Skip to content
Search

MachineLearning:Terms

lastmod 2024-11-06created 2015-12-2335 edits

A collection of terms related to machine learning.

The alphabetical list below was drafted from Artificial Intelligence (AI) & Machine Learning (ML) Dictionary - WikiDocs.

A

  • A2C: Advantage Actor-Critic
  • A3C(asynchronous advantage actor-critic)
  • AGI: Artificial General Intelligence (Korean: 인공일반지능)
  • AI Winter (Korean: AI 겨울)
  • AI alignment (Korean: AI 정렬)
  • AI safety (Korean: AI 안전)
  • AI: Artificial Intelligence (Korean: 인공지능)
  • ANN: artificial neural network (Korean: 인공신경망, 뉴럴네트워크)
  • ARIMA: AutoRegressive Integrated Moving Average (Korean: 자기회귀 통합 이동 평균)
  • ASR: automatic speech recognition (Korean: 자동음성인식)
  • AUC: area under the curve (Korean: 곡선 아래 면적)
  • Actor-Critic (Korean: 액터-크리틱, 행위자-비평자)
  • AdaBoost: Adaptive Boosting
  • AdaGrad
  • Agglomerative Clustering (Korean: 병합 군집 분석)
  • AlexNet (Korean: 알렉스넷)
  • Alpaca
  • AlphaGo (Korean: 알파고)
  • Autograd (Korean: 자동 미분)
  • Autoregressive Transformer (Korean: 자동회귀 트랜스포머, 자기회귀 트랜스포머)
  • accuracy (Korean: 정확도)
  • acoustic scene classification
  • action (Korean: 행동)
  • action-value function (Korean: 행동가치 함수)
  • activation function (Korean: 활성화함수)
  • agent (Korean: 에이전트)
  • anomaly detection (Korean: 이상 탐지)
  • association analysis (Korean: 연관성 분석)
  • attention (Korean: 어텐션, 주의, 주목)
  • audio generation (Korean: 오디오 생성)
  • autoencoder (Korean: 오토인코더, 자동부호기)
  • auxiliary classifier (Korean: 보조 분류기)

B

  • BART: Bidirectional and Auto-Regressive Transformers
  • BERT: Bidirectional Encoder Representations from Transformers (Korean: 버트)
  • BGD: Batch Gradient Descent (Korean: 배치 경사 하강법)
  • BLIP: Bootstrapping Language-Image Pre-training
  • BLSTM: Bidirectional Long Short-Term Memory
  • BM25
  • BPTT: Backpropagation Through Time
  • Bellman equation (Korean: 벨만 방정식)
  • Binary Cross Entropy (Korean: 이진 크로스 엔트로피)
  • Bootstrap Sampling (Korean: 부트스트랩 샘플링)
  • Borda count (Korean: 보르다 계산법)
  • Boston Housing Dataset (Korean: 보스턴 주택 가격 데이터세트)
  • backbone (Korean: 백본)
  • backpropagation (Korean: 역전파)
  • bagging (Korean: 배깅)
  • balanced MSE
  • baseline model (Korean: 베이스라인 모델, 기준선 모형)
  • batch normalization (Korean: 배치 정규화)
  • bias (Korean: 편향, 치우침, 바이어스)
  • biclustering, block clustering, co-clustering, two-mode clustering (Korean: 이중 클러스터링)
  • bilinear interpolation (Korean: 쌍선형 보간)
  • binary classification (Korean: 이진 분류)
  • bivariate normal distribute (Korean: 이변량정규분포)
  • boosting (Korean: 부스팅)

