Breakout Sessions

  1. Text Analytics as a Component in Cognitive Solutions – at the core of all text-based cognitive assistants (including assistants that have to read complex technical documents) is some kind of text analysis. Nouns and verbs are identified, context is built, phrases are clustered. What text analytics platforms do we depend upon? Which algorithms are we rewriting over and over again? Can we converge to some common services?
  2. Conversation/Dialog as a Component in Cognitive Solutions - a large number of the first generation of cognitive solutions depend on a dialog between a human (or a team of humans) and a cognitive assistant. It could be as "simple" as a chatbot, or as complex as a mediated discussion of a complex business investment or acquisition. What are the best practices for incorporating dialog into cognitive solutions, especially so we don’t have to reinvent the dialog wheel every time we build a new solution?
  3. Video understanding as a Component in Cognitive Solutions - deep understanding of audio-visual streams captured in videos is today at the forefront of AI and cognitive science research. We will explore how CHN centers can work together in this domain; which tasks we should pursue (action recognition, visual question-answering, emotion recognition etc.), data curation and annotation, and algorithmic approaches.
  4. Spatial intelligence as a Component in Cognitive Solutions - cognitive solutions know about numbers and about text and about structure data and about images...but eventually they must understand our 3d world to help humans in all their daily endeavors. What are the basic constructs, concepts, and models needed by industry applications to reason about the space we live in?
  5. Tools and Environments to Build Cognitive Systems (tools, microservices, orchestration) - cognitive systems are often built like a hydra: curating data sources, applying sophisticated machine learning techniques, innovating in the user experience to achieve a natural collaboration between human and machine, repeatedly training and testing to achieve acceptable accuracy of the ML models, and all the problems that come with developing a large conventional code base developed on microservices. How do we support cognitive developers so that "normal" developers can perform as experts? How do we support teams of domain experts and developers to produce reliable solutions?
  6. Improving Solution Performance - systems that learn require vast amounts of data … and vast amounts of computation to be trained on all that data. What systems architectures are best for supporting training of models and then the execution of those models? How best to characterize the performance of those systems? How best to take advantage of the post-CPU kind so hardware to accelerate the training and execution? Do static training and learning in real time require different hardware and software stack support?
  7. Improving Efficiency of Deep Learning - in recent years deep learning models have proven to significantly perform better than other approaches on various tasks in domains such as vision, speech and text. However, this comes at a price: these models contain millions of parameters and thus they are difficult to train, requiring a significant amount of time and labeled data. In addition, they require high memory and energy consumption. We will discuss approaches to address these issues and improve the efficiency of deep neural networks.
  8. What defines Good Cognitive Research - what are the best practices to publish cognitive research for real impact? Are the best venues the established conferences or new workshops? Or is the real impact in publishing open source code and data? If cognitive success requires combining technology from domain experts, machine learning researchers and data scientists, how do each of them impact their own field (and get credit for the publication)? If our technology depends on reusing services from other researchers (or other providers), where is the research in combining services that already exist? How do we prioritize AI vs subject domain research? What are the compelling killer apps that demonstrate the value of our cognitive technologies?
  9. Using Humans in CS research (challenges of IRBs, privacy) - development of cognitive solutions for healthcare and wellbeing requires data from humans. From collection to utilization of the human data, we face unique challenges including ethics, privacy and dealing with IRBs. In this session we will discuss challenges associated with performing research using data generated by human subjects. For example, collection, assessment, evaluation of personal intervention, regulation, privacy, etc.
  10. Improving the User Experience of Cognitive Solutions - one critical dimension of cognitive systems is to interact with humans as other humans do: use sight, sound, voice, video, even taste and smell to be part of the human "team". Beyond the multi-modal requirements we place on cognitive solutions, what other expectations do we have for the user experience? Are there maturity models (ranging, perhaps from Eliza up to Commander Data)? What methodologies, services, and tools do developers need to meet these expectations?
  11. Security - given that modern cognitive solutions depend on training data and the orchestration of many different microservices, they present new opportunities for attack by evil doers. How best to protect the reliability and accuracy of these systems? On the other side of the equation, do cognitive techniques provide new algorithms or approaches to protect generic systems and solutions?