5. Why Now?
Second –Perfectly timed arrival of Data Infra, Compute, Network and ML
Innovation
Improved
Data Collection Infrastructure
Improved
Computing Power & Network
Advancement in
ML Frameworks
Hadoop w/Apache Spark Faster GPUs, CPUs & Network TensorFlow, PyTorch, Caffe, Theano
6. Beyond Hype…Solving Biz Needs AcrossAll IndustryVerticals..
Retail
Security and
Defense
Media and
Entertainment
HealthcareFinanceAI Task
Business
Question
Is "it" present
or not?
What type of
thing is "it"?
To what
extent is "it"
present?
What is the
interpretation?
What is the
likely
outcome?
What will
satisfy the
objective?
Detection
Classification
Segmentation
Natural
Language
Processing
Prediction
Recommenda
tions
Identify
Access
Anomalies
Fraud
detection
Sentiment
Analysis
Chatbot
Advisors
Credit
Profiling
Algorithmic
Trading
Indication of
Anomaly in
Scan
Diagnostics –
Tumor?
Condition
Analysis
Expert
Diagnosis
from Notes
Length of Stay
Forecasting
Treatment
Recommenda
tions
Content
Based Search
Content
Labeling
Improved
Product
Placement
Video
Captioning
Targeted
Content
Generation
Content
Recommenda
tions
Identify
Security
Breaches
Facial
Recognition
Crowd
Analytics
Real Time
Language
Translation
Equipment
Health
Assessment
Risk
Management
Events in
Store
Surveillance
Returning vs
New
Shoppers
Segment by
Customers
Actions
In Store
Personal
Assistants
Customer
Churn and
Retention
"Magic Mirror"
Manufacturing
Detect
Manufacturing
Flaws
Robots to
Track Objects
Sort
Components
by Quality
Assembly Build
Instruction
Translation
Proactive
Machine
Maintenance
Assembly
Process
Improvements
7. High Level View of Workflow When Using AI/ML In Production
Data
Infrastructure
Prepare Data Train a Model
Evaluate
the Model
Deploy, Inference
& Improve
Data is at the heart.
More data trumps
better algorithms
Data and business
questions determine
ML algorithm(s)
Data can come from
anywhere and is not
usually in a state
where it is ready to
use for training ML
models
Build a model and
feed the model with
prepared training data
so that it can learn to
make inferences
Test the trained model
performance and
accuracy by analyzing
inference feedback
Deploy Model for
inferencing. Evaluate and
Improve accuracy by either
selecting different
algorithms and/or retraining
model with more data
Big Data
ML/DL Framework
Training
Inference
Digital disruption and the resulting explosion in data, applications, and devices are disrupting traditional business models and making it harder to compete successfully than ever.
30M new devices connected every week
As just one security indication of the security impact of all of those new devices, 4.2Billion web filtering blocks are created per day
Each of those devices will be creating large quantities of data. As a matter of fact, 5TB+ of data will be created per person by 2020
That’s clearly just the tip of the iceberg, 277X data will be created by IoE devices vs end-user
And 40% of all data will be created from sensors in IoT devices by 2020
Insights held within these massive volumes of data have the potential to deliver a powerful competitive advantage.
The disruption caused by unprecedented growth in devices, mobile apps, data, use of cloud, and security risks are forcing organizations to redefine business processes to compete more effectively, which, in turn, is driving the need for unprecedented agility in their IT architectures and governance. To enable this digital transformation, IT must create/manage an infrastructure that extends from the enterprise data center − to the cloud− to the edge to deliver new levels of speed and efficiency and lay the foundation for ongoing innovation.
The business value of AI/ML isn’t a debate
Why is ML emerging now?
There are a couple key reasons. One, it’s getting really cheap to store data. From cheaper hardware to the spread of open source technology like Hadoop to the growth of the cloud, it is now very cheap to store data, resulting in the rise of Big Data. Much of this data is excellent training material for learners and is being packaged for easy use, such at ImageNet and data.gov.
