The IBM Watson AI XPRIZE® UCSD CENI competition team
Enabling machine ‘thinking’ and new ideas in cognitive computing systems. And using these capabilities to put humans on Mars.
The Center for Engineered Natural Intelligence (CENI) at UC San Diego has been selected to take part in the IBM Watson AI XPRIZE ®. The competition aims to accelerate the development and adoption of artificial intelligence (AI) technologies that are truly scalable and have the capacity to solve grand challenges facing society.
Our goal is to give some of the world’s most sophisticated computing systems the ability to come up with new ideas and perform “creative machine thinking.” And ultimately to allow machines to say things in natural language that they might not have been trained for (in order to allow them to express the new ideas they came up with).
This work has applications for a wide range of areas, including contextual data analytics, transportation systems, communications networks, studying the spread of infectious diseases as well as malware, cybersecurity, genomic medicine, autonomous vehicles, and the study of the biological brain and neurological disorders. But the long term goal of our work is to use the capabilities and technologies we are building to have cognitive computing systems think creatively in order to come up with disruptive novel designs for self sustaining eco-habitats for putting humans on Mars.
Our Grand Challenge
To develop a unique and novel mathematical framework and associated algorithms that will build algorithmic and processing layers on top of existing cognitive computing systems to enable machines to ‘think’ creatively on their own in order to arrive at original ideas and thoughts about a specific problem or class of problems the system is working on. Our goal is to have cognitive computing systems be able to achieve such creative machine thinking through their own internal recourses and to be able to have them express those thoughts using natural language in its interactions with humans in real time.
We are building algorithms and software that leverage existing machine learning capabilities in order to enable creative machine ‘ideas’ by morphing the existing structure and dynamics of internal representations of knowledge graphs associated with a learned knowledge base. Our methods are fundamentally different from existing machine learning in how we mine the solution space to identify context-relevant patterns in the data, and in how we go beyond those patterns by morphing and manipulating the solution space to produce new patterns (ideas) in the machine's ‘brain’ not originally present in existing associations or data patterns. This is a simplified machine computational analogue to original thought and creativity by the human brain.
Our goal: Getting humans to survive on Mars (and helping Earth in the process)
While the capabilities we are developing have a wide variety of potential applications to many academic fields and industry verticals, the specific long term goal of our work is to use the capabilities and technologies we are building to have cognitive computing systems think creatively in order to come up with disruptive novel designs for self sustaining eco-habitats that will allow humans to survive on Mars. The challenges in achieving this are huge. And that is precisely why we think that machine generated ideas to help humans survive on Mars for extended periods of time is a great application of our work.
But helping humans to survive on mars and what we learn in the process will also enable us to help improve humans' sustainability right here on earth and the sustainability of our own planet. Disruptive and innovative improvements in building and construction designs, materials properties, energy efficiency, monitoring and control in buildings, and designs intended to withstand extreme environmental conditions could aid in the safety and improvement of urban areas, allow humans to build in more inhospitable places on our planet in an environmentally safe way, and improve quality of life. These are all issues critical to our survivability that we can impact with the technologies were are developing for the XPRIZE competition here at home.
Our biggest asset is the incredible group of faculty, students, programmers, engineers and scientists that have come together to solve these challenges. The team includes faculty, students, and postdocs from bioengineering, physics, mathematics, computer science and engineering, cognitive sciences, psychology, neurobiology, and neurosciences at the University of California San Diego, the Salk Institute for Biological Sciences, and Scripps Institution of Oceanography at UC San Diego.
