Machine Learning Through the Pattern AI Lens
Seeing the Pattern Beneath the Algorithm
Machine Learning is often introduced through algorithms.
Students learn Linear Regression, Logistic Regression, Decision Trees, K-Nearest Neighbours, Support Vector Machines, Naive Bayes, Random Forests, Clustering, PCA, Boosting and Neural Networks.
They learn formulas.
They write Python programs.
They import libraries.
They train models.
They calculate accuracy.
But beneath all these apparently different techniques lies a remarkably useful way of seeing Machine Learning:
Machine Learning is Pattern Learning.
This is not proposed here as a replacement for the formal mathematical definition of Machine Learning. It is a Pattern AI Lens—a conceptual framework for understanding what many ML methods are accomplishing underneath their algorithms and mathematics.
A more rigorous formulation is:
Machine Learning, through the Pattern AI Lens, can be understood as the computational process of learning useful statistical, structural, or representational regularities from data, representing them through a model, and applying that model to new inputs to produce useful outputs.
This distinction is important.
Machine Learning provides the mathematics and computational mechanisms. Pattern AI provides a way of seeing the common structure underneath them.
From Algorithm-First to Pattern-First Thinking
A conventional introduction to Machine Learning often looks something like this:
DATA → ALGORITHM → MODEL → PREDICTION
There is nothing wrong with this representation.
But it hides an important conceptual layer.
What is the algorithm trying to learn from the data?
Through the Pattern AI Lens, we can expand the pipeline:
DATA
↓
DISCOVER / LEARN USEFUL STRUCTURE
↓
REPRESENT THAT STRUCTURE
↓
MODEL
↓
NEW INPUT
↓
APPLY THE LEARNED MODEL
↓
OUTPUT
↓
EVALUATE
The word pattern becomes a useful umbrella for statistical regularities, relationships, similarities, boundaries, clusters, distributions, representations and other learnable structure.
This gives us a deeper interpretation:
Data provides examples.
The learning process searches for useful structure.
The model represents what has been learned.
Inference applies the model to new cases.
Evaluation determines whether what was learned generalizes.
That is the essence of the Pattern AI Lens.
What Does “Pattern” Mean Here?
We must be precise.
When people hear the word pattern, they sometimes imagine something visually obvious:
2, 4, 6, 8, ...
or perhaps a repeated design.
That meaning is far too narrow for Machine Learning.
Within the Pattern AI Lens, pattern is an accessible umbrella term for useful learnable structure or regularity in data.
That could include:
relationships between variables,
similarities between observations,
probability distributions,
decision boundaries,
clusters,
associations,
temporal dependencies,
useful representations,
combinations of features,
or regularities connecting situations, actions and outcomes.
A learned pattern also need not be easily interpretable by a human.
A neural network containing millions or billions of parameters can encode complex distributed representations that cannot be reduced to one simple human-readable rule.
The Pattern AI Lens therefore does not claim:
“Every ML model discovers a simple pattern that humans can see.”
It makes the broader claim:
Machine Learning systems learn exploitable structure from data. Pattern is a useful conceptual language for thinking about that structure.
That distinction keeps the framework technically grounded.
The Pattern Comes Before the Algorithm
Imagine a student is given a new Machine Learning problem.
The usual instinct is:
“Which algorithm should I use?”
Pattern AI suggests asking an earlier question:
“What kind of pattern or structure am I trying to discover?”
Consider several problems.
| Problem | Pattern AI question |
|---|---|
| Predict house price | What relationship connects characteristics to price? |
| Detect spam | What characteristics distinguish spam from non-spam? |
| Segment customers | Which customers exhibit similar behaviour? |
| Detect fraud | What behaviour deviates significantly from expected behaviour? |
| Recommend products | What preferences or associations connect users and products? |
| Forecast demand | What temporal regularities help predict future demand? |
| Recognize images | What representations distinguish one visual category from another? |
Suddenly Machine Learning stops looking like a catalogue of algorithms.
