Tools and Frameworks - Cambridge Wireless · 12. What about improving the tools, knowledge graph inference, generative models? 13.

Tools and Frameworks

Dan Neil, BenevolentAI

3 October 2017

We are

Our mission is to invent and apply AI and

Machine Learning technology to accelerate

scientific discovery that benefits society.

Founded in 2013, $BN valuation, 80+ team of

world-class scientists and technologists

Focused on accelerating Drug Discovery as a first scientific domain

Drug Discovery is broken!

- On average it costs about $2.6bn to develop a new drug

- It requires 12-15 years of R&D from start to market

- 97% of drug programmes fail

- Less than 40% of known diseases are currently treatable

- Novel discoveries are rare and is mainly done by academic research

Eroom’s Law and Moore’s Law

5https://buildingpharmabrands.com/2013/08/16/erooms-law-of-pharma-r-d/

6

1 scientific paper is published every 30 seconds

Individual scientists access a tiny fraction of available

data in their lifetime - much less than the data

generated in 1 day

All biological databases combined are less than 5% of

available data

The BenevolentAI platform extracts facts and reasons

from all relevant databases and literature, structured

and unstructured

Power of the AI technology – Knowledge & Reasoning

7

Unstructured Data

Large Scale NLP,

Machine Learning,

Deep Learning, and

Neural Reasoning

TechnologyStructured Data

Rich proprietary knowledge

Proprietary Knowledge and Inference Models

BenevolentAI finds potential treatments for ALS

8

How do we put together an AI system? Can we leverage the tools out there?

9

Face and Image Recognition

10https://www.clarifai.com/demo

Text Analysis, NLP, Translation, Sentiment, Voice

11https://cloud.google.com/natural-language/

Generalized Recommendation and Prediction

12https://static.bigml.com/pdf/BigML_Classification_and_Regression.pdf?ver=44e4f71

What about improving the tools, knowledge graph inference,

generative models?

13

Every Language Has a Framework

Python Tensorflow, Theano, Keras, CNTK, MXNet, Pytorch, …

Java DeepLearning4J, H2O, Weka, …

R Tensorflow, Keras, H2O, Deepnet, …

Matlab Caffe, MatConvNet, CNTK, …

C++

Javascript

MatConvNet, CNTK, Torch, …

MXNet, deeplearn.js, ConvNetJS, …

The Rising Dominance of Python – Stack Overflow

15https://stackoverflow.blog/2017/09/06/incredible-growth-python/, https://stackoverflow.blog/2017/09/14/python-growing-quickly/

0

2

4

6

8

10

12

Jan-12 Oct-12 Jul-13 Apr-14 Jan-15 Oct-15 Jul-16 Apr-17

Overall Percent of Search Queries, by Language

Python Java JavaScript c# PHP C++

0

5

10

15

20

25

30

Jan-15 May-15 Sep-15 Jan-16 May-16 Sep-16 Jan-17 May-17

Seasonal Traffic, by US and UK Universities

Java Python

0

10,000

20,000

30,000

40,000

50,000

60,000

70,000

80,000

Q2 2015 Q3 2015 Q4 2015 Q1 2016 Q2 2016 Q3 2016 Q4 2016 Q1 2017 Q2 2017 Q3 2017 Q4 2017

Stars on Github

Caffe Theano DL4J MXNet CNTK Pytorch

0

10,000

20,000

30,000

40,000

50,000

60,000

70,000

80,000

Q2 2015 Q3 2015 Q4 2015 Q1 2016 Q2 2016 Q3 2016 Q4 2016 Q1 2017 Q2 2017 Q3 2017 Q4 2017

Stars on Github

Caffe Theano DL4J MXNet Tensorflow CNTK Pytorch

Introduction of Tensorflow

Should I use one framework for everything?

18

Importance Weighting

1. Machine Learning Simplicity

− Many problems can be formulated as

model.fit(train, test), and it’s really fast

to use those. Some can’t.

2. Standard Algorithm Similarity

− ConvNets are cake. Recursive parse nets

aren’t.

3. Deployment Ease

− It runs on your data scientists’ computers.

How do you put it on a phone, or build a

web API around it with less than a second

of latency?