The faculty leadership for the CENI XPRIZE team are:
Henry Abarbanel, Professor of Physics at UC San Diego and a research physicist at the Scripps Institution of Oceanography
Gert Cauwenberghs, professor of Bioengineering and Neurobiology (Integrated Systems Neuroengineering lab)
Fan Chung Graham, Professor of Mathematics and Computer Science and Engineering
Jeffrey L. Elman, Distinguished Professor of Cognitive Science, and Chancellor’s Associates Endowed Chair
Timothy Gentner, Professor of Psychology and Neurobiology (Gentner lab at UC San Diego)
Terry Sejnowski, Professor of Neurobiology at UC San Diego and Professor at the Salk Institute for Biological Sciences (Computational Neurobiology lab at the Salk)
Gabriel A. Silva, Professor of Bioengineering and Neurosciences. Silva leads the Mathematical Neuroscience lab @ UCSD and is the Founding Director of Center for Engineered Natural Intelligence
The team also includes a number of very talented graduate students, postdoctoral researchers, research scientists and programmers, including Nirupama (Pam) Bhattacharya (bioengineering research scientist); Vivek George (bioengineering graduate student); Nick Grayson (electrical engineering graduate student); Francesca Puppo (BioCircuits Institute and CENI postdoc); Brad Theilman (neurosciences Ph.D. student); Kai Chen (bioengineering Ph.D. student); Tim Sainburg (psychology graduate student); Nirag Kadakia (physics graduate student); Daniel Breen (physics graduate student); Alexandra Sherman (physics graduate student); Bruno Pedroni (bioengineering graduate student and IBM graduate fellow); Jun Wang (bioengineering graduate student); Hesham Mostafa (Institute for Neural Computation postdoc); Sinan Askoy (mathematics graduate student); and Josh Tobin (mathematics graduate student).
The technical details
Our own lab's core technical contribution to this work is the construction and theoretical analysis of a framework that models the competing dynamics of incident signals into nodes along directed edges in a network. From a neurobiological perspective we are abstractly modeling the interplay of postsynaptic spatial and temporal summation. The framework is constructed in order to explicitly analyze the dynamics between the offset in the latencies of propagating signals, which are the result of the geometry (morphology) of the edges and conduction velocities, and the internal refractory dynamics and processing times of the downstream node. We are using this framework to construct a new class of recurrent artificial neural network called geometric dynamic networks that will allow us to encode data and concepts as internal representations of network states and structures (connectivity and physical geometry). We can then dynamically morph these networks using other algorithms we are developing in order to achieve modified representations that encode new machine concepts and 'ideas' related to but different than the input data that created the networks in the first place.
Signals traveling between connected neurons in the brain, and the information they represent, can only travel through the specific pathways formed by the physical connections in the network. The totality of all these pathways represent the full solution space over which patterns of information can be represented and encoded by the brain. But not all neurons are capable of being receptive to incoming signals from upstream neurons they are connected to at every moment in time, which in turn means not all possible patterns are accessible or realizable at every moment. The internal state of a neuron, independent of the incoming signals they are receiving, determines whether at a given moment in time that neuron can be receptive to any such incoming signals. It is the complex (combinatorial) interplay between different neurons ‘internally doing their own thing’ while at the same time (unknowingly!) contributing to the huge number of paths through which information can flow at the network scale, that ultimately determines the patterns of activity in the network. This constrained information flow and the resulting emergent patterns determine, and indeed define, what the network can learn and how it manipulates information. Emergent patterns in brain networks ultimately represent different objects, different thoughts, patterns in data, and the transition from one thought to another. They are also what allows the brain to extrapolate beyond what (patterns) it has learned. These core neurobiological principles are what our theory and algorithms attempt to capture mathematically, and what makes them fundamentally different than the learning rules of existing machine learning methods based on statistical or stochastic minimization of objective cost functions. By using our algorithms to systematically morph, or change, the patterns in the data and information in the machine's ‘brain’, we can allow it to extrapolate beyond what it has learned in order to create new ideas (new patterns) that are contextually meaningful and relevant, bounded by the data the system has ingested (in order to avoid nonsensical ideas- but the more we relax this property, the more 'creative' the ideas will be).
