Artificial Intelligence [in Cognitive Neuroscience]

Artificial intelligence definition, techniques, and applications for cognitive neuroscience students (June 14, 2016).

Transcript

1. Cognitive Neuroscience Artificial Intelligence Morteza Ansarinia Institute for Cognitive Science Studies June 2016
2. Harold Cohen (1928-2016) AARON (1973-2016) Marvin Minsky (1927-2016)
3. Outline ✦ Modern History + Alternative History ✦ A Working Definition ✦ Knowledge Representation (Semantic Networks and Ontologies) ✦ Uncertainty and Reasoning • Bayesian Inference • Fuzzy Systems ✦ Machine Learning - Reinforcement Learning - ANN, SVM and PCA “AI attempts not just to understand but also to build intelligent entities.” — Russell and Norvig (2010) ✦ Evolutionary Algorithms and Emergent Intelligence ✦ Artificial Life ✦ Agent_Zero ✦ Future of AI ✦ Search, Planning, and
 Constraint Satisfaction Problems ✦ Chaos Theory and Fractals ✦ Expert Systems ✦ Rough Sets ✦ Natural Language Processing
4. Topics Authors 4 aitopics.org (2016)
5. Modern History of AI 5 The Dartmouth workshop marks the birth of AI. 1956 Rosenblatt’s perceptron, a two layer binary classifier. 1958 Franklin (2014) GPS demonstration by Newell, Shaw, and Simon. DENDRAL, and expert in organic chemistry to identify the molecular structure. 1967 Minsky & Papert published Perceptrons on limits of simple ANNSs. 1968 MYCIN, a rule-based system by KR and inference in medical diagnosis and therapy (First expert system). 1974 Cohen demonstrated AARON at the AAAI. 1978 Brooks’ Cog project, major advances in machine learning, NLP, vision, games, and translation. 1990s Commercial AI and robots, sociable machines (KISMET), autonomous vehicle, and semantic web. 2000s Cognitive architectures and self-driving cars. Today
6. Alternative History
 Why information processing and computers? 6 Preserved in religious texts, humans were formed from clay or dirt, which an intelligent god then infused with its spirit. That spirit explained our intelligence. … Zarkadakis (2016) Invention of hydraulic engineering. 300 BCE Hydraulic model of intelligence; The flow of different fluids in the body (cardinal humors) accounted for our physical and mental functioning 300 BCE, for 1600 years Automata powered by springs and gears had been devised. 1500s Humans are complex machines (Leibniz, Descartes). Thinking arose from small mechanical motions in the brain. 1600s Discoveries about electricity, chemistry, and communications. 1700s New analogies regarding human intelligence. Helmholtz compared the brain to a telegraph. 1700s and 1800s Dawn of computer technologies and information processing. 1940s Epstein (2016) Neumann and Miller stated the function of the nervous system is algorithmic and prima facie digital. 1958
7. A Working Definition ✦ A working definition of a term, like intelligence, is a definition to be used as the goal of a research field and related research projects. ✦ Human beings differ from animals and machines significantly in their mental ability, which is commonly called intelligence, and AI is the attempt to reproduce this ability in computer systems. - Not identical but similar. - Not a duplication of human body, but only intelligence. ✦ Similarity between C and H, when
 they are described at a certain level
 of abstraction. Wang (2008) Human Agent (t is time) Percepts Actions Internal States Machine Agent 7
8. A Working Definition Evaluating Similarity Cognitive functions similar to those observed in human FUNCTION AI Function Similar normative behaviors PRINCIPLE AI Principle Similar to the human brain STRUCTURAL AI Structure Similar to the human mind BEHAVIORAL AI Behavior Human capability of practical problem solving CAPABILITY AI Capability Intelligence is the capacity to adapt under insufficient knowledge and resources (Wang, 2009). 8
9. A Working Definition Evaluating Similarity 1. (of 5) By Structure - Best-known intelligence is produced by the human brain. - AI can be achieved by emulating a brain-like structure — massive neuron-like processing units working in parallel. - “The ultimate goals of AI and neuroscience are quite similar.” - Examples: Artificial Neural Networks, EU Human Brain Project, OpenWorm, Blue Brain, CLA, HMAX
 2. By Behavior - Intelligence seems to be more about the human mind than the brain. - It requires behavior similarity between AI system and human mind. - Both systems are treated as black box. - Examples: Turing Test, Soar, ACT-R, Loebner Prize Competition (chatbots) Wang (2008) “Brains cause mind.” — Biological Naturalism (Searle, 1992; Searle, 2002) 9
10. A Working Definition Evaluating Similarity 3. By Capability (e.g. Deep Blue, IBM Watson) - Intelligent people are the ones that can solve hard problems. - AI should aim at expert systems in various
 domains of application. 4. By Function (e.g. IBM Watson) - Intelligence depends on cognitive functions like reasoning, learning, and problem solving. - AI should aim at cognitive architectures with these functions. 5. By Principle (e.g. NARS) - Human mind follows normative principles that are not followed by computers. - AI should aim at the specification and implementation of these principles (bounded rationality and optimality). Wang (2008) “The only way … is to be the machine and to feel oneself thinking!” — Alan Turing 10
11. DIKW Hierarchy 11 Context Undestanding Data Information What? Experience Knowledge How to? Why? Consumer Producer Cognition What to do? Context
12. Knowledge and Reasoning Logic, Inference, and Representation ✦ The primary goal of knowledge representation is to allow an agent to make intelligent decisions about its environment. ✦ Semantic Network: useful way to describe relationships between a number of objects. It contains facts and relationships between that knowledge to link them (typically IS_A and A_KIND_OF). ✦ Frames: Semantic Network + Generic Classes + Instances + Triggers - Scripts: type of frame that is used to describe a timeline. ✦ Propositional logic (propositions and connectives) and first order logic (variables, constants, functions, compound sentences, and quantifiers). ✦ Semantic Web: an effort to change the way that the web is defined and interpreted by adding semantic labels (in RDF or OWL formats). ✦ Ontologies: Formally defined as specification of conceptualization, an ontology is a model that represents a set of concepts within a specific domain as well as relationships between those concepts. 12 Russell & Norvig (2010)
13. Knowledge and Reasoning Examples 13 Washington & Lewis (2008) Semantic Network Ontology
14. Bayesian Statistics ✦ Bayesian Statistics is a system to describe epistemological uncertainty using probability. ✦ Bayesian methods start with existing prior beliefs, and update these using observed data to give posterior beliefs, which may be used as the basis for inferential decisions. ✦ When to use? - Lack of data on some other aspects of the model, - multiple source of evidences with some shared distribution, - and where a huge joint probability model is already constructed. 14
15. Naive Bayes Classifier ✦ Find out the probability of the previously unseen instance belonging to each class, then simply pick the most probable class (with naive assumption of independence between every pair of features). ✦ Advantages: - Fast to train (single scan), fast to classify, - not sensitive to irrelevant features, - handles real and discrete data, - and handles streaming data well. ✦ Disadvantage: - Assumes independence of features, - bad estimator. 16 New Data Activation Level Brain Region (PC1) ?
16. Bayesian Statistics ✦ Bayesian Belief Network is a probabilistic graphical model, which can simultaneously represent a multitude of relationships between variables in a system. - Nodes represents variables, and arcs represents conditional relationships. 15 Mihaljević et al. (2014)
17. Fuzzy Systems ✦ Fuzzy Logic: A logical system to formalize human capability to reason and make decision in uncertainty, and imprecision of information, and partiality of knowledge and class membership.
 
