- 23:08appropriate description for using Latent
- 23:11Dirichlet Allocation (LDA) in document
- 23:13analysis?
- 23:16A: Assign every document to only one
- 23:18identical topic.
- 23:20B: It is a classifier that only
- 23:22determines the correctness of each
- 23:24document.
- 23:25C: Model documents as a mixture of
- 23:27multiple topics, where words appear for
- 23:30each topic.
- 23:32D: It is a method for resizing images.
- 23:44The correct answer for Question 23 is C
- 23:46: Model documents as a mixture of
- 23:48multiple topics, where words appear for
- 23:50each topic.
- 23:52The topic model LDA is a generative
- 23:54model that assumes each document has a
- 23:56mixture ratio of multiple topics, and
- 23:58each topic has a distribution of word
- 24:00occurrences.
- 24:01A denies the mixture of multiple topics
- 24:03. B is binary classification. D is a
- 24:06description of image processing.
- 24:09Question 24: Which is the most
- 24:11appropriate description of collaborative
- 24:13filtering used for recommendations on EC
- 24:15sites?
- 24:17A: Recommend products by utilizing user
- 24:19behavior history and similarity in
- 24:21preferences.
- 24:23B: Recommend by analyzing only the words
- 24:26in product descriptions.
- 24:27C: Display product names by sorting them
- 24:30in alphabetical order.
- 24:32D Change all products to the same price
- 24:34without using teacher labels.
- 24:45The answer to Question 24 is A:
- 24:47Recommend products using user behavior
- 24:49history and similarity in preferences.
- 24:52Collaborative filtering provides
- 24:54recommendations using interactions
- 24:56between users and products, such as
- 24:58ratings, views, and purchases. You can
- 25:00reference the behavior of similar users
- 25:02or items.
- 25:03B is close to content-based filtering, C
- 25:06is sorting, and D is price changing;
- 25:08none are recommendation methods.
- 25:12Question 25: Which is the most
- 25:14appropriate description of the
- 25:15cold-start problem in recommendation
- 25:17systems?
- 25:20A: The problem where all recommended
- 25:22products are physically cold.
- 25:24B: The problem where all user ratings
- 25:26are too high to calculate errors. C: The
- 25:28problem where a model's temperature
- 25:30parameter cannot be initialized.
- 25:33D: The problem where it is difficult to
- 25:35ensure recommendation accuracy for new
- 25:37users or products due to a lack of
- 25:39history.
- 25:48The answer to Question 25 is D: It is
- 25:50difficult to ensure recommendation
- 25:52accuracy for new users or products due
- 25:55to a lack of history.
- 25:57The cold-start problem is the difficulty
- 25:59in making recommendations because there
- 26:01is not enough behavior history or
- 26:03ratings for new users or items.
- 26:06Content-based methods that use product
- 26:07attribute information can be a
- 26:09supplement.
- 26:10A is a literal translation of the term,
- 26:12B is about evaluation metrics, and C is
- 26:14about different parameters; all differ
- 26:16from the recommendation cold-start
- 26:18problem.
- 26:19Question 26: Which is the most
- 26:21appropriate description of the basic
- 26:23structure of reinforcement learning?
- 26:27A: Learn using only pairs of correct
- 26:29labels and input images.
- 26:31B: An agent acts in an environment and
- 26:33learns a policy based on the rewards it
- 26:35receives.
- 26:37C: Unsupervised clustering of known text
- 26:40.
- 26:42D: Human-fixed agent actions without
- 26:44using rewards.
- 26:54The answer to Question 26 is B: An agent
- 26:56acts in an environment and learns a
- 26:58policy based on the rewards it receives
- 27:00.
- 27:01In reinforcement learning, the agent
- 27:03observes the state of the environment,
- 27:05takes actions, and receives rewards. It
- 27:07learns a policy for actions that
- 27:09maximizes the cumulative future reward.
- 27:12A is supervised learning, C is
- 27:13unsupervised learning, and D differs
- 27:15from the reinforcement learning concept
- 27:17that uses rewards.
- 27:20Question 27: What is the most
- 27:22appropriate combination of the basic
- 27:24elements of a Markov Decision Process (
- 27:27MDP)?
