BIQUANTS
TECHNOLOGIES
DESCRIPTION
INTRODUCTION
TO MACHINE LEARNING & AI
PART #1
THE PROGRESSION OF ANALYTICS
DIFFICULTY / TECHNOLOGY VA LUE What happened?
DESCRIPTIVE ANALYTICS
Why did it happen?
DIAGNOSTIC ANALYTICS
PREDICTIVE ANALYTICS What will happen?
PRESCRIPTIVE ANALYTICS How can we make
it happen?
INSIGHTS & FORESIGHT
Statistics, Clustering, Classification, Regressions
OPTIMIZED HUMAN DECISION-MAKING
Neural Network & Deep Learning THE LEARNING & INTELLIGENT ENTERPRISE
Reinforcement Learning, AI
AI VS MACHINE LEARNING VS DEEP LEARNING
*machine learning – ML deep learning - DL
LEARNING DEEP MACHINE LEARNING ARTIFICIAL INTELLIGENCE
A technique which enables machines to mimic human behavior
Subset of AI technique which use statistical methods to enable machines to improve with experience
Subset of MACHINE LEARNING which make the computation of multi-layer neural network feasible
MACHINE LEARNING
A NEW
PROGRAMMING PARADIGM
RULES
VS
CLASSICAL
PROGRAMMING MACHINE
LEARNING
DATA
ANSWERS
DATA ANSWERS
RULES
MACHINE LEARNING =
STATISTICAL + DEEP LEARNING
INPUT FEATURE
EXTRACTION DECISION TREE
NOT CAR CAR
OUTPUT
STATISTICAL LEARNING
DEEP LEARNING
INPUT FEATURE EXTRACTION
+ CLASSIFICATION OUTPUT
CAR
NOT CAR
HAS MORE HIDDEN NEURAL NETWORK LAYERS CAN DO FEATURE ENGINEERING / EXTRACTION
EXAMPLE OF DEEP
REPRESENTATION LEARNED
HIERARCHICAL
REPRESENTATION OF FEATURES
EDGES CONTOURS OBJECT
PARTS OBJECT IDENTITY
HOW DEEP LEARNING WORKS
LAYER
data transformation
LAYER
data transformation
WEIGHTS
WEIGHTS
PREDICTIONS
Y
INPUT X
TRUE TARGETS
Y
LOSS FUNCTION
LOSS SCORE OPTIMIZER
Weight update
A NEURAL NETWORK IS PARAMETRIZED BY ITS WEIGHTS
LOSS FUNCTION MEASURES THE QUALITY OF THE NETWORK’S OUTPUT
THE LOSS SCORE IS USED AS A FEEDBACK SIGNALTO ADJUST THE WEIGHTS
SCHEMATIC DEPICTION OF DEEP REINFORCEMENT LEARNING
REINFORCEMENT LEARNING
Currently, reinforcement learning is mostly a research area and hasn’t yet had significant practical successes beyond games
SUPERVISED LEARNING IS ”TEACH BY EXAMPLE”
REINFORCEMENT LEARNINGS IS “TEACH BY EXPERIENCE”
AGENT is represented by a neural network.
KEY CHALLENGE:
SIMULATION OF ENVIRONMENT AND REWARD
The neural network interacts directly with the environment.
It observes the current STATE of the ENVIRONMENT
and decides which ACTION to take (e.g. move left, right etc.) on basis of the current STATE and the past experiences.
Based on the taken ACTION the AI Agent receives a REWARD The amount of the REWARD determines the quality of the taken ACTION with regards to solving the given problem (e.g. learning how to walk).
The objective of an AGENT is to learn taking ACTIONS in any given circumstances that maximize the accumulated REWARD over time.
