1, AI
Artificial intelligent method on the machine: machine learning, computer vision, planning and decision-making, natural language processing, cognitive reasoning, efficient search
2, the three schools
Symbolism
Connectionism: CNN
Behaviorism
3, two routes
Structure imitation
Mimics
4, AI chip
The human brain structure
Brain Function
5, relations
Artificial Intelligence> Machine Learning> MLP> deep learning
6, the type of learning
Rule learning
Sample learning
Self-learning
7, sample learning
Classification: linear, nonlinear
Regression: curve fitting sample
Airspace extension
Extending the time domain
8, sample learning problems
Data Quality: sparsity, small sample size, prejudice, imbalance
Training Efficiency: Convergence, sample selection, parallelization
Scale model: model to quantify, model cutting, model migration
9, rule learning
From learning rule to follow the rules
--- expert system knowledge engineer
Conditional language, logical language
Rules table == "data mining
Deterministic logic = "fuzzy logic
10, rule learning problem
Sensitivity: robustness, repeatability, reproducibility
Simulation high cost: delay reward, simulation error, low efficiency of learning
Behavioral biases: Awards prejudice, local extreme fraud conspiracy
11, independent study AutoML
12, process: raw data == "data processing ==" features works == "model design ==" Model Validation
13, independent learning: sparse network, multitasking support, technical multiplexing, dynamic building
------ target equalization. . . Catastrophic forgotten
14, the intelligent system: self-organization, self-learning, adaptive
15, the depth of study: tensors, tensor operations, calculation map, automatic differentiation tool calculates extended package
16、PaddlePaddle
PaddleDection:Mask R-CNN 、YOLO V3、FPN、Casade RCNN、TridentNet
PaddleSeg: V-Net, Deep LabV3, ICNet (real-time semantic segmentation)
Mobile end: MobileNet
Server: XCeption
17. Process: needs analysis, technology selection, model training, model tuning, hardware deployment, actual production
18, PaddleSeg: lane dividing line, land division
19、PaddleVideo
End-to-End
Two-State
Video Targeting: BMN, BSN
Visual inspection: appreciated, edit, create,
20、PaddleNLP
WordEmbedding
ELMO
GPT
BERT: Basic Semantic Modeling Language Unit
ERNIE: enhance knowledge-based semantic modeling
Text similarity, matching questions and answers, sentiment analysis, natural language inference
Morphology, structure, semantics
Knowledge forecast, ordering the sentence, the sentence prediction logic (since it is, if it is, and, in spite of it)
Search CTR forecast, intelligent searching questions and answers
ERNIE development
Tasks: generation tasks, text, match
Industry: legal, medical, financial
21, federal study
Federal honeycomb learning platform
Google Input GBoard
TensorFlow Federated Pysyft
Federal learn, secure multi-party computation, the difference privacy
FATE
Search advertising --- logistic regression
Homomorphic encryption: homomorphic addition, multiplication homomorphism
France Snips
OWKIN
Risk analysis, marketing analysis
Life insurance recommendation system, customer loans overdue rate prediction system
XGBoost
LR