Dr. Peter Olausson started his career as a cognitive neuroscientist and spent over a decade at Yale University researching how our memories, motivation and cognitive control together affect decision-making. Before starting COGNITUUM, Peter was focusing on new breakthroughs in the information solutions that shape the human experience, including cognitive computing, data analytics, neuromanagement, and knowledge networks. Peter received his PhD in neuropharmacology at the University of Gothenburg in Sweden and his postdoctoral training at Yale University.
COGNITUUM has developed a general intelligence framework that provides a viable pathway towards human-level machine intelligence. The platform features continuous and real-time learning from any data source.
2. Peter Olausson PhD
Yale University Faculty (2003-2013)
Cognitive Neuroscientist
Research on Learning, Memory, Motivation and
Decision-making
Sevak Avakians
Cognitive Physicist
AI Researcher
Research on Quantum Information & Information Theory
3. Using Cognitive Neuroscience to create AI
This talk will discuss how we can create intelligent machines
instead of machines that behave intelligently
What cognitive functions are needed to create AI?
How does the brain produce these cognitive functions?
How do cognitive functions produce intelligence?
How do we use these insights to create true AI?
4. Emergence of Intelligence
Intelligence is the result of cognitive processes that results
in the ability to complete complex tasks
Biological Intelligence vs Programmed Intelligence vs Artificial Intelligence
Intelligent Behavior vs Intelligence
No clear neurobiological substrate of intelligence. Multidimensional. Not a simple
mechanism that can be programmed.
5. What is Cognitive Neuroscience?
Cognitive neuroscience is the study of the neurobiological processes
that underlie cognition.
The specific focus is on the neural connections in the brain which are
involved in cognitive functions and behavior.
Cognitive Neuroscience addresses the questions of how cognitive
activities are affected or controlled by neural circuits in the brain.
6. The core functions of the brain is
information processing, decision-making and action
The brain is a general information processing framework that has
the capacity to derive predictions and decisions from any sensory
and non-sensory information and act on the information observed
The functional domains are specialized nodes of this framework
that use a general architecture and mechanisms
7. Learning & Memory
Learning: Learning is the process of acquiring new, or modifying
existing, knowledge, behaviors, skills, values, or preferences
Memory: An active system that receives, stores, organizes,
alters and recovers information
Learning and Memory enables the ability to use knowledge and
experience to influence behavior and is a central component in
cognition
8. There are many types of Memories, all required
for a coherent cognitive system
The neurobiological substrate of memory is general, the neurons/brain region that is
processing each memory type differ
9. Learning and how memories are formed
Acquisition and Consolidation
Short-Term Memory Long-Term Memory
can be disrupted
for several hours
after learning
10. How memories are used and maintained to be
relevant
Retrieval and Reconsolidation
Short-Term Memory Long-Term Memory
11. What to Learn? Motivation & Emotion
Primary vs Incentive Motivation
Emotional valence and strength (eg. happiness, anxiety and fear)
13. Memories are used to make Predictions and can
identify Prediction Errors to support new learning
14. What to Learn about? Working Memory and Attention
The ability to selectively process information (attention) and to retain information
in an accessible state (working memory) are critical aspects of our cognitive
capacities.
15. Reinforcement Learning
But knowing a lot about our environment is only one aspect of intelligence.
We also need to understand causality and the consequences of our actions
so that we can influence our environment
16. When the world changes: The ability to adapt
Reversal Learning Attentional Set-Shifting
Behavioral Flexibility & Control
17. Decision-Making
The ultimate result of the coordination of cognitive processing of our environment and our
actions is to make better decisions that facilitate survival, and allows us to thrive and adapt
to a changing environment.
The cognitive process resulting in the selection of a belief or a course of action among
several alternative possibilities.
Advanced Decision-Making requires response integration from a coherent cognitive system
The ability to make good decisions based on the current goals and context may
be regarded as one of the most important outputs of intelligence
18. Using Cognitive Neuroscience to create AI
• Intelligence is emergent from complex cognitive processes
• Process information/general mechanisms
• Learn about Observations or Sequences of Observations
• Store in knowledgebase for pattern-matching/pattern-separation and prediction error signals
• Goal-driven purpose (Motivation & Emotion, natural drives, survival and ability to thrive)
• Adaptive when the world changes
• Contextual awareness for relevance
• Causal learning & Explainability
• Common communication mechanism/language
• Decision-Making - Ability to integrate cognitive functions and select from multiple co-existing options
What is the core themes from biological intelligence?