4. Debugging Ease

− Tensorized, delayed execution is

notoriously hard to debug.

09/10/2017 19

Three Optimization Points (that we use)Simplicity

Breadth

SupportDeploy

DebugH2O

TensorFlow

Pytorch

H2O

TensorFlow

Pytorch

Developed and supported by

H2O.ai (with enterprise support)

Simple API; supports Java and

Python

Plugs into existing Big Data

infrastructure (Hadoop, Hive, Pig)

Rapid, simple, scalable machine

learning for our simple predictions

Dedicated development, feature

rich

Cross-platform – including mobile

Performs heavy-lifting for NLP,

knowledge graph inference, and

factorization

Extremely fast for prototyping

research algorithms

Extremely flexible

Rapid prototyping for emerging

research topics: e.g., GANs,

complex seq2seq models

Fast, Simple Predictions: H2O Example (1/2)

21

import h2oh2o.init(ip="192.168.1.10", port=54321)-------------------------- ------------------------------------H2O cluster uptime: 2 minutes 1 seconds 966 milliseconds H2O cluster version: 0.1.27.1064H2O cluster name: H2O_96762H2O cluster total nodes: 4H2O cluster total memory: 38.34 GBH2O cluster total cores: 16H2O cluster allowed cores: 80H2O cluster healthy: True-------------------------- ------------------------------------path_train = ["hdfs://192.168.1.10/user/data/data_train.csv"]path_test = ["hdfs://192.168.1.10/user/data/data_test.csv"]train = h2o.import_file(path=path_train)test = h2o.import_file(path=path_test)#Parse Progress: [##################################################] 100%#Imported [hdfs://192.168.1.10/user/data/data_train.csv'] into cluster with 60000 rows and 500 cols#Parse Progress: [##################################################] 100%#Imported ['hdfs://192.168.1.10/user/data/data_test.csv'] into cluster with 10000 rows and 500 cols

http://docs.h2o.ai/h2o/latest-stable/h2o-py/docs/intro.html

Fast, Simple Predictions: H2O Example (2/2)

22

from h2o.estimators.gbm import H2OGradientBoostingEstimator

model = H2OGradientBoostingEstimator(ntrees=100, max_depth=10)

model.train(x=list(set(train.names)-{"label"}), y="label", training_frame=train, validation_frame=test)#gbm Model Build Progress: [##################################################] 100%

preds = model.predict(test)model.model_performance(test)

http://docs.h2o.ai/h2o/latest-stable/h2o-py/docs/intro.html

Debug anywhere: Pytorch

23

import torch

class MyLoss(torch.autograd.Function):def forward(self, x):return x.view(-1).sum(0)def backward(self, x):import pdbpdb.set_trace()return x

v = torch.autograd.Variable(torch.randn(5, 5), requires_grad=True)loss = MyLoss()(v)loss.backward()

https://discuss.pytorch.org/t/is-there-a-way-to-debug-the-backward-method-of-function-class/2246

For everything else, there’s Tensorflow

24

0

10,000

20,000

30,000

40,000

50,000

60,000

70,000

80,000

Q2 2015 Q3 2015 Q4 2015 Q1 2016 Q2 2016 Q3 2016 Q4 2016 Q1 2017 Q2 2017 Q3 2017 Q4 2017

Caffe Theano DL4J MXNet Tensorflow CNTK Pytorch

A Note About Deployments

1. Remember your deployment during

development

− Serializable? Cross-platform? On-device?

2. Tensorflow eases everything:

− Tensorflow Serving

− Universal protobufs

3. JVM is dead, long live Docker

− Package your code into Docker images

− Docker works across your scientists’

computers, your AWS instances, your local

compute

− Consider Kubernetes-style compute

management for heavy-AI companies with

local data requirements

25

1. Python is now the de facto language

of machine learning

− Deployment difficulties are offset by

Docker

− Runtime challenges are fixed by efficient

back ends

2. Consider the aim:

− Research: Pytorch

− Standard Big Data ML: H2O

− Long-term: Tensorflow

3. All still in flux

− Correct abstraction not yet determined

− Proper software engineering for ML under

development

Conclusion