Machine Learning

Machine Learning#

  1. Basics

  2. ML Models

  3. Workflow

back to top

Basics#

programmed * computers learn from data to discover patterns and make predictions

Supervised Learning#

  • type of ML technique

  • every training sample from the dataset has corresponding label or output value

  • the algorithm learns to predict labels

  • e.g. predict sales price or classify object in an image

  • discrete: term from statistics referring to an outcome that takes only a finite number of

values * label: data that already contains the solution * Categorical Label

  • has a discrete set of possible values

  • performing a classification task

  • Continuous Label

    • does not have a discrete set of possible values, potentially unlimited number of

    possibilities - performing a regression task

Unsupervised Learning#

  • type of ML technique, no labels for training data

  • algorithm tries to learn underlying patterns or data distribution

  • unlabeled data: do not provide the model with any kind of label or solution while the

model is being trained * Clustering

  • to determine if there are any naturally occurring groupings in the data

  • in Supervised and Unsupervised learning, models inspect data to discover patterns and humans

use the patterns learnt by the model to gain new understandings or make predictions

Reinforcement Learning#

  • learning what actions to take in a situation to maximize reward

  • learns through consequences of actions in an environment

  • Agent: entity being trained

  • Environment: the world where agent perform certain actions

  • Rewards: given to agent for performing good actions

  • e.g. traffic signaling, self-driving cars

  • can give reward for good actions, punish for poor actions, or combine both approaches

  • rewarding a good action will increase the probability of doing it again, and punishing a

poor action, such as removing it, will ensure it won’t do it again * Policy: define actions an agent should take for a given state * Deterministic Policy

  • direct relationship between action and state

  • use when agent has a full understanding of the environment

  • will always perform the same action for a given state

  • Stochastic Policy

    • range of possible actions with probabilities for a state

    • an action for a state is selected based on the probability distribution

  • Value Function

    • determine possible future rewards given the current policy

    • adjust to encourage desirable actions while discouraging others, also called policy

    update

  • PPO

    • Proximal Policy Optimisation, uses on-policy learning

    • learn only from observations made by current policy, most recent and relevant data

    • need more data as it does not consider historical data

    • can produce more stable model in short-term

  • SAC

    • Soft Actor Critic, uses off-policy learning

    • use observations from previous policies, old data

    • data efficient, need less new data as it consider historical data

    • can produce less stable model in short-term

  • traditional programming requires humans to make a program for solving a problem

  • machine learning is created using statistics, applied math, and computer science, but each

fields might use different formal definitions * machine learning has a flexible component called the model * ML model is as a block of code or framework, that can be modified to solve different but related problems, based on provided data

back to top

ML Models#

most common libraries

Linear Models#

  • simply describe the relationship between a set of inputs to outputs through a linear

function * classification tasks often use a strongly related logistic model, which adds an additional transformation mapping the output of the linear function to the range [0, 1] * fast to train and give a great baseline against which to compare more complex models * it is better to start with a simple model for most new problems

Tree-based Models#

  • learn to categorize or regress by building an extremely large structure of nested if/else

blocks * split the world into different regions at each if/else block * training determine where the splits happen and what value is assigned at each leaf region * e.g XGBoost is commonly used as an off-the-shelf implementation * try tree-based models to quickly get a baseline before moving to more complex ones

Deep Learning Models#

  • popular and powerful, also called neural networks

  • composed of collections of neurons, simple computational units/models, connected together

by weights * weights: mathematical representations of how much information is allowed to flow from one neuron to the next * training involves finding values for each weight * FFNN

  • Feed Forward Neural Network, most straightforward way of structuring neural network

  • structures neurons in a series of layers

  • each neuron in a layer contain weights to all neurons in the previous layer

  • CNN

    • Convolutional Neural Network

    • represent nested filters over grid-organized data

    • most commonly used type of model when processing images

  • RNN/LSTM

    • Recurrent Neural Networks and the related Long Short-Term Memory

    • to effectively represent for loops in traditional computing, collecting state while

    iterating over some object - can be used for processing sequences of data

  • Transformer

    • more modern replacement for RNN/LSTMs

    • enables training over larger datasets involving sequences of data

back to top

Workflow#

Defining Problem#

  • define a specific task

  • identify the ML task to use to solve the problem, as it helps understand the data needed

better * Machine Learning Tasks

  • output of a task can be different and classified into different groups based on the

task - characteristics of the input data can help to define which ML task to be used - Supervised and Unsupervised learning are two common tasks