- 27:31A: State, action, state transition,
- 27:33reward, discount rate B: Teacher label,
- 27:36regression coefficient, mean squared
- 27:39error, decision tree C: Principal
- 27:41component, singular value, cluster
- 27:44center, dendrogram D: Precision, recall
- 27:47, confusion matrix,
- 27:51AUC
- 28:00The correct answer for Question 27 is A.
- 28:02State, action, state transition, reward
- 28:05, and discount rate.
- 28:07An MDP represents an environment using
- 28:10states, actions, transition
- 28:11probabilities, and rewards, and uses a
- 28:14discount rate to evaluate future rewards
- 28:16.
- 28:17This is the fundamental framework of
- 28:18reinforcement learning. B is mainly
- 28:20related to supervised learning, C to
- 28:22unsupervised learning, and D relates to
- 28:24evaluation metrics for classification
- 28:26models.
- 28:27Question 28: Which is the most
- 28:29appropriate description of the
- 28:31difference between the state-value
- 28:33function V (s) and the action-value
- 28:35function Q (s, a) in reinforcement
- 28:36learning?
- 28:38A: V does not consider rewards, and Q
- 28:41only uses today's rewards.
- 28:43B: V evaluates only actions, and Q
- 28:46evaluates only states.
- 28:48C: V represents the value of a state,
- 28:50and Q represents the value of a
- 28:52state-action pair.
- 28:54D: V and Q are always the same value, so
- 28:56there is no point in distinguishing them
- 28:58.
- 29:07The correct answer for Question 28 is C.
- 29:09V represents the value of a state, and Q
- 29:11represents the value of a state-action
- 29:13pair. The state-value V (s) represents
- 29:16the expected future reward from a state
- 29:18, while the action-value Q (s, a)
- 29:20represents the expected future reward
- 29:22when choosing a specific action in that
- 29:24state.
- 29:25A's explanation of learning information
- 29:27is incorrect. B's objects of evaluation
- 29:29are reversed.
- 29:32D denies the distinction between V and
- 29:34Q. Question 29: In reinforcement
- 29:36learning, where the discount rate gamma
- 29:38is set between 0 and 1, which is the
- 29:40most appropriate general effect of
- 29:42making gamma smaller?
- 29:45A: It becomes easier to prioritize
- 29:47future rewards over current rewards.
- 29:50B: Rewards are no longer used at all
- 29:52during learning.
- 29:53C: The number of states automatically
- 29:55becomes zero. D: The contribution of
- 29:58rewards in the distant future becomes
- 30:00relatively smaller.
- 30:09The correct answer for Question 29 is D
- 30:11: The contribution of rewards in the
- 30:13distant future becomes relatively
- 30:14smaller. With a smaller discount rate,
- 30:17the further away in time a reward is,
- 30:18the smaller its contribution to the
- 30:20present value, meaning rewards in the
- 30:22near future are relatively prioritized.
- 30:24A represents a line of thinking closer
- 30:26to when the discount rate is large. B
- 30:28does not mean that rewards are not used
- 30:30, and C is unrelated to the number of
- 30:32states.
- 30:35Question 30: Which is the most
- 30:36appropriate characteristic of Q-learning
- 30:39?
- 30:41A: Q-values are updated using the
- 30:43maximum value of the action-value in the
- 30:45next state.
- 30:47B: Each piece of data is always assigned
- 30:49to a specific class.
- 30:51C: Only the cross-entropy for the
- 30:53correct label is calculated.
- 30:55D: The variance in the direction of the
- 30:57principal component is maximized.
- 31:08The correct answer for Question 30 is A
- 31:10: Q-values are updated using the maximum
- 31:12value of the action-value in the next
- 31:14state.
- 31:15Q-learning is a representative
- 31:17value-based reinforcement learning
- 31:18method that updates Q-values using the
- 31:20maximum action value obtained in the
- 31:22next state. It is an off-policy method
- 31:24where the behavior policy and the target
- 31:26policy can differ. B describes K-means
- 31:29clustering, C describes loss in
- 31:31supervised learning, and D describes PCA
- 31:33.
