API World + AI DevWorld

IBM Watson Deep Learning as a service (DLaaS), a cloud-based deep learning platform that provides a deep learning software stack with leading edge GPU hardware for cloud environments in a secure, scalable and fault-tolerant manner. It supports a wide range of deep-learning frameworks such as Tensorflow, PyTorch, Caffe, Torch, Theano, and MXNet etc. These frameworks reduce the effort and skill set required to design, train, and use deep learning models.  DLaaS is running in 14 production data centers in multiple regions such as Tokyo, South Korea, Sydney, Frankfurt, London, Washington DC and Dallas providing Deep Learning service capabilities to customers worldwide with 24X7 availability. The DLaaS DevOps process is fully automated to provide updates to all data centers in a timely and zero down-time fashion. DLaaS is being used by many customers as well as Watson services such as Watson Assistant, Visual Recognition, Natural Language Classifier and Speech Recognition etc.

This talk will introduce you to DLaaS and Watson services built based on DLaaS. It will introduce how you can leverage DLaaS to build your deep learning model in a timely manner and leverage Watson technologies to build your AI applications.

Transfer learning enables using pretrained deep neural networks trained on various large datasets and adapt them for various tasks. Fine-tuning such pre-trained models in computer vision has been a far more common practice than training from scratch. In NLP, however, due to the lack of models pretrained on large corpus, the most common transfer learning technique had been fine-tuning pretrained word embeddings. These embeddings are used as the first layer of the model on the new dataset, and still require training from scratch with large amounts of labeled data to obtain good performance. Finally in 2018, several pretrained language models (ULMFiT, OpenAI GPT and BERT) emerged.These models are trained on very large corpus, and enable robust transfer learning for fine-tuning many NLP tasks with little labeled data. In this talk we'll learn the architecture of these pretrained language models. In particular, we'll share how different transfer learning techniques have been used with BERT to solve various downstream tasks in the NLP community.

The amount of data produced in every minute makes it challenging to store, manage, utilize, and analyze it faster. Companies like Intel are building hardware accelerators to boost software application performance significantly. However, not many big data applications are really known them or optimizing to use it. We created an Opensource DARS (Data Analytics Reference Stack) (https://clearlinux.org/stacks/data-analytics-stack-v1) based on open source Clear Linux, and well utilizes the hardware accelerated libraries in the big data service stack.
The DARS comprises of Apache Spark, Apache Hadoop, OpenJDK and utilizes both MKL, OpenBLAS math libraries all built on ClearLinux and complete stack includes ClearLinux OS along with containerized environment images such as Docker images. This end-end optimized stack delivers up to 8x performance gain in machine learning workloads. We will also cover how to use the optimized stack, replicate on your production environments and present the performance results.

Complex machine learning models are frequently termed “black boxes”. While these models deliver highly accurate predictions, they frequently provide no insight into the factors driving the predictions. In turn, this can make it problematic to ensure that the model’s predictions are actionable. For instance, in churn prediction, explaining why the customer is predicted to churn is as important as the prediction itself, as it enables action to be taken to address highlighted issues and prevent the churn.In recent years, as operationalized ML solutions become ever more complex and prevalent in every facet of life, the interpretability of models has become a critical consideration. In the last couple of years, an increasing number of papers and open source software libraries have appeared to help data scientists explain their predictions. In this presentation, we discuss recent approaches to interpreting the results from complex machine learning models, including the well-known LIME package. We present best practices for their use and deployment and provide cautionary examples and caveats associated with leveraging these techniques.

Recent years, inspired by the huge success of convolutional networks in Computer Vision, Graph Convolutional Networks has emerged and attracted great attention for its potentials in graph-based data world, such as social networks, recommendation, biological graphs and traffic prediction. In this talk, we present our innovation on fraud prediction of using Graph Convolutional Networks (GCNs) in Walmart eCommerce system where it predicts risk scores for entities (e.g. customer, device) in real time on large-scale transaction-based graphs. We demystify how it smartly leverages graph topology via adjacency matrix and learn the weight matrices to nonlinearly propagate information across multi-hop graphs, especially on graphs condensed of fraudulent transactions. We also describe how the semi-supervised classification GCNs is trained with a weighted softmax loss function on partially labeled graph snapshots. By comparing with models that do not rely on graph structures, we show the noticeable performance lift of GCNs in various aspects. To demonstrate how well it works, we evaluate the prediction results through real-world graph examples at WalmartLabs. At the end, we show the significant business/engineering value our entity scores can bring to the company and point out multiple directions in which the methodologies can be generalized and further improved.

