Expansive AI: Convergence of Deep Learning and Cognitive Knowledge Computing
by Spyro Mousses PhD, CEO of Systems Oncology and Co-Founder of Systems Imagination
Genomics and big-data have the promise to make a major impact against human disease, but the reality is that the data we now have access to is too big and too complex for conventional computational approaches. Over the last decade we have worked to build the next generation of intelligent machines that can augment human intelligence and greatly accelerate our ability to generate new scientific knowledge. The dream is that this new paradigm of ‘man + machine’ will lead us to rapid progress against the diseases that plague our society. The good news is we are closer than ever to realizing that dream.
The field of machine learning has been making slow and steady progress for decades. It wasn’t until 2012 that the field reached a major turning point when Geoff Hinton PhD, a computer scientist at the University of Toronto, published some extraordinary results. Briefly, Dr. Hinton was able to show that the use of deep convolutional neural networks can leverage massive datasets of annotated images to train models which were surprisingly effective at learning how to classify objects in images. Since this seminal discovery, this new field of ‘deep learning’ has achieved rapid evolution, with diverse applications ranging from the game “Go” to diagnosing diabetic retinopathy, demonstrating unprecedented accuracy that approximates and often surpasses the abilities of human experts. This activity has led to a true renaissance in artificial intelligence, and many are now wondering where the field is going next.
In parallel, our team has been making progress in advancing the field of cognitive computing and knowledge representation. Specifically, our goal was to develop a computational framework capable of representing the complex biomedical models . By leveraging the flexibility and scalability of nested hypergraphs, we were able to represent complex multi-dimensional relationships across multiple scales of biomedical knowledge. This new approach has completely transformed our ability to not only mine large complex datasets to recover relevant knowledge, but more sophisticated representation strategies also enabled us to explore new approaches to machine learning. This is leading us to the creation of machines that are capable of autonomously generating new knowledge. This idea, that computer systems can ‘imagine’ novel beliefs and then test and justify those computer-generated beliefs through the use of machine learning and big-data, represents a new paradigm for how scientists can use machines to massively amplify their ability to generate new knowledge. This new ability to imagine truly novel beliefs is leading to systems that generate new knowledge, an ability has been the exclusive domain of the human mind. The difference between how humans and computer-generated knowledge is of course scale. Scientific knowledge computing can now generate new scientific knowledge many orders of magnitude faster than what is possible by humans.
Recently, the parallel paths of deep learning and scientific knowledge computing converged. This convergence started when we began to apply the rapidly emerging deep learning tools to mine lined data stored in a hypergraph knowledge base. In one example, we used the features from the hypergraph knowledgebase as multi-dimensional inputs to an artificial neural network model. Specifically, we were able to use known synthetic lethal pairs of genes (representing potential cancer drug targets) to train the model. We also fused in multiple other functional classifiers and dimensions of relationships across the entire cancer genome, which are naturally captured in the structures of the hypergraph knowledge network. The results of the experiment were impressive both from the output and in terms of the computational efficiency. The model was able to classify known synthetic lethal pairs, but also identified completely novel gene pairs which we are now investigating as new drug targets. Unexpectedly, we also learned that combining these two converging technologies produced unexpected computational synergy. Namely, the hypergraph framework allowed us to achieve unexpected computational efficiency and flexibility in extracting high dimensional features on an unprecedented scale, so models can be trained and optimized much more efficiently.
From an epistemological perspective, this convergence of technologies is enabling a new type of consilience, and fundamentally changing the science of knowledge itself. In other words, as we begin to generate knowledge on scale, we began to realize two things. The first is that imagination is at the heart of new knowledge generation. So, the rate limiting step to knowledge generation is not just data, but the number of beliefs or hypotheses that a system can generate and test. This is why we need computer imagination - to explore a massively more expansive universe of hypotheses. The second is that knowledge is not a single static entity to be discovered and archived. Instead all scientific knowledge evolves rapidly. Think of evolving knowledge like a dynamic network of beliefs that are constantly converging with evidence from multiple sources. This type of continuous consilience will disrupt our conventional notions of science and evidence-based medicine, but if done right, it will also transform our ability to make an impact.
Taken together, we now see the combination of deep learning and cognitive knowledge computing as a very powerful new platform capable of revolutionizing science. One that will enable us to explore a massively more expansive universe of knowledge – hence the name, ExpansiveAI. Systems Imagination will continue to advance the ExpansiveAI platform technology and Systems Oncology has begun to apply this approach in the field of cancer therapeutics to achieve scientific breakthroughs beyond human imagination.
1. Krizhevsky, Alex; Sutskever, Ilya; Hinton, Geoffrey E. (2012-12-03). "ImageNet classification with deep convolutional neural networks". Curran Associates Inc.: 1097–1105.
2. Farley T, Kiefer J, Lee P, Von Hoff D, Trent JM, Colbourn C, Mousses S. (2013) The BioIntelligence Framework: a new computational platform for biomedical knowledge computing. J Am Med Inform Assoc. 2013 Jan 1;20(1):128-33. Epub 2012 Aug 2.