- 31:35Question 31: Which is the most
- 31:37appropriate explanation for Q-value
- 31:39updating in SARSA?
- 31:41A: Updates by assuming only the best
- 31:43action in the next state.
- 31:46B: Updates using the Q-value of the next
- 31:48action actually selected.
- 31:51C: Updates using only teacher labels
- 31:53without using Q-values for all actions.
- 31:57D: Selects the same action every time
- 31:58without observing the state.
- 32:09The correct answer to Question 31 is B:
- 32:11Updates using the Q-value of the next
- 32:13action actually selected.
- 32:16SARSA is a representative on-policy
- 32:18method that updates using information on
- 32:20the current state, action, reward, next
- 32:23state, and the next action actually
- 32:25selected.
- 32:26A is an explanation close to the update
- 32:28rule of Q-learning.
- 32:30C confuses reinforcement learning with
- 32:32supervised learning, and D does not
- 32:34describe policy-based learning.
- 32:39Question 32: Which is the most
- 32:41appropriate explanation of the
- 32:43epsilon-greedy policy?
- 32:48A: Always chooses the action with the
- 32:50least reward.
- 32:51B: Always executes all observed actions
- 32:54simultaneously.
- 32:56C: Usually chooses the action with high
- 32:58estimated value and tries a random
- 33:00action with a certain probability.
- 33:03D: Does not change the action once
- 33:05chosen thereafter.
- 33:15The correct answer to Question 32 is C:
- 33:17Usually chooses the action with high
- 33:19estimated value and tries a random
- 33:21action with a certain probability.
- 33:24The epsilon-greedy policy usually
- 33:26selects the action estimated to be best
- 33:28currently, while choosing a random
- 33:30action with probability epsilon to
- 33:32balance exploration and exploitation.
- 33:35A is not a strategy to increase value.
- 33:38B does not assume simultaneous execution
- 33:40. D is an explanation that eliminates
- 33:42exploration.
- 33:45Question 33: Which is the most
- 33:46appropriate feature of the UCB policy
- 33:49used in the bandit problem?
- 33:52A: Never chooses untried actions.
- 33:56B: Uses only the average of estimated
- 33:58rewards for each action, ignoring the
- 34:00number of trials.
- 34:02C: Uses only the classification accuracy
- 34:04of correct labels as a criterion. D:
- 34:07Selects actions by adding an upper bound
- 34:09that considers uncertainty to the
- 34:11estimated reward.
- 34:20The correct answer for Question 33 is B
- 34:22: Select actions by adding an upper
- 34:24bound that accounts for uncertainty to
- 34:26the estimated reward.
- 34:29UCB stands for Upper Confidence Bound.
- 34:32By adding a bonus for uncertainty to the
- 34:35estimated reward, it allows for the
- 34:36exploration of less-understood actions.
- 34:39A hinders exploration.
- 34:41B ignores uncertainty derived from the
- 34:43number of trials, and C is a
- 34:45classification metric, not a bandit
- 34:47action selection method.
- 34:49Question 34: Which of the following is
- 34:51the most appropriate description of the
- 34:53fundamental characteristics of the
- 34:54multi-armed bandit problem?
- 34:57A: Balance the exploration of uncertain
- 34:59rewards with the exploitation of
- 35:01promising choices.
- 35:04B: Always predict image labels.
- 35:07C: Perform regression on multiple time
- 35:10series simultaneously.
- 35:12D: Update only one cluster center.
- 35:24The correct answer for Question 34 is A
- 35:26: Balance the exploration of uncertain
- 35:28rewards with the exploitation of
- 35:30promising choices.
- 35:33In multi-armed bandit problems, because
- 35:35it is unclear which option yields a
- 35:37higher reward, balancing exploration and
- 35:39exploitation is crucial. An example is
- 35:41selecting which advertisements to
- 35:43display. B describes image
- 35:45classification, C describes the VAR
- 35:47model, and D is close to the K-means
- 35:49method.
- 35:50Question 35: Which of the following is
- 35:52the most appropriate description
- 35:54regarding policy gradient methods?
- 35:57A: Always combine decision trees for
- 35:59classification.