C

  • CART: Classification And Regression Tree (Korean: 분류 및 회귀 트리)
  • CBOW: Continuous Bag of Words
  • CGAN: conditional GAN (Korean: 조건부 GAN)
  • CLIP: Contrastive Language-Image Pre-training (Korean: 대조적 언어-이미지 사전 학습)
  • CNN: Convolutional Neural Network (Korean: 합성곱 신경망)
  • CRFs: conditional random fields (Korean: 조건부 랜덤 필드)
  • CTC: Connectionist Temporal Classification (Korean: 연결 시간 분류)
  • CVAE: Conditional Variational Autoencoder (Korean: 조건부 VAE)
  • Caffe (Korean: 카페)
  • CatBoost
  • ChatBIT
  • Confusion Matrix (Korean: 혼동 행렬)
  • Coreference Resolution (Korean: 상호참조해결, 공지시 해결)
  • Curse of Dimensionality (Korean: 차원의 저주)
  • casual inference (Korean: 인과 추론)
  • catastrophic forgetting (Korean: 파멸적 망각)
  • checkpoint (Korean: 체크포인트)
  • class (Korean: 클래스)
  • classification (Korean: 분류)
  • clustering (Korean: 클러스터링, 군집화)
  • collaborative filtering (Korean: 협업 필터링)
  • computer vision (Korean: 컴퓨터 비전, 컴퓨터비전, 컴퓨터 시각)
  • conformer (Korean: 컨포머)
  • context window (Korean: 컨텍스트 윈도우, 문맥 윈도우, 문맥 창, 문맥 윈도, 콘텍스트 윈도)
  • continual learning (Korean: 연속 학습)
  • contrastive loss (Korean: 대조 손실)
  • coordinate descent (Korean: 좌표 하강)
  • corpus, corpora (Korean: 말뭉치, 코퍼스)
  • cross-entropy(Korean: 교차 엔트로피, 크로스 엔트로피)
  • cross-validation (Korean: 교차 검증)

D

  • DBSCAN: Density-based spatial clustering of applications with noise
  • DCGAN: deep convolutional generative adversarial network
  • DDPG: Deep Deterministic Policy Gradient
  • DDQN: Double DQN
  • DL: deep learning (Korean: 딥러닝, 딥 러닝, 심층학습)
  • DNN-HMM
  • DNN: deep neural network (Korean: 심층 신경망)
  • DPO: Direct Preference Optimization (Korean: 직접 선호 최적화)
  • DQN: deep Q-Network
  • DRL: deep reinforcement learning (Korean: 심층강화학습)
  • Define-by-run
  • Diffusion Models
  • Discriminative AI (Korean: 판별형 AI, 판별 AI)
  • Distilbart
  • The democratization of ML(Korean: 머신러닝의 민주화)
  • data augmentation (Korean: 데이터 증강)
  • data leakage (Korean: 데이터 누수, 데이터 누설, 데이터 누출, 데이터 유출)
  • decision tree (Korean: 결정 트리, 의사결정나무)
  • decoder (Korean: 디코더)
  • deductive learning (Korean: 연역적 학습)
  • deep Q-learning (Korean: 심층 Q-러닝)
  • dense layer (Korean: 밀집 계층), fully-connected layer (Korean: 완전연결층, 전결합층)
  • digital image processing (Korean: 디지털 이미지 프로세싱)
  • discount factor (Korean: 할인율, 할인 계수)
  • downstream task (Korean: 다운스트림 작업)
  • dropout (Korean: 드롭아웃)
  • dynamic programming (Korean: 동적 계획법)

E

  • EVA-CLIP
  • Entropy (Korean: 엔트로피)
  • Explained Variance
  • Exploding Gradient (Korean: 그래디언트 폭발)
  • early stopping (Korean: 조기 중단)
  • embedding (Korean: 임베딩)
  • end-to-end learning (Korean: 엔드 투 엔드 학습, 종단 간 학습)
  • ensemble learning (Korean: 앙상블 학습)
  • environment (Korean: 환경)
  • episode (Korean: 에피소드)
  • epoch (Korean: 에포크, 에폭, 세, 세대)
  • explanatory variable (Korean: 설명 변수)
  • exploration-exploitation trade-off (Korean: 탐색-활용 트레이드오프)
  • ϵ-greedy (Korean: ϵ 탐욕)

F

  • FAISS
  • FCN: Fully Convolutional Networks
  • FNN: feed-forward neural networks (Korean: 피드포워드 신경망, 순방향 신경망, 앞먹임 신경망)
  • FP: feature permutation
  • FPN: Feature Pyramid Networks (Korean: 특징 피라미드망)
  • Fashion MNIST
  • Fast R-CNN
  • Faster R-CNN
  • feature (Korean: 특징, 피처)
  • feature engineering (Korean: 특징 공학, 특성 공학)
  • feature extraction (Korean: 특징 추출)
  • few-shot learning (Korean: 퓨샷 러닝)
  • fine-tuning (Korean: 파인 튜닝, 미세 조정)
  • forward propagation (Korean: 순전파)
  • fully-connected (Korean: 완전 연결, 전결합)