It becomes a collection of different mechanisms for exploiting different forms of structure.
Linear Regression: Relationship Pattern
Consider one of the simplest ML techniques: Linear Regression.
Google's Machine Learning Crash Course describes linear regression as a statistical technique for finding the relationship between variables. In an ML context, it models the relationship between features and a numerical label. During training, weights and bias are adjusted to produce a model that minimizes prediction error.
Pattern AI therefore interprets Linear Regression as:
RELATIONSHIP PATTERN
Suppose we have:
House size → House price
The algorithm tries to learn a useful mathematical relationship between the input and output.
The Pattern AI view is:
EXAMPLES
↓
RELATIONSHIP
↓
MODEL
↓
NEW HOUSE
↓
PREDICTED PRICE
This is considerably easier for a beginner to grasp before introducing the mathematical machinery.
The mathematics then explains how the relationship is represented and learned.
Logistic Regression: Probability Pattern
Logistic Regression is different.
Google describes logistic regression as a model designed to predict the probability of an outcome. Its sigmoid function maps model output into a value between zero and one, which can then be used directly as a probability estimate or converted into a binary classification.
Through the Pattern AI Lens:
LOGISTIC REGRESSION → PROBABILITY / DECISION PATTERN
Imagine spam detection.
Email characteristics
↓
Learn relationship to probability of spam
↓
New email
↓
Probability of spam
↓
Decision, if required
This distinction is useful because students often see Linear Regression and Logistic Regression as merely two chapters with two different equations.
Pattern AI gives them a conceptual contrast:
Linear Regression → numerical relationship
Logistic Regression → probability of an outcome
KNN: Similarity Is the Pattern
K-Nearest Neighbours provides perhaps one of the clearest examples of the Pattern AI Lens.
According to the scikit-learn documentation, nearest-neighbour methods identify training samples closest to a new point and use those neighbours to predict its label. KNN classification can use the majority class among the nearest neighbours.
So KNN becomes:
SIMILARITY / NEIGHBOURHOOD PATTERN
The reasoning is wonderfully intuitive:
NEW CASE
↓
WHO IS SIMILAR?
↓
LOOK AT NEIGHBOURS
↓
WHAT HAPPENED TO THEM?
↓
PREDICT
There is an interesting technical nuance here.
scikit-learn describes nearest-neighbour methods as non-generalizing or instance-based methods because they can retain training instances rather than constructing the kind of compact parametric internal model seen in methods such as linear regression.
This nuance strengthens rather than weakens the Pattern AI perspective.
It demonstrates that:
Different ML methods exploit structure in fundamentally different ways.
SVM: Boundary Pattern
Support Vector Machines offer another way of looking at classification.
Rather than asking primarily:
“Who are my neighbours?”
an SVM classification approach can be understood as asking:
“Where should the separating boundary lie?”
Pattern AI therefore characterizes the central intuition as:
SVM → BOUNDARY / SEPARATION PATTERN
This gives students a powerful comparison.
KNN
Similarity → Neighbours → Vote
SVM
Classes → Separating boundary → Classification
Both may solve classification problems.
But they exploit different structural ideas.
This leads to one of the most important principles of Pattern AI applied to ML:
SAME TASK → DIFFERENT PATTERNS → DIFFERENT LEARNING MECHANISMS
That is why algorithm selection matters.
Clustering: Discovering Groups Without Labels
Supervised learning gives the system examples paired with targets.
But what happens when there are no supplied target labels?
Consider customer data containing:
purchase frequency,
transaction value,
products purchased,
visit frequency,
engagement behaviour.
We might not know the customer categories in advance.
A clustering algorithm attempts to uncover structure in the observations.
