Schedule

Developing theoretically principled tools to guide the use of production-scale neural networks is an important practical challenge. Michael Mahoney explores recent work from scientific computing and statistical mechanics to develop such tools, covering basic ideas and their use for analyzing production-scale neural networks in computer vision, natural language processing, and related tasks.

In many countries, policy decisions are disconnected from data, and very few avenues exist to understand deeper demographic and socioeconomic insights. Ganes Kesari and Soumya Ranjan explain how satellite imagery can be a powerful aid when viewed through the lens of deep learning. When combined with conventional data, it can help answer important questions and show inconsistencies in survey data.

Machine learning (ML) engineering differs fundamentally from traditional software engineering in the level of uncertainty and unpredictability of an idea until fully verified in production. Join Zhe Zhang to explore the deciding factor in ML-based products (e.g., recommendation, ranking)—the speed of the trial-and-error loop.

Ilya Feige explores AI safety concerns—explainability, fairness, and robustness—relevant for machine learning (ML) models in use today. With concepts and examples, he demonstrates tools developed at Faculty to ensure black box algorithms make interpretable decisions, do not discriminate unfairly, and are robust to perturbed data.

Weifeng Zhong explores a novel method to learn structural changes embedded in unstructured texts based on the Policy Change Index (PCI) framework developed by economists Julian Chan and Weifeng Zhong. He explains how an off-the-shelf application of deep learning—with an important twist—can help you detect structural break points in time series text data.

Evaluating machine learning models is surprisingly hard, but it gets even harder because these systems interact in very subtle ways. Ted Dunning breaks the problem into operational and functional concerns and shows you how each can be done without unnecessary pain and suffering. You'll also get to see some exciting visualization techniques to help make the differences strikingly apparent.

Many natural language processing (NLP) tasks require each word in the input text to be mapped to a vector of real numbers. Julien Simon explores word vectors, why they’re so important, and which are the most popular algorithms to compute them (Word2Vec, GloVe, BERT). You'll get to see how to solve typical NLP problems through several demos by either computing embeddings or reusing pretrained ones.

AI assistants are getting a great deal of attention from the industry and research. However, the majority of assistants built to this day are still developed using a state machine and a set of rules. That doesn’t scale in production. Tyler Dunn explores how to build AI assistants that go beyond FAQ interactions using machine learning and open source tools.

Laurence Moroney explores how to go from wondering what machine learning (ML) is to building a convolutional neural network to recognize and categorize images. With this, you'll gain the foundation to understand how to use ML and AI in apps all the way from the enterprise cloud down to tiny microcontrollers using the same code.

You're building a high-volume, expensive, robot-driven warehouse. Your robots need to get to the right place quickly, find the right item, and sort it to the right place without colliding with each other, the shelves, or people. But you don't have any robots, and you need to start writing the logic and training them. Paris Buttfield-Addison and Tim Nugent outline how to use a simulation to do it.

The Radeon open ecosystem (ROCm) is an open source software foundation for GPU computing on Linux. ROCm supports TensorFlow and PyTorch using MIOpen, a library of highly optimized GPU routines for deep learning. Jim Dowling and Ajit Mathews outline how the open source Hopsworks framework enables the construction of horizontally scalable end-to-end machine learning pipelines on ROCm-enabled GPUs.

Today traditional approaches to predictive maintenance fall short. Zaid Tashman dives into a novel approach to predict rare events using a probabilistic model, the mixed membership hidden Markov model, highlighting the model's interpretability, its ability to incorporate expert knowledge, and how the model was used to predict the failure of water pumps in developing countries.

The future of machine learning is on the edge and on small, embedded devices that can run for a year or more on a single coin-cell battery. Alasdair Allan dives deep into how using deep learning can be very energy efficient and allows you to make sense of sensor data in real time.

As robots and AI systems become more prevalent in enterprise, industrial, and home settings, there's an increasing need for well-maintained, reliable, and secure development tools and frameworks for the next-generation production-grade robots and systems. Cam Buscaron explains how to leverage large-scale cloud simulation and the Robot Operating System (ROS) to build such systems.