G

  • GAN: Generative Adversarial Networks (Korean: 생성적 적대 신경망)
  • GLUE: General Language Understanding Evaluation
  • GMM-HMM: Gaussian Mixture Model-Hidden Markov Model
  • GMM: Gaussian Mixture Model
  • GNN: Graph Neural Network (Korean: 그래프 신경망)
  • GPGPU: General-Purpose computing on Graphics Processing Units
  • GPT: Generative Pretrained Transformer
  • GPTs
  • GRU: Gated Recurrent Unit
  • Gemini (Korean: 제미나이, 제미니)
  • General-purpose AI (Korean: 범용 인공지능)
  • Glove
  • GoogleNet
  • Gradient Boosting (Korean: 그래디언트 부스팅)
  • generalization (Korean: 일반화)
  • generative model(Korean: 생성 모델)
  • gorilla problem (Korean: 고릴라 문제)
  • gradient (Korean: 경사, 그래디언트, 그레이디언트, 기울기, 구배, 물매)
  • gradient descent (Korean: 경사 하강법, 경사 감소법, 그래디언트 디센트)
  • grid search (Korean: 그리드 탐색, 그리드 서치, 격자 검색)
  • ground truth (Korean: 실측 자료)

H

  • HMM: hidden Markov model (Korean: 은닉 마르코프 모형)
  • Hamming score (Korean: 해밍 점수)
  • Heterogeneous data (Korean: 이질 데이터)
  • HiFi-GAN
  • Hugging Face (Korean: 허깅페이스)
  • hallucination (Korean: 환각)
  • hard negative
  • hidden layer (Korean: 은닉층, 은닉 계층, 히든 레이어, 숨은 레이어)
  • hierarchical clustering(Korean: 계층적 군집화, 계층적 클러스터링)
  • hyperbolic tangent, tanh (Korean: 하이퍼볼릭 탄젠트, 쌍곡 탄젠트)
  • hyperparameter(Korean: 하이퍼파라미터, 초매개변수, 하이퍼매개변수)
  • hyperplane (Korean: 초평면)

I

  • ICL: In-context learning
  • IG: Information Gain (Korean: 정보 이득)
  • IG: Integrated Gradients (Korean: 적분 경사, 통합 경사)
  • Image Captioning (Korean: 이미지 캡셔닝)
  • Instance Segmentation (Korean: 인스턴스 세그먼테이션)
  • Isolation Forest
  • image classcification (Korean: 이미지 분류)
  • image segmentation (Korean: 이미지 세그멘테이션, 이미지 세분화, 이미지 분할)
  • image-to-image translation
  • independent variable (Korean: 독립변수)
  • inductive learning (Korean: 귀납적 학습)
  • inference (Korean: 추론, 인퍼런스)
  • interpolation (Korean: 보간)

J

  • job (Korean: 작업, 잡)

K

  • KLD: Kullback–Leibler divergence (Korean: 쿨백-라이블러 발산)
  • Keras (Korean: 케라스)
  • k-NN: k-nearest neighbors (Korean: k 최근접 이웃)
  • k-fold cross-validation (Korean: k-겹 교차 검증)
  • k-means clustering (Korean: k 평균 군집법)
  • k-fold average ensemble (Korean: k-폴드 평균 앙상블)
  • knowledge base (Korean: 지식 베이스)
  • koNLPy