The latest AI advances have the potential to massively improve our health and well-being. However, most of the work is yet to be done. In this talk, we will explore the most important opportunities for AI in healthcare. For example, we will explore how AI can diagnose major life-threatening conditions even before those conditions emerge. We will talk about AI ability to recommend dramatically more effective and less harmful treatment plans based on AI understanding of patient's medical history and current conditions. Finally, we will talk about AI role in making our healthcare system effective and affordable for everyone. For each area, we will discuss the latest progress made as well as the challenges remaining.

Low latency applications - such as cloud gaming, immersive media streaming, Industrial IOT, Video surveillance and data reduction applications such as Video CDN are driving edge computing, more specifically distributed edge computing.  Geo-distributed computing needs geo-distributed analytics.  Bootstrapping collection, big data lakes and AI frameworks for a single site takes weeks and think of amount of time it requires to bring up big data AI framework in multiple edge locations. In this session, we will talk about AI framework, which we developed in Linux Foundation as open source. This framework can be used to many AI applications together, developed in any language and used any ML/DL frameworks. This Microservice based AI frame architecture can be deployed anywhere where Kubernetes is used. In this session, we will also talk about the configuration flexibility to deploy in heterogeneous environments. Deploying and configuration framework in tens of thousands of edge locations require automation. We will also talk about the work done on deployer intent and auto deployment of AI framework components on static edge locations as well as dynamic edges.  If you are edge provider or deploying applications in edges, you will walk away with a scalable and practical AI framework that can be used to deploy your analytics applications.

You've developed and trained your ML model, and it performs beautifully in your development environment -- but what happens when you move that into production, and it suddenly has to scale massively varying elastic workloads, compete with other models for memory and processing resources, or mesh with models deployed in other languages and frameworks?

It isn't enough to simply fire up a machine instance, write a Flask wrapper, and call it a day: properly productionizing a model requires a deep understanding of container management, load balancing, CI/CD, dynamic resource allocation, and more. In this talk, we'll look at what your team does and does not need to build in order to move from weeks of deployment time to mere minutes, while preserving elasticity, low-latency, and flexibility.

This talk will provide deep insights in comparing CPU vs. GPU vs. Cloud based solutions in real-world scenarios. We will also explore how Open Source platforms can facilitate edge computing applications, including Java and Python frameworks' capabilities on ARM and X86 architectures.

As we continually search for bigger, more complicated problems, the demands we place on the computing power continue to grow rapidly. So where next - how can we continue to build models and algorithms that scale as our data does or enable us to tackle even more intricate problems?
Edge Processing not only optimizes large-scale data mining and aggregation, (by moving the data processing portion of an application to a single unit, known as the Gateway) but also facilitates Machine Learning in modern IoT applications. Most IoT gateways however, are only capable of running basic data aggregation functions due to their limited parallel processing capabilities of their CPUs.
For IoT devices to truly simulate human intelligence, not only do they need to implement neural networks, but they also need to take advantage of their distributed resources collectively.
In comparison with traditional CPU systems a GPU is capable of far higher peak performance for parallel data streams, which is a requirement of artificial neural networks. GPUs have evolved from a fixed pipeline graphics processing hardware into powerful programmable co-processing units capable of performing general purpose computing.

In this talk, Dr Mo Haghighi will explore the challenges and opportunities in edge computing, as well as discussing various platforms and their applicability for various real-world scenarios. This talk will explore JVM capabilities on ARM and X86 architectures, as well as large scale data mining on embedded GPUs. We will also present a benchmarking demo comparing CPUs vs. GPUs for processing time-series data, using a combination of Java and Python libraries.

IoT can be divided mainly into following four stages:
1. Sensors, Actuators, Controllers (limited resources, limited intelligence)
2. Gateway, Concentrators, Routers (data aggregation, measurement)
3. Edge (analytics, preprocessing)
4. Cloud, Servers (analytics, management)

With the advent of TF lite on microcontrollers, deep learning can be applied at all four stages of IoT.
This presentation will look at evolution of performing functions at every stage of IoT. The presentation will describe how function implementations changed from conventional to rule based to conventional machine learning and finally to deep learning using TF lite.
Conventional machine learning relies on hand crafted feature sets limited to human view and uses large data sets and are not very accurate. Deep learning on the other hand automatically generates engineering feature sets that are beyond human view, provides better training and more accurate prediction that approaches human level accuracy.