- 36:01B: Parameterize the policy and directly
- 36:03update it in the direction that
- 36:05increases expected reward. C: Find only
- 36:08the axes where the variance of the
- 36:09acquired quantities is maximized.
- 36:12D: Learn multiple decision trees using
- 36:14bootstrap sampling of training data.
- 36:25The correct answer for Question 35 is B
- 36:27: Parameterize the policy and directly
- 36:29update it in the direction that
- 36:31increases expected reward. Policy
- 36:33gradient methods are techniques for
- 36:34optimizing by parameterizing the policy
- 36:36used to select actions itself.
- 36:38REINFORCE is a representative policy
- 36:40gradient method that uses episode
- 36:42rewards.
- 36:43A and D describe decision trees and
- 36:45bagging, while C describes PCA.
- 36:49Question 36: Which of the following is
- 36:51the most appropriate description of the
- 36:53actor-critic method?
- 36:56A: The actor defines the reward, and the
- 36:58critic deletes the dataset.
- 37:01B: The actor and the critic always play
- 37:03the same role.
- 37:05C: The actor learns the action policy,
- 37:07and the critic supports learning by
- 37:09evaluating values.
- 37:11D: The actor performs clustering, and
- 37:13the critic performs dimensionality
- 37:15reduction.
- 37:24The correct answer for Question 36 is C
- 37:26: The actor learns a policy for actions
- 37:28, and the critic supports learning by
- 37:30evaluating values.
- 37:33In the actor-critic method, the actor
- 37:35learns the policy for choosing actions,
- 37:37while the critic estimates state values
- 37:39to assist in updating the actor.
- 37:42It is a method that combines value-based
- 37:44and policy-based approaches.
- 37:46A incorrectly describes the roles. B
- 37:48contradicts the point that the roles are
- 37:50divided.
- 37:50D is confused with unsupervised learning
- 37:53.
- 37:54Question 37: Which is the most
- 37:56appropriate task for applying
- 37:58reinforcement learning?
- 38:00A: A process that simply follows a fixed
- 38:02rule, such as always turning on the heat
- 38:04when the temperature is low.
- 38:07B: Classifying images using only
- 38:09provided labels.
- 38:11C: Compressing data into two dimensions
- 38:14while preserving variance.
- 38:16D: A robot improving its movement
- 38:18strategy through trial and error and
- 38:20rewards.
- 38:29The correct answer for Question 37 is D
- 38:31: A robot improving its movement
- 38:33strategy through trial and error and
- 38:35rewards.
- 38:36Reinforcement learning is suitable for
- 38:38tasks where actions affect the
- 38:39environment, and rewards obtained as a
- 38:41result are used to improve continuous
- 38:43decision-making.
- 38:45A is a simple control rule, B is
- 38:47supervised classification, and C is
- 38:49dimensionality reduction such as PCA.
- 38:53Question 38: What is the most
- 38:54appropriate description of the basics of
- 38:56holdout validation?
- 38:58A: Performing training and final
- 39:00evaluation using only the same data.
- 39:03B: Dividing data into training and
- 39:05evaluation sets, then checking
- 39:07performance with unused data.
- 39:10C: Making random predictions without
- 39:12performing model training.
- 39:14D: Adding the correct answers from the
- 39:16evaluation data to the training features
- 39:18.
- 39:27The correct answer for Question 38 is B
- 39:29: Dividing data into training and
- 39:31evaluation sets, then checking
- 39:33performance with unused data.
- 39:36Holdout validation is a method of
- 39:37splitting data into training and
- 39:39evaluation sets to estimate
- 39:40generalization performance using data
- 39:42not used for training.
- 39:45A easily overlooks overfitting to
- 39:46training data, and C does not learn the
- 39:48model.
- 39:50D causes data leakage.
- 39:54Question 39: What is the most
- 39:55appropriate description of K-fold
- 39:57cross-validation?
- 39:59A: Dividing data into K partitions, and
- 40:02for each iteration, using one partition
- 40:04for validation and the rest for training
- 40:06to repeat the evaluation.
- 40:08B: Training K times without using any
- 40:10validation data at all.
- 40:12C: Using the same partition for both
- 40:14training and validation every time.
- 40:17Always ensure the sum of the K
- 40:19evaluation metrics equals 100 points.