L

  • L2 normalization (Korean: L2 정규화, L2 노멀라이제이션)
  • L2 regularization (Korean: L2 규제, L2 정규화)
  • LAS: Listen, Attend and Spell
  • LDA: Latent Dirichlet Allocation (Korean: 잠재 디리클레 할당)
  • LDA: Linear Discriminant Analysis (Korean: 선형 판별 분석법)
  • LIMA: Less Is More for Alignment
  • LLM: Large Language Model (Korean: 대형 언어 모델, 대규모 언어 모델, 거대언어모델)
  • LLaMA
  • LMM: Large Medical Model (Korean: 대규모 의료 모델)
  • LMM: Linear Mixed Model (Korean: 선형 혼합 모델)
  • LMMs: Large Multimodal Models (Korean: 대형 멀티모달 모델)
  • LSTM: long short-term memory (Korean: 장단기 메모리)
  • LangChain
  • Latent Variable Models (Korean: 잠재 변수 모델)
  • Leaky ReLU (Korean: 리키 렐루, 누출 ReLU, 누설 ReLU)
  • LightGBM
  • Llama 2
  • Llama 3 (Korean: 라마 3)
  • LoRA: Low-Rank Adaptation
  • Low Rank Approximation (Korean: 저계수 근사)
  • language model (Korean: 언어 모델)
  • learning (Korean: 학습)
  • learning rate (Korean: 학습률)
  • linear regression (Korean: 선형 회귀)
  • local minima (Korean: 극소점)
  • logistic (Korean: 로지스틱) function
  • logistic regression (Korean: 로지스틱 회귀)
  • logit (Korean: 로짓, 로지트) function
  • loss function (Korean: 손실 함수), cost function (Korean: 비용 함수)

M

  • MAB: Multi-Armed Bandits (Korean: 여러 대의 슬롯머신, lit. "multiple slot machines")
  • MAE: Mean Absolute Error
  • MCTS: Monte Carlo Tree Search (Korean: 몬테 카를로 트리 탐색)
  • MDP: Markov decision process (Korean: 마르코프 결정 과정)
  • ML: machine learning (Korean: 머신러닝, 기계 학습, 기계학습)
  • MLE: maximum likelihood estimation (Korean: 최대우도법, 최우도법, 최대우도추정)
  • MLM: Masked Language Model
  • MLP: multilayer perceptron (Korean: 다층 퍼셉트론)
  • MMLU: Massive Multitask Language Understanding
  • MNIST
  • MSE: Mean Squared Error (Korean: 평균 제곱 오차)
  • MT: Machine Translation (Korean: 기계 번역)
  • Many-Shot In-Context Learning
  • Markov process (Korean: 마르코프 과정)
  • Markov property (Korean: 마르코프 성질)
  • Markov reward process (Korean: 마르코프 보상 과정)
  • Mask R-CNN
  • Monte Carlo Methods (Korean: 몬테카를로 방법)
  • Multiclass Classification (Korean: 다중 클래스 분류)
  • Multicollinearity (Korean: 다중공선성)
  • Multimodal AI (Korean: 멀티모달 인공지능)
  • Multiple Linear Regression (Korean: 다중 선형 회귀)
  • maze (Korean: 미로)
  • meta-learning (Korean: 메타 학습)
  • mini-batch gradient descent (Korean: 미니 배치 경사 하강법)
  • model (Korean: 모델)
  • model-based learning (Korean: 모델 기반 학습)
  • model-free learning

N

  • NAG: Nesterov Accelerated Gradient
  • NER: Named Entity Recognition (Korean: 개체명 인식)
  • NLG: Natural Language Generation (Korean: 자연어 생성)
  • NLP: Natural Language Processing (Korean: 자연어 처리)
  • NLTK: Natural Language Toolkit
  • NLU: Natural Language Understanding (Korean: 자연어 이해)
  • NMF: Non-negative Matrix Factorization (Korean: 비음수 행렬 분해)
  • NNLM: Neural Network Language Model
  • Naive Bayesian Model (Korean: 단순 베이즈 모델)
  • Naver sentiment movie corpus
  • Non-parametric(Korean: 비모수)
  • Numpy (Korean: 넘파이)
  • n-gram
  • nDCG: normalized Discounted Cumulative Gain
  • neuron (Korean: 뉴런)
  • nonlinear regression (Korean: 비선형 회귀)
  • notMNIST
  • New Page

O

  • ONNX: Open Neural Network Exchange
  • On-Device Machine Learning (Korean: 온디바이스 머신러닝)
  • One-Hot encoding (Korean: 원핫 인코딩)
  • Optimization Algorithm (Korean: 최적화 알고리즘)
  • Optuna
  • object detection (Korean: 객체 탐지, 객체 검출)
  • object recognition (Korean: 객체 인식)
  • odds (Korean: 오즈, 승산비, 비율비)
  • optical character recognition (Korean: 광학 문자 인식)
  • optimization problem (Korean: 최적화 문제)
  • optimizer function (Korean: 최적화 함수)
  • overtraining, overfitting (Korean: 과적합, 과대적합, 오버피팅)