19. Defining the Phases of Machine Intelligence
Second Wave - Machine Learning (human-
designed statistical models)
Third Wave – Machine Intelligence
(true AI; contextual adaptation)
First Wave - Expert Systems (rules
crafted by humans)
20. Our Reality: Possibilities and Probabilities
True intelligence is an emergent property in a
continuously changing world where reality is defined
by possibilities and probabilities
Our reality is a quantum-like state
We can use information to make predictions, decisions
and take actions in this complex state
21. Our framework is made to predict reality using
cognitive/information patterns
Information (definition):
What is conveyed or represented by a particular arrangement or sequence of
observations.
A mathematical quantity expressing the probability of occurrence of a particular
sequence of symbols, impulses, etc., as contrasted with that of alternative sequences.
Information Pattern (Cognitive Pattern):
A cognitive pattern is network of communicating and interrelated objects and classes
that can solve a problem in a particular context, such as a specialized node in the
brain.
22. A Cognitive Information-Centric Approach to AI
A proprietary quantum information technique is used to extract
actionable/useful information directly from raw data and to
create a knowledgebase of this information (long-term memory)
The system processes the information using principles
from neuroscience to discover information patterns, make
predictions, decisions and take actions
Our scientific discoveries have allowed us to create an information-centric framework where intelligence
emerges directly from data
27. Examples of custom AI topologies with abstractions and operators
Building a General Topology for Complex
Cognitive Patterns
28. Agents have a Genome
So they can be self-optimized, evolve and breed new generations
Classifiers - Continuous or Discrete Vectors (act in real-time and outperform NN)
Recall threshold - How Confident?
Max predictions - How Open-Minded?
Persistence - How Adaptable?
Atomic operators with genes themselves allow for flexible and adaptable universal functionality
29. How the framework works
The framework determines patterns in observations and sequences, and converts
them into relevant predictions & actions
Cognitive patterns are discrete/exact.
No modeling/curve-fitting to data
30. Ensembles of predictions - Quantum states of information
The [getPredictions] API call generates an ensemble of predictions with
the strength of each scenario from each cognitive processor
32. Information Analyzer
Level 1
(P1)
Level 2 (P4) - information gain
>10x better than P1 for identifying
negative outcomes
The Information Analyzer Quantifies and Qualifies the Information in the raw data, and
provides lists of key indicators and what data that does not convey signal
Abstraction
optimal
information
33. Example - Continuously Adapting Learning
Continuous learning and adaptation in a reinforcement learning and
reversal learning task when the correct response flips every 100 trials.
Training = 6 trials
Reversal 1 Reversal 2 Reversal 3
34. Example – Healthcare Applications
Solution is 35.5% better than
surgeon estimates
Predict surgery room scheduling for process
optimization in major health system
Diagnosis of echocardiogram
Solution was completely consistent with a trained
radiologist assessment, allowing for rapid and
automated preliminary screenings of exam data
35. Performance – Healthcare Applications
COGNITUUM: 99.65% accuracy and 100% precision classifying breast cancer using
the WBDC dataset (569 cases). No false answers.
BEST PUBLISHED: 99.51%, hybrid SVM with larger training set (Borges, 2015)
Classify breast cancer diagnosis using the
Wisconsin Breast Cancer Dataset
37. Our Vision - General Networked Intelligence
COGNITUUM
Machine IQ
Sensor
s IoT
Users Devices
External
App8
Users
Intelligent
Agents & Machines
Virtual
& Web
Universal Data Inputs
App6 App7
App4
Ap
Stand-alone and Networked Applications
App5
pp3
p2 A
App1
AI
Training
Vehicles
&Robots
Universal Data Outputs
39. Key Features of the COGNITUUM framework
A contextually aware Machine Intelligence framework that makes unbiased
predictions based on information from the environment in which it operates
40. Award-Winning Technology
"for their unique approach to an Artificial General Intelligence
framework that combines aspects from both quantum
information science and neuroscience”
2018 E-Challenge Winner
41. COGNITUUM Intelligent Agent Solution Stack
Data Inputs
Interoperable Data Intelligence Contextualization Layer
Interlocking Data Application Layer
COGNITUUM Cognitive AI
Operating System
Custom AIAgents
Measurement, Prediction & Attribution Layer
Integrated Presentation Layer
(Dashboards & Scorecards)
User Application Layer
(front end)
COGNITUUM Intelligence
Layer
Application Data Layer
(back end)
42. Our Domain Experience
Predictive Analytics
Cybersecurity
Internet of Things (IoT)
Robotics
Image analysis
Behavior predictions
Predictive Analytics
Process optimization
IT Intrusion Detection
Investment analysis
Internet of Things (IoT)
Autonomous robotics
Radiology
Sales optimization
Churn/Default Detection
Healthcare outcomes
Functional Domains Applications
43. Data 2 Data 3 Data 4 Data 5
Advanced Information Processing in Deep
Learning Hierarchies
Data 1