But, training a ML is computationally expensive. Fortunately, as seen in our second chart, better CPUs, GPUs, and cloud computing have massively reduced the price of compute.
Additionally, during this time, there have been mathematical advances that have made training NNs easier and more effective and a public-private ecosystem – among Universities, companies, and government – that has encouraged rapid progress in the field.
The combination of these trends creates an environment that is prime for the growth of machine learning.
As with many cloud and digital transformation initiatives, the line of business is driving modern AI. A successful AI application starts with a compelling business question. In order to be successful selling AI, you need to start out side of IT and focus on data and business professionals. Industries are creating specific roles that are data driven, a good example is the Chief Medical Information Officer in healthcare.
Moving down and across this table:
“Is “it” present or not”, is asked in Finance to identify abnormal access to financial data
“What type of thing is “it”?”, is asked in healthcare to identify if a tumor is cancerous or benign
“To what extent is “it” present?”, is asked in Media and Entertainment to identify if a placed product is impactful
“What is the interpretation?”, is asked in Defense to translate a local dialect
“What is the likely outcome?” is asked in Retail to identify what customers will buy or not
“What will satisfy the objective?” is asked in Manufacturing to identify how to make robots work quicker and build higher quality products
Simple Learning: K-Nearest Neighbor, Ensemble Techniques, Support Vector Machines, Random Forest, Decision Trees, Linear Regression, Bayesian Techniques, SVMs
Deep Learning: Reinforcement Learning, Deep Neural Networks, Recurrent Neural Nets
Driving factor for Deep Learning: Big Data (large-scale training data, Algorithmic innovation, Compute infra)
We believe there are a few key challenges when it comes to AI/ML. The first challenge is traditional data center technology is not designed to handle the data volume, velocity, and variability of AI at production scale. IT teams are struggling to keep up with data scientists needs while data scientists are struggling to operationalize machine learning, new data sources and rapidly shifting software stacks requires. Infrastructure to constantly adapt to these new data-shaped workloads. The massive amount of data required and the ingestion speed also require bandwidth to move the amount of data consumed by AI through all stages of the data lifecycle.
“Data is the new Application and IT needs new tools to deliver AI at scale.”
The second challenge is many IT departments regard data science software as unstable, unsupported and undebuggable. They are not wrong. The AI ecosystem is still in the early stages resulting in fragmented, unfamiliar, and rapidly evolving machine learning software stacks that increase complexity and risk. There is also a skill shortage. There are not enough data scientists and data engineers.
“We’ve taken on the challenge to demystify AI/ML with proven, full-stack solutions developed with industry leaders.”
The third challenge is the data center follows the data, IT must integrate new accelerated computing technology across an increasingly distributed landscape. This increasing need for new platforms and data to be analyzed at the source, in the data center and at the edge, can spawn islands of standalone AI/ML servers. IT leaders are not being permitted to add administrative resources as these projects scale to thousands of nodes. Roughly 50% of the DC IT budget goes toward people and software to manage existing infrastructure. A wide variety of traditional servers, software systems, storage have been utilized, and implemented in silos—most system management functions are complicated—designed to try to integrate solutions that were never designed to work together. Hence, they need management solutions that use a common operating model and enable the same resources to be able to manage a distributed, scale-out infrastructure.
“The data center is where the data is. As we take you through this briefing you will see how Cisco pairs a full portfolio of AI/ML computing solutions with the simplicity and reach of cloud-managed infrastructure.”
Let’s provide some additional detail on what we are announcing. As I mentioned, in addition to the current portfolio of UCS and Hyperflex systems with GPU support, we are introducing the new UCS C480 ML for model training. Again, I have a couple of detailed slides on this system.
We are working with both software partners and channel partners to help IT simplify the deployment of. ML infrastructure. We are collaborating with software vendors to provide proven solutions for AI/ML projects while collaborating with channel partners – Presido, WWT, E+ and others, to build expertise so they can help customer close the skills gap by delivering services to accelerate ML projects.
