ML for Smart Monkeys

Amit Sethi
Faculty member, IIT Bombay

Image source: Pixabay.com

ML is…
• The practice of automating the use of related data to estimate
models that make useful predictions about new data, where the
model is too complex for standard statistical analysis, e.g.
• Improve accuracy of classification of images using labeled images
• Improve win percentage on alpha-go using several simulated game move
sequences and their results
• Improve the Turing test confusion between human and machine for NLP Q&A
using a large sample of text including Q&A
When not to use ML
• Possible inputs are countable and few
• Use look up tables
• Algorithm is well-known and efficient
• E.g. sorting, Dijkstra's shortest path
• Model is well-known and tractable
• Use statistical estimation
• There is no notion of contiguity
• Use dicrete variable methods or give up
• Lack of data
• Use transfer learning or few-shot learning, or give up
When to use ML
• Possible inputs are many or continuous
• No well-known or efficient algorithm
• Model is not well-known or tractable
• Strong notion of contiguity
• Good amount of data
• Desired output known
• Well-defined inputs
Sweet spot for ML

• Lots of structured data

• Explainability is not critical

• Prediction accuracy is the primary goal

• Underlying model is complex but stationary

Image courtesy: Pixabay.com

 ML model training and
deployment
Training on past data Prediction on future data
 ML gives a model
Supervised Machine Learning System - Training

• Elements of a model: Model Output fθ(xi)

Input xi
• Input xi

(and regularization)
Hyper-parameters

Loss function
• Function fθ(xi)
Parameters θ

• Utility of the model:

• Target output ti Target output ti

• Bring fθ(xi) close to ti Learning algorithm

• Minimize loss L(ti, fθ(xi), θ)
Components of a Trained ML System

Supervised Machine Learning System - Testing

Input xi Model Output fθ(xi)

Hyper-parameters

Parameters θ
Mathematically speaking…
§ Determine f such that ti=f(xi) and g(T, X) is minimized for unseen set T
and X pairs, where T is the ground truth that cannot be used

§ Form of f is fixed, but some parameters can be tuned:

§ So, y=fθ(x), where, x is observed, and y needs to be inferred

§ e.g. y = 1, if mx > c, y = 0 otherwise, so θ = (m,c)

§ Machine Learning is concerned with designing algorithms that learn

“better” values of θ given “more” x (and t) for a given problem

9
 Key

Parameters and Hyperparameters Concept

● Parameters: These are the variable whose values are updated during the
training process of model.
○ Feature coefficient in regression model
○ Weights of a neural network
● Hyperparameters: These are the variables/ parameter whose values are
fixed by model developer before the beginning of learning process.
○ Number of variables in a tree node
○ Height of a tree
○ Number of layers of a neural network
Type of ML problems
● Supervised learning: uses labeled data
○ Classification: Labels are discrete
○ Regression: Labels are continuous
○ Ranking: Labels are ordinal
● Unsupervised learning: uses unlabeled data
○ Clustering: Divide data into discrete groups
○ Dimension reduction: Represent data with fewer numbers
● Somewhere in between: fewer labels than one per example
○ Semi-supervised learning: some examples are labeled
○ Weakly supervised learning: groups of examples are labeled
○ Reinforcement learning: Label (reward) is available after a sequence of steps
Supervised Learning
● Predictor variables/features and a target variable (label)
● Aim: Predict the target variable (label), given the predictor variables
○ Classification: Target variable (y) consists of categories
○ Regression: Target variable is continuous

(Label)
Broad types of ML problems

Output  Categorical Ordinal Continuous

Supervised Classification Ranking Regression

(Examples) {Cats, dogs} {Low, Med, High} [-20,+10)

Unsupervised Clustering Dimension reduction

Some popular ML frameworks

Dimension
Classification Regression Clustering
reduction
Logistic Linear K-means, PCA, k-PCA,
Vector regression regression Fuzzy C-means, LLE,
SVM, RF, NN DB-SCAN ISOMAP
RNN, LSTM, Transformer,
Series, text
1-D CNN, HMM
Images 2-D CNN, MRF
Video, MRI 3-D CNN, CNN+LSTM, MRF
Recipe for ML training

• Decide on the type of the ML problem

• Prepare data

• Shortlist ML frameworks

• Prepare training, validation, and test sets

• Train, validate, repeat

• Use test data only once

Preparing data

• Remove useless data • Handle missing data

• No variance • Impute, if sporadic

• Falsely assumed to be available • Drop, if too frequent

• Reduce redundancy • Transform variables

• Correlated • Convert discrete to one-hot-bit
• Pearson and Spearman • Normalize continuous variables
 Examples of structure in the data
Product SKU Price Margin Volume

• Records A123ajkhdf $ 120 30% 1,000,000

B456ddsjh $200 10% 2,000,000

• Temporal order

• Spatial order

• Web of relationships
Images courtesy: Pixabay.com
Model choice and rigorous
validation are very important

 Underfit Sweet Overfit 

Spot
Error

Validation

Training

Model complexity
Bias-variance trade-off
Generalization of model is bounded by the two undesirable outcomes high
bias and high variance.
● Underfitting: High bias, Low variance
● Overfitting: Low bias, High variance

Bias occurs when an algorithm has limited flexibility to learn the true signal
from the dataset. High bias can cause an algorithm to miss the relevant
relations between features and target outputs (underfitting).
Variance is an error from sensitivity to small fluctuations in the training set.
High variance can cause an algorithm to model the random noise in the
training data, rather than the intended outputs (overfitting).
Regularization is a key concept in ML
● Regularization means constraining the model

● More constraints may reduce model fit on training data

● However, it may improve fit on validation and test data

● Training performance of more constrained models are more likely to

reflect test performance
Loss versus performance metric
● Loss is a convenient expression used for guiding the learning
(optimization)
● Loss is related to performance metric, but it is not the same
● Loss also includes regularization
● Performance metric is what is used to judge the model
● Performance metric on only the held-out (validation or test) data makes
sense
Preparing data for training and validation
• Data splits:
• Training  Used to optimize the parameters (e.g. random 70%)
• Validation  Used to compare models (e.g. random 15%)
• Testing  One final check after multiple rounds of validation (e.g. random 15%)
• Cross-validation:
• K-folds: One fold for validation, K-1 folds for training
• Rotate folds K times
• Select framework (hyperparameters) best average performance
• Re-train best framework on entire data
• Test one final time on held-out data that was not a part of any fold
Cross-validation
● Model performance measurement is dependent on way the data is split
● Not representative of the model’s ability to generalize
● Solution: Cross-validation, especially when data is less
● Con: more computations
ML can fail to perform in deployment
• Lack of training diversity: data had limited confounders
• Single speaker, author, camera, background, accent, ethnicity, etc.
• Data imbalance between high-value rare and more common examples
• Proxy label leak during training:
• E.g. Only speakers A and B provide emotion “anger,” so ML confused their
voice characteristics with “anger”
• Too much manual cleansing of training data
• Too little training data, and very complex models
• Concept drift: The assumptions behind training are no longer valid
ML life stages

Most ML courses
Relation of ML to other fields

Artificial Intelligence

Machine Learning

Neural Networks

Deep Learning
Relation of ML to other fields
Probability
Machine and Statistics
Learning
Optimization

Programming
Linear Data Science
Algebra