P

  • PAL: Program-aided Language Models (Korean: 프로그램 보조 언어 모델)
  • PCA: principal component analysis (Korean: 주성분분석)
  • PEFT: Parameter-Efficient Fine-Tuning
  • PPO: Proximal Policy Optimization
  • PReLU: Parametric ReLU (Korean: 매개변수적 ReLU)
  • PaLM: Pathways Language Model
  • Pareto optimization (Korean: 파레토 최적화)
  • Point-E
  • Pong (Korean: 퐁)
  • Py-hanspell (Python Korean spell-checking library. Uses the Naver spell checker)
  • PyKoSpacing (Automatic Korean word spacing with Python)
  • PyTorch (Korean: 파이토치)
  • parameter (Korean: 매개변수, 파라미터)
  • parameter (Korean: 모수)
  • pattern recognition (Korean: 패턴 인식)
  • perceptron (Korean: 퍼셉트론)
  • perplexity
  • policy (Korean: 정책)
  • policy gradient (Korean: 정책 경사, 정책 기울기)
  • polynomial regression (Korean: 다항 회귀)
  • pooling (Korean: 풀링)
  • pre-training (Korean: 사전 학습)
  • precision (Korean: 정밀도)
  • prompt engineering (Korean: 프롬프트 엔지니어링)

Q

  • Q value (Korean: Q값)
  • Q-Learning (Korean: Q러닝)
  • QA: question answering (Korean: 질의응답)
  • QLoRA: Quantized Low-Rank Adapter

R

  • R-CNN
  • RANSAC: Random Sample Consensus (Korean: 랜덤 샘플 합의)
  • RBM: restricted Boltzmann machine (Korean: 제약 볼츠만 머신)
  • REDNet
  • REINFORCE
  • RL: reinforcement learning (Korean: 강화학습)
  • RLHF: Reinforcement Learning from Human Feedback
  • RMSProp
  • RNN: Recurrent Neural Network (Korean: 순환 신경망)
  • ROC curve (Korean: ROC 곡선)
  • ROI Pooling
  • ROUGE: Recall-Oriented Understudy for Gisting Evaluation
  • RSS: Residual sum of square (Korean: 잔차제곱합)
  • ReAct (Reason + Act)
  • ReLU: rectified linear unit (Korean: 렐루, 정류 선형 유닛, 정류 선형 단위)
  • Regularized Linear Regression (Korean: 규제 선형 회귀)
  • ResNet: Residual Network (Korean: 레즈넷)
  • Rule-based system (Korean: 규칙 기반 시스템)
  • random forest (Korean: 랜덤 포레스트)
  • recommender systems (Korean: 추천 시스템)
  • regression (Korean: 회귀)
  • regression line (Korean: 회귀직선)
  • regularization (Korean: 정규화, 정칙화, 규제화, 규칙화)
  • representation learning (Korean: 표현 학습, 표상 학습)
  • residual (Korean: 잔차)
  • residual connection (Korean: 잔차 연결, 리지듀얼 커넥션)
  • response variable (Korean: 반응 변수)
  • reward (Korean: 보상)
  • rollout (Korean: 롤아웃)

S

  • SGD: stochastic gradient descent (Korean: 확률적 경사 하강법)
  • SHAP
  • SLAM: Simultaneous Localization and Mapping
  • SMOTE: Synthetic Minority Over-sampling Technique
  • SOTA: State of the Art
  • SSD: Single Shot MultiBox Detector
  • SVD: Singular Value Decomposition (Korean: 특잇값 분해)
  • SVM: support vector machine (Korean: 서포트 벡터 머신)
  • Skip Connection (Korean: 스킵 연결)
  • Sora (Korean: 소라)
  • Stable Diffusion
  • Superintelligence (Korean: 초지능)
  • Swish
  • saddle point (Korean: 안장점)
  • safetensors
  • saturate (Korean: 포화)
  • segmentation (Korean: 세그먼테이션, 세그멘테이션, 세분화, 분할)
  • self-supervised learning (Korean: 자기 지도 학습)
  • semantic segmentation (Korean: 의미론적 세그멘테이션)
  • semi-supervised learning(Korean: 준지도 학습, 지도/자율 학습)
  • sentiment analysis (Korean: 감성 분석)
  • seq2seq: Sequence to Sequence (Korean: 시퀀스 투 시퀀스)
  • sequence data (Korean: 시퀀스 데이터)
  • sigmoid (Korean: 시그모이드)
  • simulated annealing (Korean: 시뮬레이티드 어닐링)
  • skip connection (Korean: 잔차 연결, 건너뛴 연결)
  • softmax (Korean: 소프트맥스)
  • soynlp
  • sparse matrix (Korean: 희소 행렬)
  • sparsity problem (Korean: 희소성 문제)
  • speech recognition (Korean: 음성 인식)
  • stacking (Korean: 스태킹)
  • standard deviation (Korean: 표준 편차)
  • state (Korean: 상태)
  • state transition (Korean: 상태 전이)
  • stemming (Korean: 어간 추출)
  • snowball stemming
  • super resolution (Korean: 초해상도)
  • supervised learning (Korean: 지도학습)

