3. What is machine learning
Articificial
Intelligence
Machine
Learning
Deep
Learning
Broad field of study dedicated
to complex problem solving
Artificial Intelligence (AI)
Machine learning (ML)
Data driven approach focused on creating
algorithm that has the ability to learn from
data without being explicitly programmed
Subfield of ML focused on deep neural
networks
Deep learning (DL)
4. 3 types of machine learning
Machine learning
KNN, Trees, logistic
regression, Naïve
Bayes, SVM
Linear/Polynomial
regression, Decision
trees, Random Forest
Supervised learning
Learn a general rule that maps
inputs and outputs
Unsupervised learning
Find hidden pattern, feature
learning
Reinforcement learning
Rewards/punishments
Provide inputs and their
desired outputs
y = f(x)
Datasets without labels
f(x)
Making an agent
interacting with a
dynamic environment
y = f(x) given z
Categorical Continuous
Association analysis,
hidden markov model
SVD, PCA, Kmeans
Categorical Continuous
5. 3 types of machine learning
Supervised learning
Predict
Unsupervised learning
Improve data quality, cluster
Reinforcement learning
Drive an agent in a dynamic environment
11. Why is machine learning hot?
Data
analysis
algorithms
Big data
Cheap and
fast GPUs
and RAM
Machine
learning
Data
analysis
algorithms
Data
Data
mining
12. ML use cases / lessons learned
• Business will not ask for it. They just want better result, process…
• Skillset is wide but data science platforms will foster the raise of citizen data scientist
• Better to start on with simple data/model set with a human assistance objective
• Integrated platform speeds up innovation and avoid spreading data around
• Models are assets –proven model should be captured, protected and shared as a service
Machine learning (ML) is usually considered as a subfield of Artificial Intelligence (AI)
Traditionally programming has been about defining every single step for a program to reach an outcome
With machine learning we provide the outcome and the program learns the steps to get there
Deep learning is a type of machine learning using Deep Neural Networks algorithms
There is 3 types of machine learning:
Starting right, reinforcement learning is programming an agent to interact with dynamic environment. Feeding it with penalty and reward when it is learning
Unsupervised learning is close to what we use to call data mining. We have the output and we want to analyse it in order to create new features, informations, we will see an example of clustering in a minute
Supervised learning is providing an output knowing the input for the machine to learn the best function with the less error, the best generalizing
Here are examples of algorithms and applications
On the left we have supervised algorithms like Linear regression
Linear regression is the most used model at Uber!
Decision trees
KNN is also an interesting classification algorithm that calculates the distances between the point you try to predict and its closest know neighbors. In a multi dimensional space of features
In the center, clustering of values
And at the right we teach a machine to play Mario Bros with renforcement learning
Model readability is also very important. You will probably feel more confortable explaining the decision tree.
And actually it bcomes a regulatory issue as Insurance refuses to use deep learning model as they must be able to explain why they refuse a customer.
3 slides on deep learning to complete the overview.
DL is based on perceptron. Perceptron takes matrix as input.
Matrix can be used for anything pictures, text…
Then it multiplicates this matrix of value with a matrix of weights and send this to an activation functions that output a number
What is great is that it can learn using a mechanism called backpropagation.
Backpropagation takes the output compare it with the expected output and correct the weight matrix accordingly
You can then chain the perceptron also called neuron and build a more or less complex neurol network.
Deep learning network are the ones using a lot of hidden/intermediary layers
Sorry to say that all this is not new. Perceptron are from the sixties, neural network middle eighties.
But what changed is that the computing limitation that stopped the neural network and other algorithm success is not anymore.
We went from data to Big data that can be analyzed using cheap, on demand resources
Business will not ask for it BUT we have the chance to have very smart people in house, that are building smart tools using VBA, Python, open sources framework. We should support them better.
Skillset is wide: Domain knowledge, ML algorithms, data engineering, IT industrialization (5 legs sheep)
BUT the maturing data science platform if boosted by an internal change in training content can foster the raise of Citizen datascientist. For instance Price will train 10% of its staff to data science.
To include business, it is better to start with a simple model we can explain with the objective to assist and not replace.
I believe that disposing of an integrated development platform will speed up innovation and globalisation of innovation. A data hub and a cleaned feature store is one step.
Cloud: Amazon MXNet – Microsoft CNTK – Google TensorFlow. In house sparkML, TensorFlow
I also believe, discovering all the intelligence dispersed everywhere, on personal spreadsheet, that we would gain gretaly finding a way to capture those models to protect, enhance and share them between employees or between countries.