Through the Pattern AI Lens:
CLUSTERING → GROUPING / STRUCTURAL PATTERN
The conceptual flow becomes:
UNLABELLED DATA
↓
SIMILARITY / STRUCTURE
↓
DISCOVER GROUPS
↓
INTERPRET GROUPS
The key difference is powerful:
Supervised Learning
Examples + target labels → learn task-relevant mapping
Unsupervised Learning
Observations without target labels → discover useful structure
Pattern AI gives students a simple conceptual bridge between the two.
PCA: Dominant Variation Pattern
Principal Component Analysis is another concept that becomes easier when viewed through patterns—but terminology matters.
It would be tempting to say:
“PCA discovers the most important information.”
That is too broad.
scikit-learn describes PCA as linear dimensionality reduction and identifies its principal axes as directions of maximum variance in the data. Components are ordered by the amount of variance they explain.
Therefore the more technically responsible Pattern AI description is:
PCA → DOMINANT VARIATION PATTERN
Conceptually:
MANY FEATURES
↓
IDENTIFY DIRECTIONS OF MAJOR VARIATION
↓
TRANSFORM REPRESENTATION
↓
FEWER DIMENSIONS
↓
RETAIN AS MUCH RELEVANT VARIATION AS THE SELECTED COMPONENTS ALLOW
This is a good example of why Pattern AI must remain anchored in ML science.
A memorable metaphor is useful only when it does not distort the underlying mathematics.
Deep Learning: Patterns of Representation
Deep Learning makes the Pattern AI Lens particularly interesting.
A common oversimplification is:
“Deep Learning finds more complicated patterns.”
There is truth in that statement, but we can do better.
Goodfellow, Bengio and Courville describe representation learning as allowing machine-learning systems to discover representations useful for subsequent tasks. They explain deep learning as building complex concepts from simpler representations through multiple layers.
A simplified image-recognition intuition might be:
PIXELS
↓
EDGES / SIMPLE FEATURES
↓
MORE COMPLEX REPRESENTATIONS
↓
OBJECT-RELATED REPRESENTATIONS
↓
CLASSIFICATION
The technically safer Pattern AI interpretation is therefore:
DEEP LEARNING → LAYERED / HIERARCHICAL REPRESENTATION PATTERNS
The important word is representation.
The system is not merely finding one obvious rule.
It is learning transformations and representations that make the target task possible.
This leads to another Pattern AI insight:
Sometimes learning the right representation is itself a major part of learning the useful pattern.
Representation Determines What Is Easy to Learn
This may be one of the deepest connections between Pattern AI and established Machine Learning.
Goodfellow, Bengio and Courville note that information-processing tasks can become easier or harder depending on how information is represented, and describe a good representation as one that makes a subsequent learning task easier.
Pattern AI can translate this into a memorable proposition:
Representation affects Pattern Visibility.
This does not mean that representation literally makes a hidden pattern visually appear.
It means that a suitable representation can make useful structure more accessible to a learning algorithm.
That gives us:
REAL-WORLD PHENOMENON
↓
DATA COLLECTION
↓
REPRESENTATION / FEATURES
↓
LEARNABLE STRUCTURE
↓
MODEL
This is why data representation and feature engineering matter enormously.
What Exactly Is a Model?
The Pattern AI Lens also requires care around the word model.
It is attractive to say:
“A model is a stored pattern.”
That is memorable, but technically too loose.
Different ML models can contain very different mathematical objects:
coefficients,
weights,
biases,
tree structures,
support vectors,
probability parameters,
learned components,
neural-network parameters,
and other representations.
Google's Linear Regression material, for example, explicitly distinguishes model parameters such as weights and bias, which are learned during training.
A stronger formulation is:
A trained ML model is a mathematical or computational representation learned from data that captures structure useful for performing a task.
Pattern AI can then provide the intuitive version:
The model carries forward what the learning process has captured from the training data.
That retains the simplicity without sacrificing technical accuracy.
Training Is Not the End: Generalization Is the Test
Perhaps the most important contribution of the Pattern AI Lens is the way it makes generalization intuitive.
Suppose a model performs brilliantly on its training examples.