T

  • TCN: Temporal Convolutional Network
  • TDM: Term Document Matrix
  • TDM: temporal difference model (Korean: 시간차 모델)
  • TF-IDF
  • Tavily (Korean: 타빌리)
  • Taxonomy
  • Tensor (Korean: 텐서)
  • TensorFlow Lite (Korean: 텐서플로 라이트)
  • TensorFlow.js
  • task (Korean: 과업, 작업, 과제, 태스크)
  • text classification (Korean: 텍스트 분류)
  • text mining (Korean: 텍스트 마이닝)
  • text similarity (Korean: 텍스트 유사도)
  • time series anaylisis (Korean: 시계열 분석)
  • time series classification
  • time series forecasting (Korean: 시계열 예측)
  • token (Korean: 토큰)
  • training (Korean: 훈련, 트레이닝, 학습)
  • training dataset (Korean: 훈련 데이터셋)
  • trajectory (Korean: 궤적)
  • transfer learning (Korean: 전이 학습)
  • transformer (Korean: 트랜스포머)
  • transition
  • triplet loss (Korean: 삼중항 손실)
  • triplet mining

U

  • U-Net
  • UCN: universal correspondence network
  • unsupervised learning (Korean: 비지도 학습)

V

  • VAE: Variational Autoencoder (Korean: 변분 오토인코더)
  • VITS
  • VQA: Visual Question Answering
  • VSR: Visual Speech Recognition
  • Vanishing Gradient (Korean: 그래디언트 소실, 경사 소실, 기울기 소실)
  • ViLT: Vision-and-Language Transformer
  • ViT: Vision Transformer (Korean: 비전 트랜스포머)
  • Vicuna
  • validation data (Korean: 검증 데이터, 유효성 데이터)
  • value (Korean: 가치)
  • value function (Korean: 가치 함수)
  • voting (Korean: 투표, 보팅)

W

  • WRMSE: Weighted Root Mean Square Error
  • weight (Korean: 가중치)
  • weight decay
  • word embedding (Korean: 단어 임베딩)
  • word patch alignment
  • word2vec

X

  • XAI: eXplainable AI (Korean: 설명 가능 인공지능)
  • XGBoost(eXtra Gradient Boost)

Y

  • YOLO: You Only Look Once

Z

  • ZSL: Zero-Shot Learning (Korean: 제로샷 학습)

Batch Learning

In batch learning, the system cannot learn incrementally. It must be trained using all the available data. This takes a lot of time and resources, so it is usually done offline.

Batch Size

The size of the data that goes into a single computation. See the page for details.

Epoch (Korean: 에포ㅋ; era)

The number of times every piece of data in the training dataset has passed through the model once; that is, the number of times the entire training dataset is trained on. See the page for details.

Bias

Determining the operating point of a voltage or current in advance. In short, a constant.

Curse of dimensionality

Refers to various problematic phenomena that arise when organizing data in high-dimensional spaces.

Drop-Out and Drop-Connect

A very important method that is highly effective against overfitting and the local minima problem. Dropout randomly excludes neurons in each layer of a deep learning neural network from computation with a set probability. Drop-connect, unlike excluding the neurons themselves, probabilistically excludes the connections between neurons from computation. Though simple, these methods have proven very effective and are widely used.

Feature Extraction

Ground truth

A term for the actual true value that serves as the reference when evaluating how accurate a (computed, estimated, predicted) value is. See the page for details.

