Skip to content

Latest commit

 

History

History
328 lines (307 loc) · 15.4 KB

basics.md

File metadata and controls

328 lines (307 loc) · 15.4 KB
Raw

Basics of Machine Learning

Types of Problems:

  • Classification
  • Regression
  • Clustering
  • Rule Extraction

Supervised Learning

  • Labels associated with the training data are used to correct the algorithm.

Unsupervised Learning

  • The model has to be set up right to learn the structure in the data.

Representation Learning Algorithms

  • Feature learning. Algorithm identifies important features on its own.

Deep Learning

  • Algorithms that learn what features matter.
  • Neural Networks
    • Most common class of deep learning algorithms.
    • Used to build representation learning systems.
    • Composed of neurons (binary classifiers)
    • Wide
      • Better for memorization
    • Deep
      • Better for generalization

Neurons

  • Apply 2 functions on inputs.
  • Best values of W and b found by using cost function, optimizer, and training data.
  • Back Propagation

Linear (affine) transformation

  • Like linear regression.
  • X1 * W1 + b
    • W = Weights
      • Shape of W
        • First dimension is equal to number of dimensions of feature vector.
        • Second dimension is equal to the number of params required to be tuned. (same goes for b)
    • B = Bias
      • Determined during training process.

Activation Function

  • Helps to model non linear functions. (Logistic regression)
  • Introduces non-linearity into the network.
  • Ex.
    • ReLu (Rectified Linear Unit)
      • Max(Wx + b, 0)
    • SoftMax
      • Multi-class logistic regression
      • Layer right before output.
        • Must have same number of nodes as the output layer.
      • Use candidate sampling for large # of output classes.
    • Sigmoid
      • Binary classification logistic regression

Failure Cases for Backpropagation

Vanishing Gradients

  • Gradients for lower layers (closer to input) can become very small.
  • Leads to very slow training, if at all.
  • ReLu as an activation function can help prevent vanishing gradients.

Exploding Gradients

  • Weights in a network are very large, the gradients for the lower layers involve products of many large terms.
  • Gradients get too large to converge.
  • Batch normalization and/or lowering learning rate can prevent this.

Dead ReLU Units

  • Weighted sum of a ReLU falls below 0, ReLU can get stuck.
  • Outputs 0 activation, contributing nothing to network’s output, and gradients can no longer flow through it during backpropogation.
  • Lowering learning rate can help keep ReLU from dying.

Modeling Linear Regression

  • 1 neuron with just an affine transformation.
  • Y = Ax + b
  • Minimize Least Square Error

Optimizers for Best Fit

  • Method of Moments
  • Method of Least Squares
  • Maximum likelihood Estimator

Reducing Loss

  • Convergence: When loss stops changing or at least changes extremely slowly.
  • Gradient is a vector.
  • Learning rate is a scalar.
  • Gradient is multiplied by the learning rate.

Hyperparameters

  • Configuration settings used to tune hot the model is trained.
  • Steps
    • Total number of training iterations. One step calculates the loss from one batch and uses that value to modify the model’s weights once.
  • Batch Size
    • Number of examples (chosen at random) for a single step.
    • Total # of trained examples = Batch Size * Steps
  • Learning rate

Stochastic Gradient Descent

  • Random samples from data set to estimate.
  • Uses a batch size of 1 per iteration.
  • Works (given enough iterations), but noisy

Mini-Batch Stochastic Gradient Descent

  • Compromise between full batch and SGD
  • Typically between 10 – 10K examples chosen at random.
  • Reduce noise, but still more efficient than full-batch.

Periods

  • the # of training examples in each period = batch size * steps / period
  • Controls granularity of reporting.
    • If periods = 7 and steps = 70, the loss value will be output every 10 steps.
  • Modifying period value does not alter what model learns.

Generalization

  • The less complex an ML model, the more likely that a good empirical result is not just due to the peculiarities of the sample.
  • Overfitting occurs when a model tries to fit the training data so closely that it does not generalize well to new data.
  • Identify Overfitting
    • Loss for the validation set is significantly higher than for the training set. (look at loss curve (loss/iterations))
    • Validation loss eventually increases with iterations.
  • If the key assumptions of supervised ML are not met, then we lose important theoretical guarantees on our ability to predict new data.
  • 3 Basic Assumptions
    • We draw examples independently and identically at random from the distribution. I.e. examples don’t influence each other.
    • The distribution is stationary; that is it does not change within the data set.
    • We draw examples from partitions from the same distribution.

