Cognitive Computing Forum

Dates: August 20 – 21, 2014
Location: San Jose, CA

In recent years, numerous technological advancements have combined to give machines a greater ability to understand information, and to learn, to reason, and act upon it. These advancements have reached such sophistication that in some cases machines may even appear to think. As a result, the broad term used to describe this emerging capability is Cognitive Computing.

Agenda

August 20
8:30

Cognitive Computing and the Future

Chris Welty

Research Scientist, IBM

At WWW 2011, soon after the notable performance of Watson on Jeopardy, I laid out the skeleton of a new computing paradigm, which IBM has since dubbed “Cognitive Computing”. Over the three years since then, cognitive computing is indeed proving to be a radical shift in software and information technology, that disrupts previously understood terms like “performance”, “feature”, “debugging”, and even “truth”. In this talk I will give a personal perspective on how cognitive computing has progressed, and is re-shaping the software solution business and ecosystem, and our very expectations of what computers are capable of.

9:15

Machine Learning Platforms

Joshua Bloom

CTO, Wise.io

As CTO, Josh is responsible for establishing, driving, and communicating the overall technical direction for Wise.io. He is a self-described “data-driven scientist” who leveraged the power of machine learning to enable landmark astronomy discoveries. He is a professor at the University of California, Berkeley where he teaches astrophysics and Python for data science. Josh has been a Sloan Fellow, Junior Fellow at the Harvard Society, and Hertz Foundation Fellow. He has published over 250 journal articles, and in 2010, was awarded the Pierce Prize from the American Astronomical Society. Josh holds a PhD from Caltech and degrees from Harvard and Cambridge. He serves on the Berkeley Startup Cluster Advisory Committee and is an active mentor and guest lecturer at the Haas School of Business.

9:55

Break

10:20

Deep Neural Networks in Practice

Dave Sullivan

Primary Architect, Ersatz Labs

Deep learning has recently achieved impressive results in several areas of machine learning. However, practical advice on how to use neural networks continues to be difficult to find. This talk is geared towards practitioners that would like to avoid the theory of how deep neural networks work and focus more on how and why they might use them in practice.

11:00

Cognitive Computing Is High Dimensional and Holographic

Pentti Kanerva

Project Scientist, UC Berkeley’s Redwood Center of Theoretical Neuroscience

The brain’s circuits suggest that the basic unit to compute with be a large pattern or a multicomponent vector (e.g., a 10,000-bit word) rather than a number, and that information encoded into a vector be distributed over all its components. Neural nets and deep learning are a step in that direction but need to be complemented with operations on multicomponent vectors that make fully general computing possible.

11:40

Lightning Talks: Ideas at the Speed of Thought

The field of Cognitive Computing is rich with ideas and innovation, so in this session we’re inviting both established ventures and startups to share their ideas and tell us about their approaches in a series of 5-minute lightning talks. Contact Tony Shaw at tony@dataversity.net if you’re interested in applying for a presentation slot.

12:20

Lunch

Lunch will be provided for registrants.

1:30

The Future of Communication: Human Insight at Machine Scale

Kristian Hammond

Chief Scientist, Narrative Science

We have entered a new era with regard to our relationship with the machine. Computers have become integrated with our everyday lives, especially with the advent of mobile and multi-device technology, and they finally know more than we do about many of the things we care about. Unfortunately, this now leaves us with massive amounts of data that is exponentially growing as companies and organizations meter and monitor everything – every transaction and every interaction. Surprisingly, however, the machine’s ability to communicate the insights contained in that data has been strikingly limited. Data in its raw form is overwhelming. Tables and spreadsheets require calculation and correlation. Charts and graphs require interpretation. What people really need is the story. In this talk, we will look at Narrative Science’s Quill™, an Artificial Intelligence engine that understands data and uses it to craft narratives that explain what that data mean in clear, concise English language. Based on a foundation of Narrative Analytics, the technology can analyze any data set, identify important insights and then automatically generate a communication for a variety of audiences, skill levels and delivery formats. Put simply, Quill provides human understanding and insight at machine scale.

2:10

Cognitive Computing with Associative Memories: Reasoning by Similarity

Paul Hofmann

CTO, Saffron Technology

Cognitive Computing with Associative Memories: Reasoning by Similarity We combine two very powerful ideas, Associative Memories and Kolmogorov Complexity for Cognitive Computing in order to make meaning from huge data sets in real time. Associative Memories mimic how humans learn and think but much faster and more powerfully. The Associative Memory functions as a universal NoSQL graph representation of structured and unstructured data. The universal cognitive distance based on Kolmogorov Complexity is used for reasoning by similarity on top of the NoSQL store. We’ll show use cases from health care @Mt Sinai Hospital in NY – automatic diagnosis of echocardiograms in real time, from global risk @The Bill and Melinda Gates Foundation – real time threat scoring reading incoming emails, and from maintenance and repair @Boeing – predicting before a part breaks. Additional Info: This presentation shows real world applications (Boeing, The Gates Foundation and Mtn Sinai) for finding pattern in large data sets combining a NoSQL graph representation (Associative Memories) with state of the art machine learning.