From a more technical perspective, it is the interplay between spatial (connectivity) and temporal offsets in the dynamic signaling that produces differing patterns of information flow and states the data can represent at any given moment. Based on these principles, our algorithms efficiently identify and morph only subsets of the total solution space that represent allowable patterns in complex data sets that are actually realizable given the history (states) that preceded it, encoded as dynamic geometric graphs. By dynamic geometric graphs we mean a network in which physical connectivity, the positions of vertices in space, and the path lengths of the edges, in addition to dynamic signaling and the internal models and states of individual nodes, matter. The identification of contextually and situationally relevant subsets of the total solution space is what provides context to a specific situation, and therefore what will ultimately allow cognitive computing systems to come up with contextually and situationally relevant new ideas at a specific moment in time under a particular set of circumstances. Our algorithms identify only the patterns (solutions) in the data that are possible among all the patterns (solutions) that could be possible, in effect, allowing the machine to come up with relevant and timely novel ideas under a given set of unique circumstances. These ideas are encoded as novel dynamic internal graph representations derived from but uniquely different than the data presented to it during the ‘ingestion’ phase.
And more broadly, the team is pursing a number of theoretical and computational directions related to the dynamics of artificial neural networks and their relationship to neurobiology that will inform models and algorithms we are developing to further enhance and refine our capabilities. This work includes contributions from non-linear dynamics, data assimilation, and cognitive theories of learning.
Beyond machine learning: How the biological brain learns and why it matters
The biological brain represents, learns, and manipulates information very differently than the way existing artificial neural networks, other forms of machine learning, and statistical analytics methods learn and find patterns in data. As powerful as those methods are in finding hidden and subtle patterns and relationships in often very complex input data sets, these methods and algorithms are not generally designed to go beyond the data. They are generally dependent on exposure to lots of data to train them appropriately so they can learn, and continuously require significant computational resources and the consumption of large amounts of energy in order to carry out their most sophisticated and complex tasks. Most existing methods search for patterns in data by stochastically minimizing a cost function. This means that they compute the difference, or error, in a mathematical function between the learned or computed best guess solution and example data presented to it (the training set). When randomly sampling the entire solution space, if the error shrinks, the algorithms know they are learning ‘in the right direction’ (i.e. approaching the true solution), and incrementally improve their ability to recognize the same patterns presented to them in the training set. Even the most sophisticated existing methods at their core accomplish their learning and pattern recognition in this way. As a result, most existing methods are limited by a set of fundamental engineering challenges inherent to statistical learning, including a general lack of robustness and ability to adapt beyond the training sets. Existing machine learning technology based on artificial neural networks (ANNs) represent 1950’s neuroscience combined with present day computing power.
In contrast, the biological brain primarily learns by analogy and by abstracting beyond the immediate training sets presented to it. For example, the two images below of a cow are completely physically different, but a toddler very quickly learns they represent the same concept and subject, something that is simply beyond the capabilities of existing artificial neural networks and machine learning techniques. The biological brain is capable of robustly adapting to different situations and contexts it may not have previously encountered. This is even more pronounced given the minimal computational resources and energy efficiency with which our brains achieve this - using about 20 watts of power, barely enough to power a dim light bulb, in about 3 lbs. of ‘wetware’ that occupies a volume equivalent to a 2-liter bottle of soda. In contrast, a robot with the computational power of the human brain would require about 10 megawatts, or a million times more energy.
Is it a cow?
These two images of a cow are physically completely different, but a toddler very quickly begins to recognize they represent the same concept, and begins to understand that new and different images of a cow represent the same concept. Existing artificial intelligence cannot accomplish this with the efficiency and internal representation flexibility of the human brain.
The biological brain’s ability to learn by analogy and by extrapolating beyond an often limited training set, is in fact the basis for creativity and original thoughts. Machine learning and other current forms of artificial intelligence cannot achieve these higher order processes. And even though the recent rapid growth and advancement of cognitive systems has enabled primary and higher level processes such as hypothesis generation, these systems are not currently capable of knowledge extrapolation or creation.