- 40:30The correct answer to Question 39 is A:
- 40:32Divide the data into K parts, and repeat
- 40:34the evaluation using one part for
- 40:36validation and the rest for training
- 40:38each time.
- 40:40In K-fold cross-validation, data is
- 40:42split into K parts, and we train and
- 40:44evaluate K times while rotating the
- 40:46validation partition, then aggregate the
- 40:49results.
- 40:50B does not satisfy the purpose of
- 40:52cross-validation. C does not separate
- 40:54training and validation within the same
- 40:56iteration. D is an incorrect rule
- 40:58regarding evaluation metrics.
- 41:00Question 40: Which statement best
- 41:02describes the relationship between
- 41:04training error and generalization error
- 41:06?
- 41:07A: If the training error is low, the
- 41:09error on unknown data will also be low.
- 41:12B: Both are always identical by
- 41:14definition.
- 41:15C: Even if training error is small,
- 41:18generalization error on unknown data may
- 41:20be large.
- 41:21D: Generalization error does not need to
- 41:23be considered for trained models.
- 41:34The correct answer to Question 40 is C:
- 41:36Even if training error is small,
- 41:38generalization error on unknown data may
- 41:40be large.
- 41:42If a model overfits to the training data
- 41:44, it may have low training error but
- 41:45fail to perform well on unknown data.
- 41:48This is a classic case of overfitting.
- 41:51A and B equate training performance with
- 41:53generalization performance. D denies the
- 41:55importance of generalization performance
- 41:57, which is vital in real-world
- 41:58operations.
- 41:59Question 41: Which of the following
- 42:02corresponds to a false positive (FP) in
- 42:04binary classification?
- 42:07A: Actually positive and predicted as
- 42:09positive.
- 42:11B: Actually negative and predicted as
- 42:13negative.
- 42:15C: Actually positive but predicted as
- 42:17negative.
- 42:19D: Actually negative but predicted as
- 42:21positive.
- 42:31The correct answer to Question 41 is D:
- 42:33Actually negative but predicted as
- 42:35positive.
- 42:38A false positive is a case where the
- 42:40reality is negative, but the model
- 42:42predicts it as positive.
- 42:44FP affects calculations such as
- 42:46precision.
- 42:48A is a true positive (TP), B is a true
- 42:51negative (TN), and C is a false negative
- 42:54(FN).
- 42:56Question 42.
- 42:58Out of 20 cases that a classification
- 43:00model predicted as positive, 15 were
- 43:03truly positive.
- 43:05What is the precision in this case?
- 43:10A: 60%, B: 75%, C: 80%, D: 93.75%
- 43:27The correct answer to Question 42 is B:
- 43:3175%.
- 43:32Precision is the ratio of actual
- 43:34positives among those predicted as
- 43:35positive.
- 43:3715 divided by 20 equals 0.75. Therefore
- 43:41, it is 75%. A, C, and D all differ from
- 43:44the result of calculating 15 divided by
- 43:4620.
- 43:47Question 43.
- 43:49There are 40 cases that are actually
- 43:51positive, and of those, the model
- 43:53correctly predicted 30 as positive.
- 43:56What is the recall in this case?
- 44:01A: 75%B: 60%C: 80%D: 90%
- 44:17The correct answer to Question 43 is A,
- 44:2075%.
- 44:23Recall is the proportion of actual
- 44:24positives that were correctly predicted
- 44:26as positive.
- 44:2930 divided by 40 equals 0.75, so it is
- 44:3375%. B, C, and D are all different from
- 44:35the result of 30 divided by 40.
- 44:38Question 44. A binary classification
- 44:42model has a precision of 0.8 and a
- 44:44recall of 0.5.
- 44:46Which value is closest to the F1-score,
- 44:48the harmonic mean?
- 44:52A: 0.4 B: 0.5 C: 0.615 D: 0.8
- 45:08The correct answer to Question 44 is C,
- 45:120.615.
- 45:13The F1-score is 2 times precision times
- 45:15recall divided by precision plus recall
- 45:18.
- 45:192 times 0.8 times 0.5 divided by 1.3
- 45:22equals approximately 0.615. A, B, and D
- 45:25do not match the result of the harmonic
- 45:27mean calculation.
