Sanjib Basak works as the director of data science at Digital River and has over 15 years of experience in analytics. He is fascinated by machine learning and deep learning, and has been experimenting with TensorFlow for 6-7 months. In his presentation, he introduces TensorFlow and builds basic models using it, including language models and RNNs. He discusses advantages of TensorFlow over Python/NumPy and takes questions at the end.
2. Working as a director of data science at
Digital River
Working in analytics field for 15 plus
years
Built analytical applications in Retail and
Healthcare
Fascinated by Machine Learning and
Deep Learning
Playing with Tensor Flow for 6-7 months
Host of another meet up Twin cities Big
Data Analytics
3. Introduction to Tensor Flow
A basic model with Tensor Flow
Language Model
RNN with Tensor Flow
Advantages of Tenor Flow over Python Numpy
Question and Answer
4. TensorFlow is an open source software library for numerical computation using
data flow graphs.
Open sourced by Google in Nov,2015
The flexible architecture allows to deploy computation to one or more CPUs or
GPUs in a desktop, server, or mobile device with a single API.
One of the most popular project in open source community with 30,000+ Stars
and 13,000+ Forks in Git Hub
5. Tensors provide a natural and concise mathematical framework for formulating
and solving problems in areas of physics such as elasticity, fluid mechanics, and
general relativity Wolfram Mathworld
Tensors are n dimensional arrays
Scalar is 0 D tensor
Vector is 1 D tensor
Matrix is 2 D tensor
1.5
1.1
1.3
1.5, 1.6,1.7
1.1,1.2,1.3
1.3,1.6,1.7
Vector Matrix
1.5
Scalar
Tensor
6. Node: Represents Operations like
Addition, Multiplications etc.
Edges: Carriers of the operations
Kernels: Implementation of the operations
in device - CPU , GPU
Session: A session is created when client
program establish communication with TF
run time code
Tensor Flow data structure:
Rank : Rank as defined above
Shape : Number of rows and columns
Type Int,Float32
7. A linear regression model using ipython notebook
8. Language Models computes the probability of occurrences of words in a sequence
Widely used in speech and translation system
Model chooses best word ordering from various possibilities
The world is small,small is the world
9. Bi-Gram and Tri-Gram Model
Drawbacks:
In some cases window of past consecutive
words may not be sufficient
Fails to capture context
Jane and John met in the market.
They started fighting with each other.
Who started fighting?
Memory intensive
10. In 2001, Bengio et al. introduces the
first large-scale deep learning for
natural language processing model
This model capture context via
learning a distributed representation
of words
A Feed Forward Neural Network Model
11. word vectors
Loss Function: Cross Entropy
Measurement of the Model
In 2010 Mokolov et. all presented multi layered Recurrent Neural Network (in NLP)
12. Load Data
Add
Placeholder
Create Feed
Dictionary
Validate
Model
Test Model
Add Embedding
(embed_size = 50)
Create and
update Hidden
Layer
(Hidden layer of
size =100)
Create and
update
Projection
Layer
Calculate
Loss
Optimizer applied
to all trainable
variables
S
Optimization
achieved
E
Repeatuntil required optimization
Train Model
13. Problem Definition
Test 1 Trained on 20 sections of WSJ data of Penn Tree Bank
Test 2 Trained on Harry Potter and Deathly Hallows
16. On Deep Learning/NLP
http://cs.stanford.edu/people/karpathy/
http://www.wildml.com/
Mikolovs thesis:
http://www.fit.vutbr.cz/~imikolov/rnnlm/thesis.pdf
Richard Socherrs Thesis
http://nlp.stanford.edu/~socherr/thesis.pdf
On TensorFlow
https://www.tensorflow.org/
Comparison between TensorFlow and other frameworks
http://deeplearning4j.org/compare-dl4j-torch7-pylearn.html