1. Machine Learning 101 Intro to AI, ML, Deep Learning
Intro to Machine learning
Intro to Deep learning
Neural networks
Convoluted neural networks
Computational biology/genomics
My try at ML and deep learning
I’ll put together a few slides to present my understanding of various machine learning and deep learning algorithms and how it can be useful to us. I also gave 7 different algorithms a go and applied them to two different datasets.
Richard gave a talk last week: Machine learning applications in genetics and genomics (2015); Nicola gave one on ‘How the machine ‘thinks’’; I gave one last year on DeepSEA – which utilises convoluted neural networks to make predictions on the functional effects of a variant
I’ve read some books on this, I hope to show that we’re actually in the ML world maybe without realising.
Computerphile – Deep Learning
URL:
NVIDIA Blog

2. AI: anything that mimics human intelligence i. e
AI: anything that mimics human intelligence i.e. computations, calculations
ML: Give the PC an algorithm/paradigm/model (i.e. distinguishing features) and it’ll try and predict/classify – this is what we’re doing with our regression models
DL: Give it (clean) data and it’ll extract the features and do the classification
Source URL: towardsdatascience.com

3. Machine Learning 101
“Computers’ ability to learn without being explicitly programmed”
Make predictions using (extracted features from) data
e.g. filters – spam or not?
Supervised v unsupervised methods
Consideration: Categorical v continuous data
We’re surrounded by ML. Just to give a few examples of where ML is used: filters, smart watches, you may be interested, anything with prediction
Data

4. 1- Predicts category/cluster: Classification method (e.g. SVM, CART)
Hidden structure
Labelled data
1- Predicts category/cluster: Classification method (e.g. SVM, CART)
2- Quantify: Regression method (e.g. Linear Regression, LASSO)
1- Group data: Cluster analysis method (e.g. K-means, Hierarchical)
2- Assign value to each data point: Dimensionality reduction methods (e.g. PCA, ICA)
3- Predict outcome probability from categorical data: Bayesian methods (e.g. HMMs)
PCA: Converts a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables
CART: Classification And Regression Trees
SVM: Support vector machine
LASSO: Least Absolute Shrinkage and Selection Operator
ICA: Independent component analysis

5. Many ML algorithms
HMMs
Mention SVM procedure

6. Deep learning 101 Inspired by axons in a biological brain
Application of artificial neural networks to learning tasks that contain more than one hidden layer
Amount of data* α performance
Unlike traditional methods (e.g. logistic regression, SVM) where you observe a plateauing after a certain threshold
Due to larger neural networks being formed
Today: ML becoming synonymous with Deep Learning (e.g. Google, Kaggle) – because DL is outperforming traditional approaches in prediction
*labelled, clean/precise and relevant data

7. Neural networks 101
A simple example: univariate linear regression => one neuron
ReLU: Rectified linear unit
DeepSEA utilises what is called a CNN – and to understand CNNs, you need to know something about neural networks
A. Ng

8. More complicated housing price predictor: multivariate linear regression

9. Convoluted Neural Networks
CNNs are widely used in image recognition
Convolution can be thought as a sliding window function applied to a matrix
Convolution is the process of adding each element of the image to its local neighbours, weighted by the kernel.
In example: Window size 3x3 and stride 1
*Yellow box: Kernel (convolution matrix)

10. Averaging each pixel with its neighboring values blurs an image
Taking the difference between a pixel and its neighbours detects edges
256 x 256 pixels

11. Continued… Kernel(s) Kernel(s) Feature extraction Classification
Feature extraction output: feature vector -> (hopefully!) a summary of all the data
This is then attached to a neural network
Kernel(s)
Kernel(s)
Feature extraction
Classification

12. The classical machine learning workflow: (i) data pre‐processing, (ii) feature extraction, (ii) model learning and (iv) model evaluation.
Supervised machine learning methods relate input features x to an output label y, whereas unsupervised method learns factors about x without observed labels.
Raw input data are often high‐dimensional and related to the corresponding label in a complicated way, which is challenging for many classical machine learning algorithms (left plot). Alternatively, higher‐level features extracted using a deep model may be able to better discriminate between classes (right plot).
Deep networks use a hierarchical structure to learn increasingly abstract feature representations from the raw data.
Computational biology: variant effect predictors (e.g. DeepSEA), genome annotations, PC: population stratification, prediction
Angermueller et al., 2016

13. Working examples Is someone athletic or not? COPD or not?
7 ML algorithms
Logistic regression
LDA
CART
Naïve Bayes
K-nearest neighbour
SVM
RBM feature extraction with logistic regression classifier

14. COPD input
Could have added pack years, and of course lung function measures
Age, sex, ever smoked, air pollution, genetic risk score -> COPD or not

15. Logistic regression

17. K-Nearest neighbours classifier

19. Linear discriminant analysis

21. Appendix

22. Slide by: D. Evans. California Pacific Medical Center