Meet the new UCS C480 ML AI server. Intuitive to IT but packs the punch of a supercomputer. Highly integrated but flexible so that it provides Enterprise features like choice of OS and different storage configurations to provide data performance and protection.
If we want to map UCS to a Deep Learning project we could start with hadoop solutions built on a C240 cluster to collect and look at data to see if we have the data to support answering the compelling business question. Once we know what we want to do, instead on making a huge investment in time and money, we could deploy a small HX cluster with all of the GPUs and tools required to start development and testing of AI applications. Once we start build more complex AI applications in production we can offload training to an AI supercomputer like the C480 ML so that we can keep up with business demands. If we want to deploy trained AI applications for inference of consumption outside of the datacenter and at a remote location like a retail store of manufacturing floor, we can use a small but efficient all in one hyperconverged solution equipped with inference optimized GPUs like the HX220c M5.
Now, let’s take a look at the products that we are already shipping and some that are coming..
Stat: $1.2 trillion - Forrester - https://www.forrester.com/report/Predictions+2017+Artificial+Intelligence+Will+Drive+The+Insights+Revolution/-/E-RES133325
Stat: 8 out of 10 - Oracle - https://www.oracle.com/webfolder/s/delivery_production/docs/FY16h1/doc35/CXResearchVirtualExperiences.pdf
Stat – 40% - Accenture https://www.accenture.com/us-en/insight-artificial-intelligence-future-growth
We believe there are a few key challenges when it comes to AI/ML. The first challenge is traditional data center technology is not designed to handle the data volume, velocity, and variability of AI at production scale. IT teams are struggling to keep up with data scientists needs while data scientists are struggling to operationalize machine learning, new data sources and rapidly shifting software stacks requires. Infrastructure to constantly adapt to these new data-shaped workloads. The massive amount of data required and the ingestion speed also require bandwidth to move the amount of data consumed by AI through all stages of the data lifecycle.
“Data is the new Application and IT needs new tools to deliver AI at scale.”
The second challenge is many IT departments regard data science software as unstable, unsupported and undebuggable. They are not wrong. The AI ecosystem is still in the early stages resulting in fragmented, unfamiliar, and rapidly evolving machine learning software stacks that increase complexity and risk. There is also a skill shortage. There are not enough data scientists and data engineers.
“We’ve taken on the challenge to demystify AI/ML with proven, full-stack solutions developed with industry leaders.”
The third challenge is the data center follows the data, IT must integrate new accelerated computing technology across an increasingly distributed landscape. This increasing need for new platforms and data to be analyzed at the source, in the data center and at the edge, can spawn islands of standalone AI/ML servers. IT leaders are not being permitted to add administrative resources as these projects scale to thousands of nodes. Roughly 50% of the DC IT budget goes toward people and software to manage existing infrastructure. A wide variety of traditional servers, software systems, storage have been utilized, and implemented in silos—most system management functions are complicated—designed to try to integrate solutions that were never designed to work together. Hence, they need management solutions that use a common operating model and enable the same resources to be able to manage a distributed, scale-out infrastructure.
“The data center is where the data is. As we take you through this briefing you will see how Cisco pairs a full portfolio of AI/ML computing solutions with the simplicity and reach of cloud-managed infrastructure.”
Why Cisco?
To address this emerging opportunity and associated challenges, we are focusing on the full life cycle of an AI/ML project. First, we have experience in big data has helped many customers integrate changing data sources as part of a dynamic data pipeline. We have been helping customers extend their big data environment in AI/ML by purchasing UCS servers, populate with GPU’s, and connect to their data lake.
We have consistently taken a no compromise approach to computing by developing highly available, richly configured systems, all based on a unified architecture, that can be seamlessly managed with existing infrastructure. Our existing UCS and HyperFlex systems have robust GPU support that can address all stages of an AI/ML project, so organizations can capitalize on the adaptability and programmability of the Cisco Unified Computing System and power AI workloads at scale. In a couple weeks, we will expand the UCS portfolio with a new optimized machine learning computing system targeted specifically on the model training stage. Will cover that system in more detail in a moment
Second, we are bringing our experience from big data and the work we have done there with technology partners to bridge the gap to AI/ML. Using this proven approach, we are going to help IT demystify AI/ML in the data center with proven AI/ML computing solutions that combine a broad set of technologies and applications to help extract more intelligence out of all stages of the data life cycle while ensuring a faster, more reliable, and predictable deployment.