An evolving landscape of cyber threats demands innovation. It's time to bring AI to the fight. Carlos Rodrigues explains why it's mandatory to use bleeding-edge AI in production to improve threat detection in a worldwide company such as Siemens. The corporate network has more than 500,000 endpoint and more than 370,000 employees. The attack vectors are endless; thus, legacy approaches don't scale.

Natural language processing (NLP) is hard, especially for clinical text. Manas Ranjan Kar explains the multiple challenges of NLP for clinical text and why it's so important that we invest a fair amount of time on domain-specific feature engineering. It’s also crucial to understand to diagnose an NLP model performance and identify possible gaps.

Efforts to counter human trafficking internationally must assess data from a variety of sources to determine where best to devote limited resources. Tom Sabo explores text-based machine learning, rule-based text extraction to generate training data for modeling efforts, and interactive visualization to improve international trafficking response.

Casey Dugan and Zahra Ashktorab challenge you to guess the backdoor of a hacked classifier. Join them to learn more about novel AI technologies through the design and development of engaging games. Take a look at their latest research around improving the interactions between humans and AI systems from empathy building to feedback design.

GoFood, Gojek's food delivery product, is one of the largest of its kind in the world. Jewel James and Mudit Maheshwari explain how they prototyped the search framework that personalizes the restaurant search results by using ML to learn what constitutes a relevant restaurant given a user's purchasing history.

Any business, big or small, depends on analytics, whether the goal is revenue generation, churn reduction, or sales or marketing purposes. No matter the algorithm and the techniques used, the result depends on the accuracy and consistency of the data being processed. Sridhar Alla examines some techniques used to evaluate the quality of data and the means to detect the anomalies in the data.

Alejandro Saucedo demystifies AI explainability through a hands-on case study, where the objective is to automate a loan-approval process by building and evaluating a deep learning model. He introduces motivations through the practical risks that arise with undesired bias and black box models and shows you how to tackle these challenges using tools from the latest research and domain knowledge.

There's an exponential growth in the number of internet-enabled devices on modern smart buildings. IoT sensors measure temperature, lighting, IP camera, and more. Tuhin Sharma and Bargava Subramanian explain how they built anomaly-detection models using federated learning—which is privacy preserving and doesn't require data to be moved to the cloud—for data quality and cybersecurity.

Vignesh Gopakumar explores image mapping of the temporal evolution of physics parameters as plasma interacts with the reactor wall using a data-inferred approach. The model captures how parameters such as temperature and density evolve across space and time. By analyzing the patterns found in simulation data, the model learns the existing physics relations implicitly defined within the data.

In a future of widespread algorithmic pricing, cooperation between algorithms is easier than ever, resulting in coordinated price rises. Rebecca Gu and Cris Lowery explore how a Q-learner algorithm can inadvertently reach a collusive outcome in a virtual marketplace, which industries are likely to be subject to greater restrictions or scrutiny, and what future digital regulation might look like.

Katharine Jarmul sates your curiosity about how far we've come in implementing privacy within machine learning systems. She dives into recent advances in privacy measurements and explains how this changed the approach of privacy in machine learning. You'll discover new techniques including differentially private data collection, federated learning, and homomorphic techniques.

In every AI initiative, there’s a demand from businesses to protect or increase market share or decrease operational costs. Your competitors are a growing threat, seemingly adopting new technologies better than you. Umit Cakmak examines critical steps from countless client engagements on how to consistently deliver successful AI projects.

Paco Nathan outlines the history and landscape for vendors, open source projects, and research efforts related to AutoML. Starting from the perspective of an AI expert practitioner who speaks business fluently, Paco unpacks the ground truth of AutoML—translating from the hype into business concerns and practices in a vendor-neutral way.

The growing complexity of data science leads to black box solutions that few people in an organization understand. Mark Madsen explains why reproducibility—the ability to get the same results given the same information—is a key element to build trust and grow data science use. And one of the foundational elements of reproducibility (and successful ML projects) is data management.