2:50

Break

3:10

Understanding Cortical Principles and Building Intelligent Machines

Subutai Ahmad

PhD, VP Research and Development , Numenta, Inc.

At Numenta we aim to understand the computational principles underlying the neocortex, and build intelligent machines based on those principles. At its most basic level the cortex takes in a stream of sensory data, builds a sensorimotor model of the world and outputs a stream of motor actions. In this talk I will describe the cortical principles behind these functions, and how we can translate them into a working system. The core ideas have been validated in commercial streaming analytics applications. An optimized implementation is available in the open source project NuPIC. Although we still have much work to do, this work forms a solid foundation for building biologically inspired intelligent machines.

3:50

Expressive Machines

Mark Sagar

Director, Laboratory for Animate Technologies at the Auckland Bioengineering Institute

The Laboratory for Animate Technologies research is pioneering biologically based methods to give computers the power of expression and naturally intelligent interaction. Our research combines face simulation with computational neuroscience models to create interactive self-animated expressive Avatars which learn through interaction. Face to face interaction is vital to social learning, however detailed interactive models which capture the richness and subtlety of human expression do not currently exist. Our research aims to start to tackle the complexity of interactive expressive facial behaviours involved in social interaction and learning using a bottom up – top down approach, through the use of generative low level neurobiological models which are be modulated by higher level neural system models. To do this we are developing a general simulation framework and modeling language for the integration of established and emerging practical models from the computational neuroscience literature with interactive computer graphics animation. An example of the approach we are taking is embodied in BabyX, an experimental computer generated psychobiological simulation of an infant combining models of the facial motor system and theoretical computational models of basic neural systems involved in interactive behaviour and learning. Exploring the fundamental mechanisms of early communication will lay the groundwork for human computer interfaces of the future.

4:30

Augmented Attention: First Step to an Artificial Unconscious?

Karl Schroeder

Science Fiction Author

Science fiction narratives privilege consciousness in computer interactions. Characters “jack in” to “higher states of consciousness” in blissful union with the machine. But what if it were better for us to go the other way? Technological culture is enabling us to drive many previously manual, conscious activities into a kind of technologically-mediated unconscious. A banal but important example is cell-phone roaming, where actions of connection and disconnection that used to be deliberate and consciously undertaken are now automatic, invisible, and in large part unchosen. At the same time we can for the first time choose to make ourselves aware of previously unconscious or inaccessible interactions. Eagle Cams and facial recognition systems for animals promise to socialize previously opaque relationships with other entities in our natural environment (you can potentially know your neighborhood racoons as individuals). What is emerging is a broadening spectrum of preferences, or presets, that allow us to customize our personal and collective umwelt to a degree inconceivable to previous generations. We may be approaching a decision point where we will need to declare what about ourselves and our world we will be aware of, and what we will deliberately consign to a new, technologically mediated version of the Unconscious.

5:15

Search, Structure and Knowledge on the Web

RV Guha

Fellow, Google

A significant fraction of the pages on the web are generated from structured databases. A longstanding goal of the semantic web initiative is to get webmasters to make this structured data directly available on the web. The path towards this objective has been rocky at best. While there have been some notable wins (such as RSS and FOAF), many of the other initiatives have seen little industry adoption. Learning from these earlier attempts has guided the development of schema.org, which appears to have altered the trajectory. Three years after its launch over 5 million Internet domains are using schema.org markup. Google has leveraged structured data in delivering its Knowledge Graph, and users are able to start asking more complex questions and find more relevant information more quickly than ever before. In this talk, we recount the history behind the early efforts and try to understand why some of them succeeded while others failed. We will also discuss some of the interesting research problems being addressed in the context of current efforts.

August 21
8:30

Never-Ending Language Learning

Tom Mitchell

Professor, Carnegie Mellon University

We will never really understand learning by machines or by people until we can build machines that learn many different things, over years, and become better learners over time. We describe our research to build a Never-Ending Language Learner (NELL) that runs 24 hours per day, forever, learning to read the web. Each day NELL extracts (reads) more facts from the web, into its growing knowledge base of beliefs. Each day NELL also learns to read better than the day before. NELL has been running 24 hours/day for over four years now. The result so far is a collection of 70 million interconnected beliefs (e.g., servedWtih(coffee, applePie)), NELL is considering at different levels of confidence, along with millions of learned phrasings, morphological features, and web page structures that NELL uses to extract beliefs from the web. NELL is also learning to reason over its extracted knowledge, and to automatically extend its ontology. Track NELL’s progress at http://rtw.ml.cmu.edu, or follow it on Twitter at @CMUNELL.