Training, Validation, and Test Sets

  • Training set – a subset to train a model.
  • Test set – a subset to test the trained model.
    • Must be large enough to yield statistically meaningful results.
    • Is representative of the data set as a whole. i.e. don’t pick a test set with different characteristics than the training set.
  • Doing many rounds of just using a training and test set might cause implicit fitting to the peculiarities of the specific test set.
    • Use a validation set too!
    • Flow
      • Train model
      • Use model on validation set
      • Update hyperparams
      • Repeat
      • Finally test on test set

Representation

  • Process of mapping data to useful features.
  • Discrete feature
    • A feature with a finite set of possible values.
    • Categorical feature are an example

One-Hot Encoding

  • A sparse vector in which:
    • One element is set to 1
    • All other elements are set to 0
  • Commonly used to represent strings or identifiers that have a finite set of possible values.

Feature Engineering

  • Process of determining which features might be useful in training a model, and then converting raw data from log files and other sources into said features.
  • Sometimes called feature extraction.

Qualities of Good Features

  • Avoid rarely used discrete feature values.
    • Should appear more than 5 or so times in a data set.
    • Having many examples with the same discrete value gives the model a chance to see the feature in different settings, and in turn, determine when it’s a good predictor for the label.
  • Prefer clear and obvious meanings
    • Ex. house_age_years vs. house_age
    • Some cases, noisy data causes unclear values, such as data coming from sources that didn’t check for appropriate values.
      • Ex. user_age_years: 277
  • Don’t mix “magical” values with actual data
    • Ex. quality_rating between 0 and 1.
      • If no value, it is set to -1
      • Create a Boolean feature to indicate if quality rating was defined.
    • Replace “magical” values as follows
      • For a variable that take a finite set of values (discrete variables), add a new value to the set and use it to signify that feature value is missing.
      • For continuous variables, ensure missing values do not affect the model by using the mean value of the feature’s data.
  • Account for upstream instability
    • Definition of a feature shouldn’t change over time.

Cleaning Data

  • Scaling feature vectors
    • Converting floating point feature values from their natural range (100 to 900) to a standard range (0 to 1 or -1 to 1)
    • Scaling ~= Normalization
    • If only 1 feature, little to no practical benefit.
    • Multiple features, great benefits
      • Helps gradient descent convere more quickly
      • Helps avoid NaN traps
        • One number in the model becomes a NaN (value exceeds floating point precision limit during training) and due to math operations, every other number in the model also eventually becomes NaN.
      • Helps the model learn appropriate weights for each feature. Without scaling, the model pays too much attention to features having a wider range.
  • Handling extreme outliers
    • Log scaling
      • Still leaves a tail on distribution
    • Cap or Clipping
      • Reduce feature values that are greater than a set maximum value down to that maximum value.
      • Also, increasing feature values that are less than a specific minimum value up to that minimum value.
    • Binning (Bucketing)
      • Converting a (usually continuous) feature into multiple binary features called buckets or bins, typically based on a value range.
  • Scrubbing
    • Data can be unreliable due to:
      • Omitted values
      • Duplicate examples
      • Bad labels
      • Bad feature values
    • “Fix” by removing them from data set.
    • Omitted and duplicate easy to detect.
    • Detecting bad data in aggregate by using Histograms
    • Stats can also help identifying bad data:
      • Max and Min
      • Mean and Median
      • Standard Deviation
  • Follow These Rules:
    • Keep in mind what your data should look like
    • Verify that the data meets these expectations
      • Or that you can explain why it doesn’t
    • Double check that the training data agrees with other soruces
      • i.e. dashboards

Feature Crosses

  • A synthetic feature formed by crossing (Cartesian product) individual binary features obtained from categorical data or from continuous features via bucketing.
  • Helps represent nonlinear relationships.
  • Encoding Nonlinearity
  • Crossing One-Hot Vectors

Regularization

  • Minimize loss + complexity
    • Structural Risk Minimization
    • Penalizes complexity to prevent overfitting
  • 2 Common Ways to Think About Model Complexity
    • As a function of the weights of all the features in the model
      • L2 Regularization
      • A feature weight with a high absolute value is more complex than one with a low absolute value.
      • L2 = w1^2 + w2^2 + … + wn^2
      • Consequences of L2 Regularization
        • Encourages weight values toward 0 (but not exactly 0)
        • Encourages the mean of the weights toward 0, with a normal (bell shaped or Gaussian) distribution.
    • As a function of the total number of features with nonzero weights
  • Most developer tune the overall impact of the regularization term by multiplying it by a scalar known as lambda (regularization rate)
    • Minimize(loss function + lambda(regularization function))
    • When choosing a lambda value, the goal is to strike the right balance between simplicity and training-data fit
      • Lambda too high
        • Model will be simple but run the risk of underfitting data.
      • Lambda too low
        • Model will be more complex and run the risk of overfitting data.