Has it learned something useful?
Not necessarily.
Google's ML Crash Course emphasizes that a model should make good predictions on new data. It defines overfitting as fitting the training data so closely that performance on new data suffers.
Pattern AI translates this into a fundamental question:
Does the learned pattern survive beyond the examples from which it was learned?
That gives us three possibilities.
UNDERFITTING
The model has not captured enough useful structure.
GOOD GENERALIZATION
The learned structure transfers adequately to new data from the relevant distribution.
OVERFITTING
The model has fit training-specific structure or noise too closely and performs poorly on new data.
Google identifies overly complex models and training data that inadequately represent real-world data as broad causes of overfitting.
Therefore:
Finding a pattern is not enough. ML seeks models whose learned structure generalizes usefully.
This is a critical refinement of the slogan “Machine Learning is Pattern Learning.”
Data Quality Determines Pattern Quality
Pattern AI also makes another ML reality obvious.
If the learning process depends on data, the quality and representativeness of that data matter enormously.
Google's current ML curriculum places substantial emphasis on dataset construction, data reliability, transformation, feature engineering, training/validation/test splits and generalization.
Pattern AI can summarize the dependency as:
WORLD
↓
OBSERVATIONS
↓
DATA
↓
LEARNED STRUCTURE
↓
MODEL BEHAVIOUR
If the observations are incomplete, systematically skewed, incorrectly measured or unrepresentative, the resulting model may learn undesirable or unreliable relationships.
So:
A model can only learn from the evidence made available to it.
More data is not automatically the answer either.
The composition, quality and relevance of that data matter.
Pattern Does Not Mean Truth
This deserves special emphasis.
Suppose a machine discovers a strong statistical relationship.
Does that mean the relationship is true in every context?
No.
Does it mean the relationship will continue forever?
No.
Does it establish causation?
No.
Does it mean we should use that relationship to make consequential decisions?
Again, no.
The Pattern AI Lens therefore needs a hierarchy:
PATTERN EXISTS IN AVAILABLE DATA
↓
PATTERN IS STATISTICALLY USEFUL
↓
PATTERN GENERALIZES TO RELEVANT NEW DATA
↓
PATTERN IS INTERPRETABLE
↓
PATTERN IS CAUSAL?
↓
PATTERN IS APPROPRIATE TO USE?
These are different questions.
Machine Learning can provide strong predictive performance without necessarily explaining the causal mechanisms behind the phenomenon.
That distinction prevents a dangerous mistake:
Prediction is not automatically explanation.
Bias: A Discoverable Pattern May Be an Undesirable Pattern
Pattern AI becomes especially valuable when discussing responsible Machine Learning.
Google's ML fairness material emphasizes that ML models are not automatically objective. Human choices in collecting and curating training data can expose models to biases, including historical, reporting and selection biases.
Google also recommends auditing data and evaluating model behaviour by relevant subgroups rather than relying solely on aggregate metrics.
Pattern AI translates this into a powerful principle:
DISCOVERABLE PATTERN ≠ DESIRABLE PATTERN
Historical data may contain genuine statistical regularities.
But those regularities may reflect:
historical inequalities,
sampling problems,
measurement practices,
institutional processes,
missing populations,
or other undesirable conditions.
The ML system may faithfully learn such structure.
That does not mean society should faithfully reproduce it.
Therefore:
PATTERN DISCOVERY
must be followed by
PATTERN JUDGMENT
This is where human responsibility enters the ML lifecycle.
Evaluation Is Pattern Judgment
Students often encounter evaluation metrics as formulas:
Accuracy
Precision
Recall
F1
MAE
MSE
RMSE
ROC-AUC
But Pattern AI asks:
What are these metrics really helping us determine?
They provide evidence about whether model behaviour is useful for the task and data being evaluated.
Google defines loss as a numerical measure of how wrong predictions are relative to actual labels and describes model training as seeking to reduce loss.
