Please enable JavaScript.

Coggle requires JavaScript to display documents.

Machine Learning (tensorflow (data type (categorical (text), numerical,…

- - - - learning
        
        bias
        
        weights
        
        loss
        
        squared loss
        
        Mean Square Error
        
        \[ MSE = \frac{1}{N} \sum_{(x,y) \in D} (y-prediction(x))^{2}\]
        
        reducing
        
        iterative
        
        Compute parameter updates
        
        gradient descent
        
        8 more items...
        
        loss function
        
        hyperparameter
        
        log loss
        
        \[Log Loss = \sum_{(x,y) \in D}-ylog(y') - (1-y)log(1-y')\]
        
        cross entropy
        
        categorical
        
        binary
      - empirical risk minimization
    - - continuous
      - linear
      - goldilocks principle
      - logistic
        
        generate
        
        probability
        
        value to
        
        binary category
        
        classification threshold
        
        2 more items...
    - - discrete
      - evaluating
        
        accuracy
        
        class-imbalanced data set
        
        precision
        
        \[Precision = \frac{TP}{TP+FP}\]
        
        recall
        
        \[Recall = \frac{TP}{TP+FN}\]
      - receiver operating characteristic curve
        
        ROC curve
        
        plot two parameters
        
        TPR
        
        FPR
        
        \[FPR = \frac{FP}{FP+TN}\]
        
        Area Under the ROC Curve
        
        AUC
        
        desirable reason
        
        classification-threshold-invariant
        
        scale-invariant
      - prediction bias
        
        to reduce
        
        calibration layer
        
        root cause
        
        Incomplete feature set
        
        Noisy data set
        
        Buggy pipeline
        
        Biased training sample
        
        Overly strong regularization
- - - - level
        
        reusable libraries
        
        common model components
        
        tf.layers, tf.losses, tf.metrics
        
        high
        
        tf.estimators
        
        scikit-learn
        
        Python
        
        NumPy
        
        graph
        
        2 more items...
        
        sklearn
        
        linearRegressor
        
        lower
        
        TensorFlow
  - - - data structure
        
        DataFrame
        
        imagine
        
        relational data table
        
        Series
        
        column
      - column-oriented
      - data analysis API
  - - - text
- - - - divide two subsets
        
        training set
        
        test set
        
        two conditions
        
        large enough
        
        the part for the whole
      - three subsets
        
        add
        
        validation set
- - - - vocabulary
        
        vector
        
        length
        
        multi-hot encoding
        
        one-hot encoding
        
        sparse representation
      - out-of-vocabulary bucket
  - - - standard range
    - - log
      - clipping
    - - =bucketing
- - - - activation functions
        
        sigmoid function
        
        \[z\]
        
        log-odds
        
        \[y = \frac{1}{1+e^{-z}}\]
        
        ReLU
        
        \[F(x) = max(0,x)\]
        
        softplus
        
        \[y=log(1+e^x)\]
        
        hyperbolic tangent
        
        \[y = \tan h (x)\]
      - hidden layers
      - training
        
        algorithm
        
        backpropagation
        
        failure cases
        
        vanishing gradients
        
        exploding gradients
        
        dead ReLU Unit
        
        dropout
        
        partial derivative
      - multi-class
        
        one vs. all
        
        softmax
        
        full softmax
        
        candidate sampling
        
        assume
        
        exactly one class
      - convolutional
        
        image recognition
        
        ILSVRC
        
        ImageNet
        
        VGG
        
        GoogLeNet
        
        ResNet
        
        output
        
        image feature
        
        down sampling
        
        pooling
        
        max pooling
        
        two-dimensional
        
        flatten
        
        one-dimensional
        
        segmentation
        
        pixel
        
        FCN
        
        semantic image segmentation
        
        instance segmentation
        
        stride
        
        kernel
        
        padding
      - Recurrent
        
        Long Short Term Memory
      - hopfield network
        
        cross talk
        
        boltzmann machine
        
        restricted boltzmann machine
        
        deep belief network
- - - - sum of the squares of all the feature weights
        
        \[ \left.\right.||w||_2^2 \]
- - - - standard dimensionality reduction techniques
        
        PCA
      - Word2vec
        
        google
        
        distributional hypothesis
- - - - output of results
      - high level neural networks API
        
        keras
        
        keras.util
        
        np_utils.categorical()
        
        Conv2D
        
        MaxPooling2D
  - - - artificial neural network
        
        artificial intelligence
      - multilayer perceptron
        
        neural nets
  - - - R-CNN
        
        Faster R-CNN
        
        selective search
        
        region proposal
        
        Region with CNN features
    - - speed
        
        FPS
        
        frame per second
      - bounding box
        
        IoU
        
        Intersection Over Union
        
        \[IoU = \frac{Area of Overlap}{Area of Union}\]
      - class
        
        mAP
        
        mean average precision
- - - - AWS Deep Learning AMI
        
        Packer
        
        Ansible
      - Jupyter Notebook
  - - - Nvidia-Docker
- - - - Amazon S3
  - - - Amazon Elastic Container Registry
- - - - inference code
- - - - Train a model
        
        train model
        
        evaluate model
        
        Deploy the model
        
        Deploy to production
        
        monitor, collect data, evaluate
      - fetch
        
        clean
        
        prepare
- - - - Training parameter
        
        Training
- - - - general categoires
      - 고객사별 모델
        
        example
        
        jeans
        
        Clothing > Pants
        
        Fashion > Men > Jeans
- - - - Inception v3
        
        retrain.py
- - - - gensim
- - - - Gaussian Process
    - - Expected Improvement
- - - - WSABIE
        
        weight approximate-rank pairwise loss
        
        WARP loss
      - metric learning
        
        fail label co-occurrences
      - skip-image
        
        skip-gram model
        
        CNN-RNN framework
      - Fast0Tag
        
        zero shot scenario
      - RIA
        
        combine CNN, RNN
        
        predict the number of object
        
        fail to account non-object descriptive tags
      - DeViSE
      - S-CNN-RNN
        
        semantic regularization
        
        differ to recall more of the complex, non-visual tags
    - - anchor, positive, negative
  - - - annotation
        
        label
        
        caption
    - - must be inferred
        
        weather
        
        location
      - not clearly present
        
        depicted event
        
        football game, baseball game, concert, vacation
      - a life of their own
        
        tbt
        
        Throwback Thursday
- - - - 2 important drawbacks
        
        propose
        
        log-sum-exp pairwise
        
        LSEP loss
        
        theoretical analyses
        
        Bayes consistency
        
        risk minimization
        
        per-class threshold estimation
        
        hinge loss
        
        difficult to optimize
        
        non-smooth
        
        ranking objective
        
        not fully optimize
        
        label decision part
        
        top-k and thresholding
        
        ignore image content
    - - multiple binary label