![見出し画像](https://assets.st-note.com/production/uploads/images/93205158/rectangle_large_type_2_b45fdafeeddd308292bfe085375b149e.jpeg?width=1200)
人工知能検定3級の例題
Q7. ensemble learning
Read the following statement and choose the option that best fits the blank.
Ensemble learning is a learning method that integrates multiple weak learners to build one highly accurate learner to improve generalization performance. In (a), one of these ensemble learning methods, each weak learner is constructed for multiple datasets recovered and extracted from the training data, and their training results are integrated to produce the final result. The (b) effect is achieved by adding diversity to the data sets. This representative model, (c), uses a decision tree as the weak learner.
The options for (a)
Bagging
Bootstrapping
Boosting
Stacking
Option (b)
Lowering the variance of the model's predictions
Decrease the bias between model predictions and correct labels
Decrease the covariate shift
Make weights a sparse solution
Alternative (c).
Random forest
SVM
Gradient Boosting Tree
AdaBoost
Q9. ingenuity in deep learning learning
Choose the inappropriate method to solve various challenges in deep learning (DL) learning.
One method to prevent overlearning is dropout, in which a certain percentage of nodes are randomly eliminated.
Models trained using dropout need to store which nodes were used for learning because inference is performed using the nodes that were not eliminated during training.
One method to prevent overlearning is early stopping.
Since early termination of the optimization of the loss function can be regarded as equivalent to constraining the weight parameters, early stopping is said to have a regularization effect equivalent to a norm penalty.
Batch normalization is used to solve the phenomenon that learning becomes difficult due to changes in the input distributions of each layer during training (covariate shift).
In batch processing, where all the training data are trained at once, if the data size is huge, the DL calculation volume will increase and it will be difficult to make progress in updating the weights.
Therefore, mini-batch learning divides the training data into several groups for learning, which makes the frequency of weight updates finer and speeds up learning convergence.
Q10. CNN
CNN is a neural network that includes convolutional processing and is a well-known technology applied to the image processing field. Choose the most inappropriate one from the following sentences about CNN.
Convolutional processing is a technique for extracting features in an image using kernels (filters), and before CNN, convolutional processing used artificially designed kernels. Before CNNs, convolutional processing used artificially designed kernels. However, convolutional processing in CNNs can extract features with higher accuracy by learning numerical values in the kernels through back-propagation.
There is a process called pooling that compresses the input size of images and feature maps. This process has the effect of being invariant to slight differences in the color and tone of objects in the image without requiring parameters to be learned.
In the 2014 paper in which VGG was proposed, it was shown that by repeating the convolution process of only 3x3 kernels many times, it is possible to cover the same receptive field as a large kernel size, while keeping the number of parameters low.
The technique that solves the "gradient loss problem due to multilayering," which is a problem faced by VGG, is the skip structure (jump structure, skip connection) proposed in the 2015 ResNet paper. ResNet has achieved a super multi-layered structure with 152 layers by learning the difference between the input and output of each Residual Block, which is a component of ResNet.
Q12. Image Processing
Read the following passage and choose the option that best fits the blank.
In recent years, deep learning has been very active in the field of image processing. Among them, segmentation is a (a) technology that has attracted attention for its application to medical image diagnosis and automatic driving. Deep learning has been playing an active role in the field of image processing in recent years. Among these, segmentation is a (a) technology that has attracted attention for its application to medical image diagnosis and automatic driving. A typical model of segmentation is (b). (An important feature of (b) is that it replaces the entire network with a convolution process by introducing an inverse convolution process and complements spatial information with shortcut connections. On the other hand, object detection is (c) technology, which is attracting attention as an analysis technology for automatic driving technology and surveillance cameras. Its representative model, (d), has overturned the conventional wisdom of object detection. (Prior to (d), object detection models had a step-by-step architecture, with a phase to detect the likely location of an object, followed by a phase to identify what that object is. However, (d) made it possible to perform the detection and identification phases simultaneously by dividing the input image into N × N (N = 7 in the paper) cells, thereby increasing the recognition speed 10 to 20 times faster than before.
(a) (c) Alternative (c)
Generate images by encoder-decoder architecture
Predicting regions or classifying objects in the image pixel by pixel
Classify what the image itself shows
Enclose and identify objects and their locations in the image using rectangular regions
(b) (d) Alternative
SegNet
U-Net
Faster R-CNN
YOLO
Q13. natural language processing field
Natural language processing is a technology for quantifying natural language (language used by humans for communication) in a way that makes it easier for computers to analyze. In recent years, the development of models using Attention in particular has been remarkable. Which of the following sentences related to Attention is false?
The model architecture of Seq2Seq applied to machine translation is divided into encoder (a model that processes input sentences) and decoder (a model that predicts output sentences), and Attention is the mechanism by which the decoder refers to and uses the encoder's output information.
Transformer is a model that uses only Attention, and has many advantages over conventional RNN/LSTM-based models, such as the ability to make effective use of the GPU.
OpenAI GPT is an architecture that applies a part of Transformer to word models, but it has some issues such as not being able to refer to future information in time series.
BERT is known as an interactive Transformer model and is characterized by its pre-training. Among them, the Masked Language Model (MLM) is famous for its task of determining whether two sentences are consecutive or not, given two sentences.
この記事が気に入ったらサポートをしてみませんか?