But no single metric answers every deployment question.
Consider two applications:
Movie recommendation
versus
Cancer screening
The cost of different errors can be dramatically different.
So evaluation is not merely:
“What is the accuracy?”
It becomes:
Is the model's behaviour good enough, robust enough, appropriate enough and fair enough for this particular purpose?
Pattern AI calls this:
JUDGE
A Pattern Taxonomy of Machine Learning
Once we adopt the Pattern AI Lens, apparently unrelated algorithms begin to form a coherent conceptual map.
| ML Technique | Pattern AI Interpretation |
|---|---|
| Linear Regression | Relationship Pattern |
| Logistic Regression | Probability / Decision Pattern |
| Decision Tree | Decision-Rule Pattern |
| KNN | Similarity / Neighbourhood Pattern |
| SVM | Boundary / Separation Pattern |
| Naive Bayes | Probabilistic Pattern |
| Clustering | Grouping / Structural Pattern |
| PCA | Dominant Variation Pattern |
| Random Forest | Ensemble of Decision Patterns |
| Boosting | Sequential Improvement Pattern |
| Neural Networks | Layered Representation Patterns |
| Reinforcement Learning | State–Action–Reward Pattern |
This table should not replace the formal definitions of these methods.
Its purpose is different.
It provides students with a cognitive map.
Instead of memorizing twelve unrelated algorithms, the learner begins asking:
What type of structure is this method exploiting, and how does its mechanism differ from another method solving the same task?
That question leads naturally toward deeper mathematical understanding.
Same Task, Different Patterns
Suppose we need to classify an observation.
Three algorithms could approach it differently.
KNN
Neighbourhood → Similarity → Vote → Class
SVM
Data → Separating Boundary → Margin → Class
Naive Bayes
Evidence → Conditional Probabilities → Compare → Class
The output may be the same:
Class A
But the reasoning mechanism is different.
This reveals a profound ML principle:
Algorithms are not merely different pieces of software. They embody different assumptions and mechanisms for extracting or exploiting structure.
That is why blindly asking an AI:
“Give me the best ML algorithm”
is usually the wrong starting point.
We first need to understand the problem, data, assumptions, objective and evaluation criteria.
Pattern-First Machine Learning
Pattern AI therefore suggests a different sequence for approaching ML problems.
Instead of:
ALGORITHM → CODE → OUTPUT
begin with:
1. PROBLEM
What am I trying to accomplish?
↓
2. DATA
What evidence do I have?
↓
3. PATTERN
What kind of structure might connect the evidence to the outcome?
↓
4. MECHANISM
What kind of learning mechanism could exploit that structure?
↓
5. ALGORITHM
Which algorithm is appropriate given the data, assumptions and objective?
↓
6. MODEL
What has the training process actually learned?
↓
7. GENERALIZE
Does performance transfer to appropriate unseen data?
↓
8. JUDGE
Are the results reliable, useful and appropriate?
↓
9. ACCOMPLISH
Does using the model actually solve the intended problem?
This takes us beyond Machine Learning as a programming exercise.
It turns ML into structured problem-solving.
Pattern AI Meets Think → Ask → Judge
Pattern AI also provides a natural way for students to learn ML with AI assistance.
Instead of asking an AI system:
“Explain Random Forest.”
students can use a three-stage learning discipline.
THINK
Ask yourself first:
What type of problem am I trying to solve?
What kind of pattern might matter?
What do I already understand?
For example:
“I understand Decision Trees. Why might one tree be unreliable?”
ASK
Now use AI as a tutor.
A stronger prompt would be:
“Teach me Random Forest by first connecting it to Decision Trees, which I already understand. Explain what changes when many trees are combined. Ask me questions as we proceed instead of giving me everything at once.”
The AI becomes part of the learning process rather than merely an answer generator.
JUDGE
Finally:
Can I explain Random Forest without AI?
Can I compare it with a Decision Tree?
Can I identify situations in which one might be preferable?