- 45:30Question 45. In a binary classification
- 45:33task using data where 99%of the total is
- 45:35negative, which is the most appropriate
- 45:37evaluation if the model always predicts
- 45:40negative?
- 45:42A: Since the accuracy is high, the
- 45:44performance in finding positives must
- 45:46also be high.
- 45:48B: Since the accuracy is high, there are
- 45:51no false negatives.
- 45:53C: The accuracy will always be 50%.
- 45:57D: Even if the accuracy is 99%, the
- 46:00recall for positives will be 0%.
- 46:12The correct answer to Question 45 is D:
- 46:15Even if the accuracy is 99%, the recall
- 46:18for positives will be 0%.
- 46:21If you always predict negative, you get
- 46:2499%of the total correct, but since you
- 46:26find zero actual positives, the recall
- 46:29is 0%.
- 46:30In class imbalance, the selection of
- 46:32metrics is important. A and B evaluate
- 46:35the positive detection ability
- 46:36incorrectly.
- 46:38C ignores the condition that 99%are
- 46:40negative.
- 46:43Question 46. Which is the most
- 46:45appropriate explanation for the ROC
- 46:47curve and AUC?
- 46:50A: The ROC curve shows the relationship
- 46:52between false positive rate and true
- 46:54positive rate, and AUC is the area under
- 46:57that curve.
- 46:58B: The ROC curve only shows the
- 47:00residuals of regression problems.
- 47:03C: AUC represents the training time of
- 47:05the classification model itself.
- 47:08D: The ROC curve does not change when
- 47:10the prediction probability threshold is
- 47:12altered.
- 47:21The correct answer to Question 46 is A:
- 47:24The ROC curve shows the relationship
- 47:26between false positive rate and true
- 47:28positive rate, and AUC is the area under
- 47:30that curve.
- 47:32The ROC curve shows the relationship
- 47:34between the false positive rate and the
- 47:36true positive rate as the classification
- 47:38threshold changes, and the AUC is the
- 47:40area under that curve.
- 47:42B is a confusion with regression
- 47:43residuals. C refers to a different
- 47:45evaluation target. D rejects the ROC
- 47:48concept of plotting points by varying
- 47:50the threshold.
- 47:53Question 47: If a regression model's
- 47:56predicted value is 24 and the actual
- 47:58value is 17, what is the Mean Squared
- 48:01Error (MSE)?
- 48:05A: 2, B: 4, C: 5, D: 10
- 48:20The correct answer to Question 47 is C:
- 48:225. The errors are 1 and -3, and the
- 48:25squared errors are 1 and 9.
- 48:27Averaging them gives (1 + 9)/2 = 5, so
- 48:31the MSE is 5.
- 48:33A, B, and D do not match the average of
- 48:35the squared errors.
- 48:38Question 48: Which is the most
- 48:40appropriate description of RMSE and MAE
- 48:42in regression models?
- 48:46A: RMSE is the mean of errors, and MAE
- 48:48is the square root of errors.
- 48:53B: RMSE is the root mean squared error,
- 48:55and MAE is the mean absolute error.
- 48:59C: RMSE and MAE are both precision
- 49:01metrics for classification models.
- 49:05D: RMSE is always smaller than MAE.
- 49:18The correct answer to Question 48 is B:
- 49:20RMSE is the root mean squared error, and
- 49:23MAE is the mean absolute error.
- 49:27RMSE is the square root of MSE, while
- 49:29MAE is the average of absolute errors.
- 49:33For the same data and error units, RMSE
- 49:35is greater than or equal to MAE and
- 49:36gives relatively more weight to large
- 49:38errors.
- 49:40A is incorrect in its definition. C
- 49:41confuses regression evaluation with
- 49:43classification evaluation, and D has the
- 49:46size relationship between RMSE and MAE
- 49:48reversed.
- 49:51Question 49: Which is the most
- 49:52appropriate common concept for the
- 49:55Akaike Information Criterion (AIC) and
- 49:57the Bayesian Information Criterion (BIC
- 49:59)?
- 50:02A: More complex models are
- 50:03unconditionally rated higher.