9:15

Compositional Model Selection

Roger Grosse

Fellow, University of Toronto

In Bayesian machine learning, models are often built by composing simpler motifs, such as clustering, factor analysis, and binary attributes. This compositional structure has allowed probabilistic models to be tailored to domains as diverse as vision, language, and medicine. Unfortunately, it also presents a challenge: identifying the right model and developing effective inference algorithms both require considerable human time and expertise. I’ll present a grammar of matrix decomposition models whose production rules correspond to simple probabilistic modeling motifs. The compositional structure of the grammar enables generic algorithms for posterior inference and model scoring across thousands of model structures. A greedy search over the grammar automatically identifies sensible models for datasets as diverse as image patches, motion capture, 20 Questions, and U.S. Senate votes, all using exactly the same code.

10:00

Break

10:20

The Promise of Prediction Markets

Robin Hanson

Chief Scientist, Consensus Point

The world is full of organizations that make bad decisions, because the people with relevant info don’t have incentive to admit or reveal that info. Prediction markets have shown a consistent ability to introduce better incentives, allowing more accurate estimates, for better decisions. We review the mechanisms that enable prediction markets, data on their use, and many practical problems with building and fielding them.

11:00

Predicting Models of Human Performance

Vivienne Ming

Chief Scientist, Gild

The dark art of identifying talented professionals for recruitment and promotion has become an obsession in the talent wars of the tech industry. Respected companies such as Google have apply enormous resources to predicting the best developers and managers, and yet the also periodically acknowledge the short comings of their existing methodology (e.g., no more brainteasers). A growing number of companies have begun “testing” candidates by giving them short-term contracts and real problems to solve as a part of the existing team. Of course, very few working professionals can take days off from an existing job for such assessments. An alternative is to build predictive models based on the true subject of interest: the real work of a subject’s actual career. For students this means using unobtrusive technology to turn their learning experiences into rich assessments, building cognitive models using unstructured data and ubiquitous sensors. For professionals it means consuming the data of an entire career to make recommendations for hiring, promotion, and team building. We will discuss the concept of continuous passive formative assessment applied both learners and professionals, from kindergärtners to (future) CEOs.

11:40

Cognitive Computing on Hadoop, Low Tech and High Tech Approaches

Ted Dunning

Chief Application Architect, MapR

With the advent of cost-effective big data technologies, it has become practical for businesses and research group of all sizes to have access to enormous volumes of data as well as to the computational resources required to process them. This has changed the way that computers interact with humans, particularly in the way that computers now can emulate human performance and competence in a number of tasks in ways that were only previously possible in science fiction stories. What is not well known is that there is an broad collection of methods for building these systems which range from extremely simple to highly sophisticated. I will describe several of these approaches to cognitive computing that exhibit different trade-offs in terms of implementation complexity and advanced capabilities and will include results and demonstrations of these techniques. The big news in this presentation is that many of these techniques are now practical even if you don’t have a sophisticated data science team packed with PhD’s. If you understand your data, you can start building some of these systems right away.

12:20

Lunch

Lunch will be provided for registrants.

1:30

Seven Myths about Human Annotation

Lora Aroyo

Associative Professor, VU University Amsterdam

Human annotation is a critical part of big data semantics, but it is based on an antiquated ideal of a single correct truth. This presentation exposes seven myths about human annotation and dispels the myth of a single truth with examples from research. A new theory of truth, Crowd Truth, is based on subjective human interpretation on the same objects (in our examples, sentences) across a crowd, providing a representation of subjectivity and a range of reasonable interpretations. Crowd Truth has allowed us to identify the myths of human annotation and paint a more accurate picture of human performance of semantic interpretation for machines to attain.

2:10

Using Artificial Imagination to get Big Answers

Patrick Lilley

CEO, Emerald Logic

Brilliant experts are working on problems that matter: curing diseases, predicting natural disasters, discovering laws of physics. But expertise is bias – the more you know, the less you look around. Further, the world is not made of straight lines. Biology, human behavior, and financial markets are driven by indirect, nonlinear relationships that cannot be discovered by traditional means. It’s imperative that we move beyond statistics and pretty visualizations, to assist experts with machines that can reverse engineer real-world systems on their own, with no preconceptions or limiting assumptions. We’ll show how this is possible, with real-world cases from a variety of domains.

2:50

Break

3:10

Panel

Details coming soon.

4:00

Close