Can I justify my reasoning using appropriate evidence?
That final question is crucial.
AI can help students encounter patterns.
Understanding requires students to judge and reconstruct those patterns themselves.
The Pattern AI Machine Learning Cycle
We can now express the complete framework.
OBSERVE
Gather examples from the world.
↓
REPRESENT
Transform observations into data usable by the learning process.
↓
DISCOVER / LEARN
Use an appropriate algorithm to identify useful statistical or representational structure.
↓
REPRESENT IN A MODEL
Capture the learned structure computationally.
↓
APPLY
Use the model on new inputs.
↓
EVALUATE
Measure performance on appropriate unseen data.
↓
JUDGE
Examine generalization, errors, assumptions, fairness, limitations and suitability.
↓
ACCOMPLISH
Use the model only where it meaningfully contributes to the intended real-world outcome.
↓
LEARN AGAIN
New evidence and feedback can lead to better data, representation, modelling and evaluation.
Machine Learning therefore becomes a learning loop, not simply an algorithm.
Pattern AI Does Not Replace Machine Learning Mathematics
This boundary is essential.
Pattern AI does not eliminate the need to understand:
probability,
statistics,
linear algebra,
calculus,
optimization,
algorithms,
data structures,
programming,
model evaluation,
experimental design,
domain knowledge.
Nor does calling SVM a “boundary pattern” mean a student understands SVM.
The learner must eventually understand margins, kernels, optimization and the relevant mathematical foundations.
Similarly:
PCA = Dominant Variation Pattern
is a conceptual entry point.
It does not replace covariance, eigenvectors, eigenvalues or SVD.
Pattern AI therefore operates at another level.
Mathematics explains HOW.
Algorithms specify HOW COMPUTATION PROCEEDS.
Pattern AI helps students see WHAT KIND OF STRUCTURE THE METHOD IS TRYING TO EXPLOIT.
Together, they produce deeper understanding.
A Better Way to Learn Machine Learning
This suggests a pedagogical change.
Instead of introducing an ML chapter as:
“Today we are going to learn KNN.”
we might begin:
“Today we are going to explore how similarity can become a prediction mechanism.”
Then introduce KNN.
Instead of:
“Today we learn SVM.”
begin:
“How could classification be performed by finding a good boundary between classes?”
Then introduce SVM.
Instead of:
“Today we learn PCA.”
ask:
“If our data has many dimensions, can we identify directions that capture much of its variation?”
Then introduce PCA.
The algorithm now becomes the answer to an intellectual problem, rather than an isolated item in a syllabus.
That is the educational potential of Pattern AI.
From Algorithm Literacy to Pattern Literacy
Students certainly need algorithm literacy.
But they also need something broader:
Pattern Literacy
Pattern literacy means developing the ability to ask:
What structure exists here?
What structure matters?
What evidence supports it?
How might an algorithm capture it?
Will it generalize?
What assumptions produced it?
When will it fail?
Should we use it?
Eventually, students may move through three stages:
Pattern Curiosity
“What pattern might exist here?”
↓
Pattern Fluency
“I can identify and compare different kinds of structure.”
↓
Pattern Instinct
“When I encounter a new problem, I naturally begin searching for the underlying structure before jumping to an algorithm.”
That may be far more valuable than memorizing an ever-growing list of algorithms.
The Most Important Insight
After fact-checking the framework against established Machine Learning sources, the strongest position is not:
“Pattern AI has created a new definition of Machine Learning.”
That claim would be unnecessary and difficult to defend.
The stronger position is:
Pattern AI provides a conceptual lens for seeing the common structure beneath Machine Learning techniques.
Machine Learning already has rigorous mathematical foundations.
Pattern AI asks us to look across those foundations and notice something educationally powerful:
Regression learns relationships.
KNN exploits neighbourhood similarity.
SVM exploits separating boundaries.
Clustering searches for grouping structure.
PCA identifies dominant directions of variation.