Traditionally, AI is considered as Batch processing. But Realtime AI needs are growing. In this session, I will cover Realtime AI with Apache Flink.
Companies like Amazon, Facebook etc are doing near real time recommendations based on machine learning models. To enable these AI models, there is a need for a framework like Apache Flink.
It is a beginner session for Apache Flink and Realtime AI.

The demands and complexity of backup storage have increased exponentially over recent years. The storage and backup needs of Oracle Cloud Infrastructure applications such as IoT, microservices, analytics, and private clouds require the data protection of modern workloads such as NoSQL DBs, VMs, containers, IoT data, and so on. These needs are demanding and diverse, requiring something beyond traditional backup solutions. In this session learn how to use machine learning to analyze workloads at runtime to identify their data protection characteristics, such as data deduplication and backup retrieval usage patterns, and then recommend appropriate backup targets such as in Oracle Cloud Infrastructure’s object store and backup appliances.

This talk will focus on implementing AI-based audio and voice strategies into your business. The session will include a live demo and key wins developers can achieve from activating an audio and voice strategy.

The growing adoption of mobility, IoT and cloud significantly increases the impact of network infrastructure on enterprise businesses, and creates new challenges for traditional human-driven network operation. In this talk, we will share empirical experiences of using ML/DL and AI to build the autonomous enterprise network operation solution that provides visibility, troubleshooting, reporting and maintenance of an enterprise network.

We first introduce the detailed architecture, which includes four main components: measurement, detection, orchestration and automated/suggested actions. Then we dive deeper into the example of anomaly detection where we use unsupervised confident deep multivariate models to automatically baseline and detect the user-impacting connectivity issues inside enterprise network together with other contextual information from scope and root cause analysis to make the AI results interpretable and actionable for end users. In addition, we share some results of the anomaly detection models in practice across thousands of sites. We also describe an anomaly detection example that we use deep learning models to automatically baseline and detect the user-impacting connectivity issues inside enterprise network, together with other contextual information from scope and root cause analysis to make the AI results interpretable and actionable for end users.

At last, the speakers will discuss some common challenges and best practices people encounter while developing ML/AI-driven enterprise applications, including:
1. Building the cloud-native analytics infrastructure for agility, scalability and visibility
2. Leveraging visualization, model interpretation, and feedback to keep humans in the loop while developing machine intelligence;
3. Building human trust in the step-by-step process of automation, augmentation and autonomy
4. Accelerating knowledge learning and sharing across enterprises in a SaaS environment without compromising individual privacy.

Knowledge Graphs have gained significant visibility with advancements in artificial intelligence and the adoption of graph technologies in the Enterprise. By combining artificial intelligence, graph databases, and dynamic visualizations, we will discuss deploying graph based AI applications as a means to help predict future events across numerous types of industries.

Knowledge creation via AI and Graphs stems from the capability to combine the probability space (i.e. statistical inference on a user’s data) with a knowledge base of comprehensive industry terminology systems. AI using Graphs are remarkable not just because of the possibilities they engender, but also because of their practicality. The confluence of knowledge via machine learning, visual querying, and graph databases not only displays links between objects, but also quantifies the probability of their occurrence. We believe this approach will be transformative across numerous business verticals.

During the presentation we will describe the Graph based AI concepts and techniques for developing Enterprise Knowledge Graph solutions.

AI shows great promise, but it has not been adopted rapidly because the barriers to entry are too high. In the same way that the Cloud became ubiquitous, AI will follow the same track.

Now, Visual AI As A Service - and all AI as services - are poised to provide the value the Enterprises seek. Right now, VIsual AI is available as a demo on the iPhone - the Chooch IC2 app - and now developers can train their instance of Chooch to recognize any object - faces, cells, art, sports actions - and then use an API to apply AI to their visual data.

The GPU infrastructure and the machine learning models, to name just two technology features, are very difficult to create and maintain. By using AI As A Service, the enterprise can leverage the value of economies of scale and finally reap the rewards of AI.

Ethical & Collaborative Networks: 5G & AI will enable mobile intelligent systems with volumetric data sets and low-latency closer to the users and the edge of the network. Enabling new use-cases, paradigms and possibilities. It amplifies questions around privacy, ethics, security and optimal network, data-set and user experience design. This talk will talk about how we may start to approach ethical design using examples from smart cities, IOT factories and location based entertainment experiences.