Building Blocks to the Future: Cognitive OpenTech
Now consider the relative importance of big Data, Cloud compute power, and new Algorithms as a Service in making progress… we can call all these factors the DCAaaS drivers of progress.
In 2011, IBM Watson Jeopardy! victory on the TV quiz game show would not have been possible without the existence of Wikipedia – big data that was crowdsourced, and represents a compilation of knowledge across human history, including recent movies, sports events, political changes, and other current events as well as historic events. Wolfram had an interesting analysis on how close “brute force” approaches were coming to this type of Q&A task, based on compiled human knowledge or facts. In many ways this is an example of GOFAI – Good-Old-Fashion AI, with a twist. GOFAI includes “people built giant knowledge graphs,” such as ConceptNet. The modern twist that is now available, but was not available in the 1980’s, is crowdsourcing the construction of the “big data.”
In 2016, Google/DeepMind AlphaGo victory in the game Go would not have been possible without synthetic data, massive amounts of data generated by brute force simulated game playing. In 2017, CMU Libratus victory in poker (Texas-Hold-Em) was also dependent on big data from simulated game playing. Generating synthetic data sets based on foundational crowdsourced data sets has been key to many recent ImageNet Challenge annual performance improvements/victories. Additional “big data” that is synthetic data generated from crowdsourced data is a hot topic with OpenAI’s Universe project (background: generated data) as well.
Three speakers explain the importance of big Data, Cloud compute, and Algorithm advances as a Service (DCAaaS) or simple “better building blocks” – see:
|Andrej_Karpathy (OpenAI) https://www.youtube.com/watch?v=u6aEYuemt0M|
|Richard_Socher (Salesforce) https://www.youtube.com/watch?v=oGk1v1jQITw|
|Quoc V. Le (Google) https://www.youtube.com/watch?v=G5RY_SUJih4|
In addition to “big data” that is (1) crowdsourced, like Wikipedia and ImageNet, and (2) machine generated (“Synthetic Data”) as in AlphaGo, Libratus, and OpenAI Universe, each of us has a stock pile of (3) personal data on our computers, smartphones, social media accounts, etc.
Rhizome’s Blog has an interesting post about Web Recorder tool. Web Recorder is a tool for greatly expanding the amount of personal data, while also aggregating it as part of a type of internet archive and our personal browsing history of things we find interesting on the web. A type of collective, digital social memory is emerging.
In sum, more and better data, compute, and algorithms are fueling the rapid pace of Cognitive OpenTech developments.
Part 2: Grand Challenges of AI/CogSci Progress
A universal architecture for machine intelligence is beginning to emerge. The universal architecture that is emerging is a dynamic memory. Imagine a dynamic memory that stores and uses information better than any processes known of in the past. This capability provides a type of episodic memory of text, pictures, and videos for question answering (see minute 50+ in the Socher video above). The dynamic memory includes both RNN (Recurrent Neural Net) models as well as large knowledge graph (as found in GOFAI) models for making inferences, and answering questions or making other types of appropriate actions.
What is a dynamic memory good for? Most of us have taken a standardized test with story questions. The test taker is asked to read a story, look at a sequence of pictures, or watch a video and then answer some simple questions. In grade school, these “story tests” are simple commonsense reasoning tasks, where the answer is always explicit in the story. As we get older, the stories get harder, inference is required beyond commonsense knowledge, tapping into “book learning” and “expert knowledge” that has been compiled for centuries. Some story questions we can answer based on short-term memory (STM), and others require long-term memory (LTM). A universal architecture that is a dynamic memory can combine appropriately both STM and LTM for question-answering.
|Context___________________________________||Right Ending||Wrong Ending|
|Gina misplaced her phone at her grandparents. It wasn’t anywhere in the living room. She realized she was in the car before. She grabbed her dad’s keys and ran outside.||She found her phone in the car.||She didn’t want her phone anymore.|
The example above is interesting, and the ConceptNet5 website FAQ (very end) reports: Natural language AI systems, including ConceptNet, have not yet surpassed 60% on this test.
As highlighted above in the Karpathy, Socher, Le videos – data in the form of sequences of text, sequences of images, as well as sections of videos (and audio recordings) – are all being used as input to tell simple stories. These stories (data) are snippettes of external reality representation – with some measure of internal model representation feedback loops – so are approaching a (1) experience representation and (2) episodic memory – what Schank called “dynamic memory” – that is beginning to be used in story processing and question-answering tasks – what Schank called “scripts, plans, goals, and understanding.”
The remaining grand challenge problems of AI/CogSci are being worked on by university, industry, and government research labs around the world, and rapid progress is expected, thanks in part to cognitive opentech – data, cloud (compute), and algorithms as a service offering, and very easy to access, including from smartphones that never leave our side as we operate in today’s world. The models being generated will have more and more universal applicability over time, and should boost the creativity and productivity of end-users who use these technologies to solve new and interesting problems, as advocated by Gary Kasparov. Kasparov, the world champion grand master player, lost chess games to Deep Blue in 1996 and again 1997. Today, noteworthy in the news, Gary Kasparov is now learning to love machine intelligence.
IA (Intelligence Augmentation) is a long-standing grand challenge that involves both people and machine intelligence together – thinking better together. IA is the key to what the NSF, JST, VTT, OECD and other organizations have started referring to as smarter/wiser service systems. IBM has made a lot of contributions to intelligence augmentation; Both intelligence augmentation and collaborative intelligence will benefit the world.