Deep Learning learns useful hierarchical representations.
Different mechanisms.
Different mathematics.
Different assumptions.
But a common intellectual question remains:
What useful structure can be learned from the available evidence, and can that learned structure help us deal with new cases?
That is the Pattern AI Lens.
A Definition Worth Keeping
A technically careful definition is:
Machine Learning, through the Pattern AI Lens, can be understood as the computational process of learning useful statistical, structural, or representational regularities from data, representing them through a model, and applying that model to new inputs to produce useful outputs.
And for teaching, communicating and remembering:
Machine Learning is Pattern Learning.
Then comes the qualification that makes the statement powerful rather than simplistic:
Good Machine Learning is not merely finding patterns in existing data. It is learning useful structure that generalizes appropriately to new situations—and judging whether that structure should be trusted and used.
The Pattern AI View in One Picture
REAL WORLD
↓
DATA
Evidence from observations
↓
PATTERN
What useful structure exists?
↓
ALGORITHM
How can that structure be learned?
↓
MODEL
How is what was learned represented?
↓
NEW INPUT
Can it handle something it has not seen before?
↓
INFERENCE
Apply the learned model
↓
EVALUATION
How well does it perform?
↓
JUDGE
Does it generalize? Is it reliable, appropriate and fair?
↓
ACCOMPLISH
Does it create the intended real-world value?
Conclusion
Machine Learning can initially appear to students as a forest of algorithms.
Linear Regression.
Logistic Regression.
KNN.
SVM.
Decision Trees.
Random Forest.
Naive Bayes.
Clustering.
PCA.
Boosting.
Neural Networks.
Learn them individually and the subject can feel fragmented.
Look through the Pattern AI Lens, however, and a unifying structure begins to appear.
Data contains evidence.
Learning algorithms exploit structure in that evidence.
Models represent what has been learned.
Inference applies the model to new cases.
Evaluation tests whether performance generalizes.
Human judgment determines whether the result should be trusted and used.
So perhaps the most useful question for someone beginning Machine Learning is not:
“Which algorithm should I learn first?”
It is:
“What pattern am I trying to learn?”
Once that question becomes habitual, algorithms stop being disconnected formulas.
They become different ways of learning from the patterns and structure hidden within data.
And that may be one of the simplest ways to begin truly understanding Machine Learning.
Sources & Further Reading
Google Machine Learning Crash Course — Google's current practical curriculum covering regression, classification, data representation, generalization, overfitting, neural networks and fairness.
Google Machine Learning Crash Course
Google — Linear Regression — supports the interpretation of linear regression as modelling relationships between features and labels and explains learned weights and bias.
Google: Linear Regression
Google — Logistic Regression — describes logistic regression as predicting probabilities of outcomes.
Google: Logistic Regression
scikit-learn — Nearest Neighbors — documents nearest-neighbour learning and KNN classification based on nearby training samples.
scikit-learn: Nearest Neighbors
scikit-learn — Principal Component Analysis — documents PCA as dimensionality reduction based on principal axes representing directions of maximum variance.
scikit-learn: PCA
Goodfellow, Bengio & Courville — Deep Learning — foundational treatment of Machine Learning, representation learning and hierarchical/deep representations.
Deep Learning Book
Google — Generalization and Overfitting — explains why performance on new data, rather than merely fitting training data, is fundamental to useful ML.
Google: Overfitting and Generalization
Google — ML Fairness — discusses biases in training data, auditing data and predictions, and responsible model evaluation.
Google: Machine Learning Fairness
Pattern AI terminology such as Relationship Pattern, Similarity Pattern, Boundary Pattern, Dominant Variation Pattern, Pattern Visibility, Pattern Literacy, and the proposition “Machine Learning is Pattern Learning” are used in this article as conceptual/pedagogical interpretations. They should not be read as replacements for the formal terminology and mathematical definitions of Machine Learning.
No comments:
Post a Comment