The complexity of real networks in the brain
The image in the upper left (panel a) shows all the white matter tracts (the high speed connections) that connect the different parts of the brain, imaged using diffusion tensor imaging (DTI) MRI. The image on the right in panel b shows a group of neuronal cell bodies (small colored circles) and the massive complexity of the connections that project from all the neurons- the neuropil- above them. The image in panel c shows a 3D reconstruction of all the objects between two apical dendrites, the connections between individual pairs of neurons. One single neuron can receive 10,000's to 100,000's of such connections from other neurons. And we have approximately 85 billion neurons in our brain and about another 86 billion other cell types in the brain, some of them (astrocyte neural glial cells in particular) which can directly affect neuronal signaling and potentially modulate information processing in the brain. (Credits: Panel a- Van Weeden, Harvard-MIT, panel b- Jeff Lichtman, Harvard, panel c- Bobby Kasthuri, Boston University and Jeff Lichtman, Harvard.)
To be or not to be: AI from the biological brain?
In the broad domain of artificial neural networks, machine learning, and AI more broadly, there are a number of philosophical schools of thoughts when it comes to considering how to develop AI, and there are many other approaches and algorithms being developed for analyzing, learning, and discovering patterns in data. In fact, one is free to derive any algorithm at any level of abstraction however necessary to accomplish the target analysis. But there are a number of unique advantages to designing algorithms based on our understanding of how the biological brain learns and manipulates data and information. First, we are able to make observations and measurements on the real biological brain to guide our thinking and algorithm development. In particular, that is a unique strength of carrying such research in an academic environment. Such experimental and empirical neurobiological data motivates and informs the development of mathematical theory and models, and their algorithms and software. The biological brain is the only naturally engineered system we know of capable of exhibiting the kinds of unique computational properties we are attempting to understand and develop algorithms for. How could one not attempt to learn from it? Secondly, we have an immediate benchmark which any derived algorithms can be compared to - the very properties in the biological brain we are attempting to capture. Attempting to understand the structure and function of the biological brain both for its own sake and in an attempt to develop next generation AI is motivating and awe inspiring but terrifyingly humbling at the same time. But one thing is for certain- it has a lot to teach us. And we are now at a stage where the advanced mathematics, computing power, and (critically missing in the past!) experimental tools and methods to deeply interrogate the structure and neurophysiology of the brain and brain networks across different scales of anatomical and functional organization.
XPRIZE, a 501(c)(3) nonprofit, is the global leader in designing and implementing innovative competition models to solve the world’s grandest challenges. XPRIZE utilizes a unique combination of gamification, crowd-sourcing, incentive prize theory, and exponential technologies as a formula to make 10x (vs. 10%) impact in the grand challenge domains facing our world. XPRIZE’s philosophy is that—under the right circumstances— igniting rapid experimentation from a variety of diverse lenses is the most efficient and effective method to driving exponential impact and solutions to grand challenges. Active competitions include the $30M Google Lunar XPRIZE, the $20M NRG COSIA Carbon XPRIZE, the $15M Global Learning XPRIZE, the $7M Shell Ocean Discovery XPRIZE, the $7M Barbara Bush Foundation Adult Literacy XPRIZE, the $5M IBM Watson AI XPRIZE, the $1.75M Water Abundance XPRIZE and the $1M Anu & Naveen Jain Women’s Safety XPRIZE. For more information, visit http://www.xprize.org/
IBM Watson AI XPRIZE
Driven by the desire to accelerate human and AI collaboration for the greater good, the IBM Watson AI XPRIZE provides an interdisciplinary platform for domain experts, developers and innovators, through collaboration, to push the boundaries of AI to new heights. One of the goals of the competition is to promote wider collaboration and support from the AI community to help all innovators create scalable solutions and audacious breakthroughs to address humanity’s grandest challenges.The IBM Watson AI XPRIZE includes four rounds. Each year, the teams will be evaluated for the opportunity to advance to the next round of the competition. The three finalist teams will take the stage at the TED 2020 conference in April 2020 to deliver talks demonstrating what they have achieved. The teams will also have an option to compete for two milestone prizes along the way. For more information, visit http://ai.xprize.org/