Cognitive Systems learn, reason understand, and interat with people – they use AI technologues
Cognitive transforming every industry where there is a lot of data, and horizontal applications
<!-- HTML Credit Code for Can Stock Photo--><a href="http://www.canstockphoto.com">(c) Can Stock Photo</a>
SHRDLU: A program for understanding natural language, (Terry Winograd, MIT) in 1968-70 that carried on a simple dialog with a user, about a small world of objects on a display screen. http://hci.stanford.edu/~winograd/shrdlu/
AARON - The First Artificial Intelligence Creative Artist (Harold Cohen, UCSD) 1973–present) The Aaron system composes and physically paints novel art work. It is a rule-based expert system using a declarative language. http://www.viewingspace.com/genetics_culture/pages_genetics_culture/gc_w05/cohen_h.htm
Carnegie Learning’s Algebra Tutor (1999–present): This tutor encodes knowledge about algebra as production rules, infers models of students’ knowledge, and provides them with personalized instruction. http://www.carnegielearning.com
Arthur Samuel demonstrated (1956) playing Checkers with the IBM 701 on Television. Major publicly visible milestone for Artificial Intelligence – tree searching, learning by playing itself
Gerald Tesauro (1994) developed a self-teaching backgammon program called TD-Gammon. Learning its strategy almost entirely from self-play, TD-Gammon achieved a human world-champion level of performance.
On May 11, 1997, IBM’s Deep Blue beat the world chess champion Garry Kasparov in a six-game match: Two wins for Deep Blue, One for Kasparov and Three draws.
AlphaGo is a computer program developed by Google DeepMind in London to play the board game Go.[1] In October 2015, it became the first Computer Go program to beat a professional human Go player without handicaps on a full-sized 19×19 board.[2][3] In March 2016, it beat Lee Sedol in a five-game match, the first time a computer Go program has beaten a 9-dan professional without handicaps.[4] Although it lost to Lee Sedol in the fourth game, Lee resigned the final game, giving a final score of 4 games to 1 in favour of AlphaGo. In recognition of beating Lee Sedol, AlphaGo was awarded an honorary 9-dan by the Korea Baduk Association.
D2_John_Kelly_ppt2003_FINAL
There is an enormous amount of data in the planet. According to
44,000,000,000,000,000,000,000 bytes 44 ztabytes by 2020 (by IDC / EMC)
Earlier AI Systems Stalled due to
Reliance on a large number of manually designed rules for specific purposes
Lack of sufficient computational power
Trouble scaling to complexities of real applications
Recent Trends are Driving Change
Probability and statistics provide a fundamental formalism for AI – probabilistic reasoning, graphical models, and Hidden Markov Models
More powerful and sophisticated machine learning algorithms
The availability of huge computing power and vast amounts of data
Individuals overwhelmed by information overload in private and professional lives
<!-- HTML Credit Code for Can Stock Photo--><a href="http://www.canstockphoto.com">(c) Can Stock Photo</a>
<!-- HTML Credit Code for Can Stock Photo--><a href="http://www.canstockphoto.com">(c) Can Stock Photo</a>
<!-- HTML Credit Code for Can Stock Photo--><a href="http://www.canstockphoto.com">(c) Can Stock Photo</a>
<!-- HTML Credit Code for Can Stock Photo--><a href="http://www.canstockphoto.com">(c) Can Stock Photo</a>
Talk about today – feature extracting and brittle code
ML: Speed, Scale, New Models
Learned Representations and Reasoning – mixing inference and statistics and probability
New Kinds of Queries
Reasoning – Mixing
ML at Scale (e.g. Comp-Stat Learning and Optimization)
Non-standard paradigms (e.g. Learning from much less data)
Deep Learning++ (e.g. hybrid architectures)
Actionable and interpretable learning (e.g. Learning causal, structural and sparse models)
ML for Knowledge Extraction, Representation, and Reasoning (e.g. Automated Knowledge Base Construction)
Semantic document representation
Rapid creation of new knowledge bases
”Automated” knowledge modeling by domain experts
Integrated symbolic and learned approximate/probabilistic reasoning
Learning on the job
Enhance Watson R&R with state-of-the-art capabilities for querying and question answering, such as improved ranking, passage retrieval, answer selection/generation, similarity search and more.
Dynamic query & retrieval models that adapt during the interaction with the user (e.g. search session or dialog)
Ontology-driven querying of annotated documents and extracted entities.