Hyperparameter

Hypothesis

The relationship (function) between the predicted data and the result values is called the Hypothesis.

Linear Fitting

Local minima or Local minimum problem

In a neural network, training is done by adjusting each of the weight values corresponding to a huge number of parameters, which controls the behavior of the whole system. Because so many parameters are being moved, the parameter adjustment meant to find the optimum where the energy is minimal may undesirably train only toward a local minimum, and training may fail to reach the parameters that give the global minimum. This phenomenon is highly sensitive to the structure of the parameters' initial values.

Loss function

Quantification of the predicted scores against the ground truth labels.

A function that computes how far the output predicted by the neural network's computation differs from the output that is the correct answer. There are various loss functions; commonly used ones are RMSE (Root Mean Squared Error), which computes every error of each output neuron to express the degree of error, and the improved Cross-Entorpy error calculation.

In a loss function, optimization is the process of finding the parameter W that minimizes the loss function.

Mean Average Precision (mAP)

Mean average precision.

See Precision and recall for details.

Precision

Precision is the proportion of items judged positive that are actually true. It is defined as 'true positives / (true positives + false positives)'.

See Precision and recall for details.

Recall

Recall is the proportion of items whose actual value is true that are judged positive. It is defined as 'true positives / (true positives + false negatives)'.

See Precision and recall for details.

Nonlinear Transformation, or Activation

Optimization

It's just optimization, duh.

Over Fitting

Prepared data may contain noise, real-world data has some errors, and training data is often less than ideal. In a neural network with too many neuron nodes or more degrees of freedom than necessary, the network can end up fitting the given training data too precisely. In such cases, even if the training data is learned well, feeding in test data or real data can produce undesired results. This is called overfitting. Overfitting used to be a chronic problem for neural networks, but deep learning has since developed various ways to avoid it, which became a major reason deep learning was so widely welcomed.

Data preprocessing (Pre-Training)

Score function

A function that maps raw data to class scores (or labels).

Simulated annealing

A kind of method for escaping the problem of getting stuck in a local minimum.

Maximum a posteriori estimation (MAP)

Maximum a posteriori (MAP) refers to the mode of the posterior probability in Bayesian statistics.

Maximum likelihood estimation (MLE)

The maximum likelihood method, or maximum likelihood estimation, is a method of finding the parameters of a random variable based on values sampled from that random variable.

Occlusion

A region where an object is partially hidden from view by another object.

Online Learning

In online learning, the system is trained by feeding it data sequentially, either one at a time or in small groups called mini-batches. Each training step is fast and cheap, so the system can learn immediately as data arrives. Online learning suits systems that receive data continuously and must adapt to rapid changes on their own. It is also suitable when computing resources are limited. Data that has already been trained on is no longer needed and can be discarded.

One important parameter of an online learning system is how quickly it should adapt to changing data. This is called the learning rate. With a high learning rate, the system adapts to data quickly but also quickly forgets old data. Conversely, with a low learning rate, the system has more inertia and learns more slowly. However, it becomes less sensitive to noise in new data or to unrepresentative data points.

Inferencing

Inference is the act or process of deriving logical conclusions from what is already known. Here, it means an AI program using the model produced after training to find the trained objects in an image.

Iteration

  • Deep Learning 6.2. Optimization (3) - Training units (Epoch, Batch size, Iteration)
  • In the Naver English dictionary, Iteration means "repetition (of a calculation or computer processing procedure)".
  • The number of batches needed to train the model on the entire data once.
  • In other words, the number of parameter updates needed to complete 1 epoch.
  • Since parameters are updated once per batch, Iteration is "number of parameter updates = number of batches".

Metadata

Data created for some purpose. In this page, it refers to data created for the purpose of training AI.

Predict vs Inference

Inference
Interested in how variables affect the outcome
Variables are added or removed for the model's explanatory power
Given a set of data you want to infer how the output is generated as a function of the data.
Predict
Interested in how good the resulting values are
The model can be changed to improve the accuracy of the results
Given a new measurement, you want to use an existing data set to build a model that reliably chooses the correct identifier from a set of outcomes.

See also

Favorite site

References

  1. Machine_Learning_Glossary_-_Google_Developers.pdf ↩

  2. CodeOnWeb_-_ml-glossary.pdf ↩

author Zer0 sitemap