GOOGLE’s FRAMEWORK FOR DEEP LEARNING
TensorFlow is a open-source software library for training deep neural networks
It is used for both research and production at Google
Google web search interest for DL
frameworks over time
Companies using TensorFlow:
VIDEO
5D tensors of shape (sample, frame, height, width, channels)
A 3D TIMESERIES DATA TENSOR
REAL WORLD EXAMPLES OF TENSORS
VECTOR DATA
2D tensors of shapes (samples, features) TIMESERIES DATA OR SEQUENCE
3D tensors of shape (samples, timesteps, features) IMAGES
4D tensors of shape (sample, height, width, channels)
THE DEEP-LEARNING SW & HW STACK
SUGGESTED AI INFRASTRUCTURE
PROGRAMMING LANGUAGES
AI HIGH LEVEL FRAMEWORKS
AI LOW LEVEL FRAMEWORKS
AI INFRASTRUCTURE
APACHE
MXNET CUDA,
cuDNN BLAS, Eigen
AMAZON EC2 CPUs
AMAZON EC2
GPUs ENHANCED
NETWORKING
TRAINING DEEP LEARNING NETWORKS
Typical structure of sequential deep learning NN
Recurrent NN structure
DATA PREPROCESSING NETWORK STRUCTURE HYPERPARAMETRS
• encoding
• normalization
• data augmentation
• class imbalance weighting
• number of layers and
connectivity between neurons
• type and order of layers
• network type
• loss function
• learning rate
• minibatch size
• number of epoch
• convolution kernel width
Data augmentation (apply different transformation to initial data)
WORKFLOW OF DEEP LEARNING
1 2
DEFINE THE PROBLEM AND ASSEMBLING A DATASET
What are we trying to predict?
Do we have training data?
What type of problem we are facing?
Binary or multiclass classification, scalar or vector regression, clustering, reinforcement learning
CHECK DATA HYPOTHESES
• outputs can be predicted given your inputs
• available data is sufficiently informative to learn the relationship between inputs and outputs
DATA PREPROCESSING
Vectorization - all inputs & targets must be tensors of floating value
Normalization- convert to 0..1 range Handling missing values
Feature engineering - using your own knowledge about the data
WORKFLOW OF DEEP LEARNING
4 3 CHOOSING A MEASURE
OF SUCCESS
Accuracy, AUC, etc.
Evaluation protocol (hold-out validation, k-fold cross-validation)
Tension between optimization and generalization Prevent a model from learning misleading or
irrelevant patterns found in the training data
• get more training data
• modulate the quantity of information
• reduce network’s size
• weight regularization (make distribution of weight values more regular )
• adding dropout (randomly dropping out number of output features of the layer during training)
Hyperparameters tuning and model regularization
OVERFITTING AND UNDERFITTING
OPTIMIZATIONrefers to the process of adjusting a model to get the best performance possible on the training data
GENERALIZATIONrefers to how well the trained model performs on data it has never seen before.
ROLES AND OPERATION OF DATA SCIENCE TEAM
PART #2
DATA SCIENCE ROLES & INTERACTION
Enables data access & utilization Enables value capture
Builds and supports the infrastructure of data pipe and all associated SW engineering infrastructure tasks
Optimizes & enables data for business and functional value capture and value creation
Analysis and interpretation od complex digital data to extract or discover knowledge and assist decision-making
Helps the business make better decisions through data
Blend of business, analytic and math skills to explore and solve challenges, bridging the data and business communities
DATA OPERATIONALIZATION
DATA MONETIZATION
DATA OPTIMIZATION
AI / DATA SCIENCE: ROLES, SKILLS, QUALIFICATION
DATA
ANALYST MACHINE LEARNING
ENGINEER DATA ENGINEER DATA SCIENTIST PROGRAMMING TOOLS
DATA VISUALIZATION & COMMUNICATION DATA INTUITION
STATISTICS / ALGORITHMS DATA WRANGLING
DATA WAREHOUSE, DATABASE SYSTEMS, ETL MACHINE LEARNING
SOFTWARE ENGINEERING
DISTRIBUTED COMPUTATIONS
MULTIVARIABLE CALCULUS AND LINEAR ALGEBRA
Very important Somewhat important Not that important
TECHNOLOGY & SKILLS MAP
SKILLS / TECHNOLOGY TECHNOLOGY
Programming Tools Python, R, Java/script, MatLab
Data Visualization & Communication GGPLOT, D3.js, Shiny apps, Tableau, Power BI, HTML, JavaScript, REST API, Swagger
Data Intuition R Studio, Python notebooks
Statistics / Algorithms Clustering, decisions trees, Regression, OpenCV
Data Wrangling SQL
Data Warehouse, Database Systems, ETL Amazon RedShift, IBM dash DB, IBM Netezza, Oracle
Machine Learning TensorFlow, Keras, Caffe, Theano, caret R, scikit-learn Python.
Software Engineering Scrum, Version control systems (git), Continuous Integration (jenkins), Docker Distributed Computations Apache MXNET, Amazon AWS, Microsoft Azure, Google cloud, IBM, Apache Hive,
Hadoop, Spark
Multivariable Calculus And Linear Algebra CUDA, cuDNN, BLAS, Eigen