Supporting natural language query interfaces as well as programmable (domain-specific) APIS
Long Term Goal (< 3 yrs)
Support for multiple retrieval pipelines
Answer Generation (NLG)
Leveraging usage data - Interactive Retrieval, Usage data analysis
Ontology driven querying
Personalized Retrieval – personalize according to user profile, intent/task and context
Talk about today – feature extraction and brittle code
ML: Speed, Scale, New Models
Learned Representations and Reasoning – mixing inference and statistics and probability
New Kinds of Queries
Reasoning – Mixing
No support for user-specific answers to be synthesized
No support for extracting quantities, semantic mapping, nor any math
Requires precise and complete answers with high confidence
Requires identifying appropriate formula, and semantic mapping of values to variables
Questions are often ill-posed
Units and types may be unspecified
Context and formula inputs required from a variety of sources
Dialog and explanation expectations
Short Term Goal (< 1 yr)
Services : Recommender Service piloted in WCA / Retail V.A. that is based on Decision Dialog and Voyager
Solutions: IBM Cognitive Recommender Engine (CoRE) for CAO, M&A, [Boson] Assisting flight crews with diversion scenarios – validated & delivered to client, Decision Agent for Disease Grading and Patient Triaging - validated & delivered to client
Long Term Goal (< 3 yrs)
Services: group decision making, decision gisting
Solutions: Watson Care Manager recommender system for care planning – transferred to Watson Health, Decision Agent for Disease Grading and Patient Triaging – Transferred to Watson Health
Goes beyond factual question answering
Helps humans make decisions and persuade others by automatically constructing pro and con arguments
Mines huge corpora of textual data. The claims are backed up with relevant evidence
The distinctive debating technologies developed in this project can have great practical use in industries such as government, legal, finance, healthcare, and sales, to name just a few. For example, automatic argument construction could serve to dramatically enhance business processes and decision making – whether by providing assisted reasoning for which treatment will work best on a patient, or by helping salespeople develop persuasive arguments when working with clients in deal negotiations, or by presenting pro and con arguments in support of or against government policies.
Old way:
User acceptance determined by usability and desirability
New way
User acceptance determined by engagement, effective communication and ease of participation
Objects aware of those interacting with them: physical and virtual embodiments:
Model, plan, represent, sense, respond
Dialog is between a person and a cognitive system and can be via different interaction modes (e.g. speech, text, gestures, etc.).
Create an architecture for integrating contextual understanding, various inference engines, language generation, and user modeling such as emotions, personalities, and other important contextual information
1900: TABULATION
Punched card tabulation
Scale, automation
Seeds of future innovation
1950: PROGRAMMING
Stored data, instructions
Languages for computing
Metrics for computation
2011: COGNITION
Massive data scale
Data for training
Real-world modalities
Cognitive Systems learn and interact naturally with people to amplify what either humans or machines could do on their own. They help us solve problems by penetrating the complexity of Big Data.
<!-- HTML Credit Code for Can Stock Photo--><a href="http://www.canstockphoto.com">(c) Can Stock Photo</a>
Cognitive systems are more challenging to develop, deploy, and manage because a critical component (model) is created from data and requires domain expertise.
Cognitive systems are more challenging to develop, deploy, and manage because a critical component (model) is created from data and requires domain expertise.
Models are new kinds of artifacts, then need to be secured, composed, trained in a context – they life in a hostile environment
Models have a lifeycle
Deep Learning Computing Platform: Big data and the explosion in compute needs of machine/deep learning has made training and inference expensive, time-consuming, and fraught with complexities.
Cloud-based training and inferencing services, with accelerators improve developer and scientific productivity.
<!-- HTML Credit Code for Can Stock Photo--><a href="http://www.canstockphoto.com">(c) Can Stock Photo</a>
Six years ago, IBM and our university partners embarked on a quest—to build a brain-inspired machine—that at the time appeared impossible. Today, in an article published in Science, we deliver on the DARPA SyNAPSE metric of a one million neuron brain-inspired processor. The chip consumes merely 70 milliwatts, and is capable of 46 billion synaptic operations per second, per watt–literally a synaptic supercomputer in your palm.
Along the we have journeyed from neuroscience to supercomputing, to a new computer architecture, to a new programming language, to algorithms, applications, and now to a new chip—TrueNorth.
Considering overall energy consumption underscores the divergence between the brain and today’s computers even more starkly. Note that a “human-scale” simulation with 100 trillion synapses (with relatively simple models of neurons and synapses) required 96 Blue Gene/Q racks of the Lawrence Livermore National Lab Sequoia supercomputer—and, yet, the simulation ran 1,500 times slower than real-time. A hypothetical computer to run this simulation in real-time would require 12GW, whereas the human brain consumes merely 20W.
To support these algorithms at ever increasing scale, TrueNorth chips can be seamlessly tiled to create vast, scalable neuromorphic systems. In fact, we have already built systems with 16 million neurons and 4 billion synapses. Our sights are now set high on the ambitious goal of integrating 4,096 chips in a single rack with 4 billion neurons and 1 trillion synapses while consuming ~4kW of power.
Technology support is a labor-intensive business – both diagnosis and field repair.
There is a large body of prior incident reports and service requests – similar symptoms might have different root causes – server down due to full file system or hardware error
There are many resolution reports and success indicators
Diagnosis often conducted iteratively in a dialog, pruning potential causes to the most likely ones
Knowledge Extraction and Representation:
Enhance Knowledge Base (with domain vocabulary, instances, constraints and rules) to help current way of working (for Explicit, e.g. dialogue and NLC).
The input to KB should be from domain experts, input from humans and historical data
Dialog:
Implicit: Create an ontology/representation to create coarse representation of concepts in the space together with tasks
Inferred: Need recorded dialogues and once we have that we can use learning techniques to estimate what happens next in the hardcoded dialogues/automations. This is used to build the ontology