Voiced-based AI continues to gain popularity among customers, businesses, and brands, but it’s important to understand that, while it presents a slew of new data at our disposal, the technology is still in its infancy. Andreas Kaltenbrunner examines three ways voice assistants will make big data analytics more complex and the various steps you can take to manage this in your company.

Karim Beguir discusses a system in which an agent that learns to pack boxes efficiently in containers while respecting multiple physical constraints. The agent is trained using reinforcement learning to minimize the wasted space. Without any human knowledge, the agent achieves superhuman performance and outperforms commercial optimization software.

Over the last few years, convolutional neural networks (CNNs) have risen in popularity, especially in the area of computer vision. Many mobile applications running on smartphones and wearable devices would benefit from the new opportunities enabled by deep learning techniques. Siddha Ganju and Meher Kasam walk you through optimizing deep neural nets to run efficiently on mobile devices.

Holger Kyas details the AI multicloud broker, which is triggered by Amazon Alexa and mediates between AWS Comprehend (Amazon), Azure Text Analytics (Microsoft), GCP Natural Language (Google), and Watson Tone Analyzer (IBM) to compare and analyze sentiment. The extended AI part generates new sentences (e.g., marketing slogans) with a recurrent neural network (RNN).

Getting machine learning models ready for use on device is a major challenge. Drag-and-drop training tools can get you started, but the models they produce aren’t small enough or fast enough to ship. Jameson Toole walks you through optimization, pruning, and compression techniques to keep app sizes small and inference speeds high.

What are the essentials steps to take in order to develop an AI solution? How long would this process would take? As machine learning is teaching us, the answers can be learned from previous experience. Walter Riviera walks you through a collection of real-life stories, looking for successful and misleading behavioral patterns.

Carlos Escapa takes a deep dive into how to identify use cases for ML, acquire cutting-edge best practices to frame problems in a way that key stakeholders and senior management can understand and support, and set the stage for delivering successful ML-based solutions for your business.

Topic Table discussions are a great way to informally network with people in similar industries or interested in the same topics.

Program chairs Ben Lorica, Roger Chen, and Alexis Helzer open the second day of keynotes.

From people you may know (PYMK) to economic graph research, machine learning is the oxygen that powers how LinkedIn serves its 630M+ members. Zhe Zhang provides you with an architectural overview of LinkedIn’s typical machine learning pipelines complemented with key types of ML use cases.

Reinforcement learning is an advanced machine learning technique that makes short-term decisions while optimizing for a longer-term goal through trial and error. Ian Massingham dives into state-of-the-art techniques in deep reinforcement learning for a variety of use cases.

Ihab Ilyas highlights the data-quality problem and describes the HoloClean framework, a state-of-the-art prediction engine for structured data with direct applications in detecting and repairing data errors, as well as imputing missing labels and values.

Walter Riviera details three key shifts in the AI landscape—incredibly large models with billions of hyperparameters, massive clusters of compute nodes supporting AI, and the exploding volume of data meeting ever-stricter latency requirements—how to navigate them, and when to explore hardware acceleration.

Machine learning solutions are revolutionizing AI, but Marta Kwiatkowska explores their instability against adversarial examples—small perturbations to inputs that can catastrophically affect the output—which raises concerns about the readiness of this technology for widespread deployment.

It's hard ignore the attention given to autonomy and robotics. The impact is significant and the reach is extensive, hitting transportation with self-driving cars, logistics and supply with mobile robots, and remote sensing applications with aerial vehicles or drones. Raffaello D'Andrea explores how autonomous indoor drones will drive the next wave of autonomous robotics development and growth.

O'Reilly AI program chairs close the second day of keynotes.

Ready, set, network! Meet fellow attendees who are looking to connect at the AI Conference. We'll gather before Wednesday and Thursday keynotes for an informal speed networking event. Be sure to bring your business cards—and remember to have fun.