- 50:06B: Unlabeled data must be classified
- 50:09into clusters.
- 50:10C: Models are evaluated solely based on
- 50:13prediction accuracy.
- 50:15D: Models are compared by considering
- 50:17both goodness-of-fit and model
- 50:19complexity.
- 50:28The correct answer to Question 49 is D:
- 50:30Models are compared by considering both
- 50:32goodness-of-fit and model complexity.
- 50:35AIC and BIC compare models by
- 50:37considering not only the fit to the data
- 50:39but also penalties for model complexity
- 50:41, such as the number of parameters.
- 50:44Generally, smaller values are preferred
- 50:46. A is the opposite of the penalty for
- 50:48complexity. B differs in that it
- 50:50concerns clustering, and C is limited to
- 50:52classification accuracy.
- 50:55Question 50: Which of the following is
- 50:57the most appropriate description of
- 50:59Occam's razor in machine learning model
- 51:01selection?
- 51:04A: If explanatory power is similar,
- 51:06prioritize considering a simpler model.
- 51:10B: The more parameters there are, the
- 51:12higher the generalization performance
- 51:14always becomes. C: If the model is made
- 51:16more complex, validation data becomes
- 51:19unnecessary. D: Outliers included in the
- 51:21training data must always be deleted.
- 51:32The correct answer to Question 50 is A:
- 51:35If explanatory power is similar,
- 51:37prioritize considering a simpler model.
- 51:40Occam's razor is the concept that if
- 51:42explanatory power is similar, one should
- 51:44not unnecessarily increase complex
- 51:46processes. It serves as a guideline that
- 51:48emphasizes simplicity in model selection
- 51:49as well. B is incorrect because
- 51:51complexity does not guarantee
- 51:52generalization performance. C does not
- 51:54make validation unnecessary. D is not a
- 51:57mandatory rule for outlier processing.
- 52:00Question 51: The accuracy on training
- 52:03data was 99%, while the accuracy on
- 52:06validation data was 65%.
- 52:09Which of the following is the most
- 52:10appropriate phenomenon to suspect first
- 52:12?
- 52:14A: Automatic conversion to unsupervised
- 52:17learning, B: Convergence of K-means, C:
- 52:19Drop in generalization performance due
- 52:22to overfitting, D: Change in the
- 52:24definition of the ROC curve.
- 52:34The correct answer to Question 51 is C:
- 52:37Drop in generalization performance due
- 52:39to overfitting. Since it shows good
- 52:41performance on training data but worsens
- 52:43significantly on validation data,
- 52:45overfitting to the training data is
- 52:46suspected.
- 52:47Additional investigation into
- 52:49differences in data distribution or
- 52:51leakage is also necessary. A, B, and D
- 52:53are not the primary phenomena directly
- 52:55suspected from this performance gap.
- 52:59Question 52: Which evaluation metric
- 53:01should be particularly prioritized if
- 53:03you want to minimize misses (false
- 53:05negatives) in primary cancer screening?
- 53:09A: Inference time only.
- 53:12B: Recall, which equals sensitivity.
- 53:16C: Mean Absolute Error (MAE).
- 53:20D: Akaike Information Criterion (AIC)
- 53:23only.
- 53:32The correct answer to Question 52 is B:
- 53:35Recall, which equals sensitivity.
- 53:38Prioritizing recall, which is the rate
- 53:40of correctly detecting actually positive
- 53:42individuals as positive, leads to
- 53:43evaluation that focuses on reducing
- 53:45false negatives. However, false
- 53:47positives and operational constraints
- 53:49should also be considered together.
- 53:51A does not directly represent the miss
- 53:52rate.
- 53:54C is a regression metric, and D is a
- 53:56model selection metric; neither is
- 53:58direct for evaluating misses in positive
- 54:00detection.
About this transcript
This page contains the full transcript of 【G検定 聞き流し 練習問題】③機械学習の概要|練習問題52問・解答解説付き ~2026年試験対応~ #ワイズデータ by ワイズデータ [聞き流しで受かるIT・データ資格チャンネル], generated from the public captions YouTube serves with the video. The transcript has 6,665 words across 1,289 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.