Third, we can help prevent architectural silos, extend administrative expertise and simplify operations with a cloud managed computing system. With Cisco Intersight, we can make it easy to adopt new technologies anywhere, eliminating islands of standalone AI/ML servers, regardless of where they are located, in the data center, multi-site remote/branch, or the edge.
One final point, in order to move the amount of data consumed by AI from its collection point to the models and back to where inferencing is taking place, who else than Cisco, the worldwide leader in networking technology, can provide the bandwidth and security required.
With big data insight is gained and recommendations are provided through the analysis of the data. For example, Amazon recommends items to purchase based on your previous purchases and what other people purchased.
With AI, it’s about machines making decisions based on the data. The data as it moves from one stage to the next, the data life cycle, requires work to prepare it for the next stage.
So, in addition to validating popular machine learning stacks such as Kubeflow from Google enabling the creation of symmetric development and execution environments between on-premise and Google Cloud. We are also creating solutions that extend and integrate Hadoop with machine learning, making data located in data lakes accessible for the next stage of analytics.
We feel it’s crucial for IT organizations, in order to have a faster path to success, to work with a vendor that has experience and track record of collaborating with a broad ecosystem and can bringing together the various components of AI in a more holistic solution.
Problem: UCS Big Data Customers want deep learning to refine data into value
Description: Hadoop loosely coupled with GPU nodes for deep learning with Jupyer notebook
Solution: CVD
Problem: For technical savvy data scientists to leverage Google DL capabilities on premise
Description: Composable, portable, scalable ML Stack Enabling rapid development AND deployment
Solution: CVD for UCS and HX
Problem: How to better integrate Hadoop and DL?
Description: Docker based application support
YARN scheduling CPU & GPU
HDFS with erasure encoding supporting tiered storage (1.7 copies instead of 3)
Hot swap drives for HDFS
Let’s reflect back to our challenges side where data volume, velocity, and variability is changing application behavior and placing new demands on infrastructure, unfamiliar software, and management of distributed and new technologies were outlined as key challenges.
We have experience in big data has helped many customers integrate changing data sources as part of a dynamic data pipeline. We have been helping customers extend their big data environment in AI/ML by purchasing UCS servers, populate with GPU’s, and connect to their data lake.
We have consistently taken a no compromise approach to computing by developing highly available, richly configured systems, all based on a unified architecture, that can be seamlessly managed with existing infrastructure. Our existing UCS and HyperFlex systems have robust GPU support that can address all stages of an AI/ML project, so organizations can capitalize on the adaptability and programmability of the Cisco Unified Computing System and power AI workloads at scale. In a couple weeks, we will expand the UCS portfolio with a new optimized machine learning computing system targeted specifically on the model training stage. Will cover that system in more detail in a moment
Second, we are bringing our experience from big data and the work we have done there with technology partners to bridge the gap to AI/ML. Using this proven approach, we are going to help IT demystify AI/ML in the data center with proven AI/ML computing solutions that combine a broad set of technologies and applications to help extract more intelligence out of all stages of the data life cycle while ensuring a faster, more reliable, and predictable deployment.
Third, we can help prevent architectural silos, extend administrative expertise and simplify operations with a cloud managed computing system. With Cisco Intersight, we can make it easy to adopt new technologies anywhere, eliminating islands of standalone AI/ML servers, regardless of where they are located, in the data center, multi-site remote/branch, or the edge.
Lastly, in order to move the amount of data consumed by AI from its collection point to the models and back to where inferencing is taking place, who else than Cisco, the worldwide leader in networking technology, can provide the bandwidth and security required.