Building Dataset#

  • working with data is the most important step

  • ML practitioners spend 80% of the time preparing the data

  • understanding the data helps select better models and algorithms to have effective ML

solutions * good, high-quality data is essential for any kind of machine learning project * impute: common term referring to different statistical tools that can be used to calculate missing values from dataset * outliers: data points that are significantly different from other data in the same sample * Data Collection

  • collect data related to the problem defined

  • search and use publicly available data in early exploration

  • the format of data, labeled or unlabeled, and availability of data will determine

the ML task

  • Data Inspection

    • inspect the integrity of the data, not all data found are high quality

    • quality of the data has a massive impact of how well the model performs

    • identify anything that are outside the norm

    • look for missing or incomplete data

    • transform or pre-process the dataset to correct format to be used by the model

    • stop words: punctuation and words that don’t have useful meaning, e.g. a, the

    • data vectorisation: convert non-numeric data, usually text, into numbers

    • bag of words: count how many times a word appears in a document

  • Summary Statistics

    • helps understand what the data is communicating or in line with the underlying

    assumptions of the ML model - allows to see insights such as scope, scale and shape of the data - there are many tools to calculate things such as mean, IQR (inner-quartile range) and standard deviation

  • Data Visualization

    • communicate the findings, such as outliers and trends, to project stakeholders

Model Training#

  • procedure to use data to shape a model

  • uses the model to process data, compares the result against the goal

  • determine what changes are required and makes small changes to the model parameters

  • the steps are repeated to bring the model closer to achieving the goal

  • model training algorithm adjusts the model to real-world data

  • the trained model can be used to predict outcomes which are not part of training dataset

  • Splitting Dataset

    • randomly split the dataset before training

    • majority of the data is in training dataset, usually 80%

    • test dataset is withheld from training and used later to evaluate the model before

    production - test the data against the bias-variance trade-off and how well the model will generalize to new data

  • feed the training data into the model, compute the loss function on the results, and

iteratively updating model parameters to minimize some loss function

  • models are trained by slowly changing model parameters to move it closer to some goal

  • Model Parameters

    • settings, knobs, configurations, weights or biases that are updated to change how

    the model behaves - weights are values that change as the model learns, specific to neural networks

  • Loss Function

    • measure how close the model is to its goal

    • used to codify the model’s distance from a goal

  • only implement model training algorithms from scratch when developing new ones

  • often use the existing frameworks that have working implementations

  • model selection: to determine which model to use, try different types while solving a

problem * Hyperparameters

  • not changed during model training

  • can affect how quickly and reliably the model trains

  • e.g number clusters to identify

Model Evaluation#

  • use statistical metrics to evaluate a model, e.g. accuracy, recall, precision, log loss,

mean absolute error, hinge loss, quantile loss, $R^2$, KL Divergence, F1 Score * Metrics

  • accuracy: how often the model predicts correctly

  • log loss: to understand model’s uncertainty about a given prediction, how strongly

the model believes its prediction is accurate - silhouette coefficient: shows how well the data is clustered by the model, score near 0 show overlapping clusters, score < 0 show data points assigned to incorrect clusters, and score near 1 show successful identification

Model Inference#

  • using a trained model to generate predictions and solve a problem

  • make sure to monitor the results produced

  • may need to reinvestigate the data, modify parameters in training algorithm, or change

the model type for training * Inference Rate

  • number of inferences per second, need to maximise

  • depend on performances such as CPU, GPU, RAM, and machine learning framework

  • cost-effective to centralise in one place, where data is fed from edge devices to a

central server for processing - inference at edge, performing inference locally is crucial for real-time devices

  • Inference Time/Latency

    • time taken to run a single inference

    • inference at edge reduces latency as data does not need to travel to the server for

    processing

back to top