Early Stopping

  • Ending training before the model reaches convergence (training loss finishes decreasing).
  • End model training when loss on a validation dataset starts to increase, that is, when generalization performance worsens.

Sparsity

  • Sparse vectors often contain many dimensions.
  • Creating a feature cross results in even more dimensions.
  • High Dimensionality -> Large Model Size -> Large RAM reqs
  • L1 Regularization
    • Penalizes absolute value of weights. (|weight|)
    • Derivative of L1 is a constant, k. (2 * weight for L2)
      • A force that subtracts some constant value from the weight every time.
    • Pushes weights toward 0
      • Efficient for wide models.
      • Reduces # of features -> smaller model size
    • May cause informative features to get a weight of exactly 0:
      • Weakly informative features
      • Strongly informative features on different scales
      • Informative features strongly correlated with other similarly informative features.

Dropout

  • Useful for neural networks.
  • Randomly dropping out unit activations in a network for a single gradient step.
  • 0.0 = No dropout regularization.
  • 1.0 = Drop out everything. Model learns nothing.

Logistic Regression

  • A model that generates a probability for each possible discrete label value in classification problems by applying a sigmoid function to a linear prediction.
  • Often used in binary classification problems, but can also be used in multi-class classification problems (multinomial regression)
  • Sigmoid Function
    • Maps logistic or multinomial regression output (log odds) to probabilities, returning a value between 0 and 1.
    • Can serve as an activation function in neural networks.
  • Loss and Regularization
    • Loss function is Log Loss
    • Regularization
      • L2 or Early Stopping
  • One vs All
    • Classification problem with N possible solutions.
    • A one-vs-all solution consists of N separate binary classifiers.

Classification

  • Classification Threshold (Decision Threshold)
    • Determines what the probability output from logistic regression is classified as.

Accuracy

  • Number of correct predictions over total number of predictions
  • TP + TN / (TP + TN + FP + FN)

Class Imbalanced Dataset

  • Labels have significantly different frequencies in a classification problem.
  • Accuracy is not enough in this scenario.

Confusion Matrix

  • An NxN table that summarizes how successful a classification model’s predictions were.
  • Useful when calculating precision and recall

Precision

  • Identifies the frequency with which the model was correct when predicting the positive class.
  • TP/ (TP + FP)
  • i.e. how many predicted cats are actually cats
  • Raising classification threshold reduces FP, thus improving precision.

Recall

  • Out of all the possible positive labels, how many did the model correctly identify.
  • TP / (TP + FN)
  • i.e. number of predicted cats out of all cats
  • Raising classification threshold will cause # of TP to decrease or stay the same and will cause the # of FN to increase or stay the same. Thus recall will either stay constant or decrease.
  • Improving precision often reduces recall and vice versa.

ROC Curve

  • Receiver Operating Characteristic Curve
  • Shows performance of classification model at all classification thresholds.
  • TP rate (TP / TP + FN) vs. FP rate (FP / FP + TN)
  • Lowering classification threshold increase TP and FP.

AUC

  • Area Under the ROC Curve
  • Provides an aggregate measure of performance across all possible classification thresholds.
  • 0 – worst model
  • 1 – best model
  • Desirable Because:
    • Scale Invariant
      • Measures how well predictions are ranked, rather than their absolute values.
    • Classification Threshold Invariant
      • Measures the quality of the model’s predictions irrespective of what classification threshold is chosen.
  • Limitations
    • Scale invariance is not always desirable
      • We may need well calibrates probability outputs and AUC won’t tell us that.
    • Classification threshold invariance is not always desirable
      • In cases where there are wide disparities in the cost of false negatives vs. false positives, it may be critical to minimize one type of classification error.

Prediction Bias

  • = average of predictions – average of labels
  • Different than bias, b, in wx + b
  • Possible root causes of prediction bias:
    • Incomplete feature set
    • Noisy data set
    • Buggy pipeline
    • Biased training sample
    • Overly strong regularization
  • Avoid Calibration Layer as a fix
    • Fixing symptoms rather than cause.
    • Built a more brittle system that you must now keep up to date.
  • Examine prediction bias on a bucket of examples

Embeddings

  • D-Dimensional Embeddings
    • Assumes something can be explained by d aspects.
  • Map items to low-dimensional real vectors in a way that similar items are close to each other.