 - Logical FL: may be viewed as a generalization of multivalued logic. - Relational FL: if X is A then Y is B (A and B are fuzzy sets carrying linguistic labels like small, medium, and large). - Membership in a fuzzy set is a matter of degree. 17 Zadeh (2008) a A Universe of Discourse Singular Value of X Granular Value of X Conditional Possibility
18. Machine Learning ✦ Machine learning is getting machines to act without begin explicitly programmed. It focuses on known properties learned from training. In three categories based on input type: - Supervised: presented with example inputs and desired outputs by a teacher. Agents generalize mappings and rules (e.g. regression and classification). - Unsupervised: Find hidden structures and patterns of the inputs
 (e.g. clustering). 18 K-Means EM DBSCAN
19. Machine Learning Reinforcement Learning ✦ Learning by interacting with an environment (rewards and punishment), or trial and error learning. ✦ The agent learns from the consequences of its actions, rather than from being explicitly taught. The agent selects its actions on basis of its past experiences and also by new choices. ✦ Time Difference Learning (TD Learning): 19 Ertel (2011) Update to Step Size (often =1) Expected Return Discount Factor State Reward
20. Machine Learning Reinforcement Learning 20 STDP Woergoetter and Porr (2007)
21. Artificial Neural Networks ✦ An algorithm inspired by structure and functional aspects of biological neural networks. ✦ Presented as an interconnected group of neurons, estimate or approximate function is stored as numeric weights of connections. ✦ Weights can be tuned based on experience, making neural nets adaptive to inputs and capable of learning. ✦ Pitfalls: ANNs cannot deal with missing data, and are not easy to use and understand. ✦ Types: Feedforward, Recurrent, Hopfield, Kohonon Self-Organizing Maps, Probabilistic, Backpropagation, Hierarchical Temporal Memory, Deep Belief Networks, Spiking, … 21 XOR SOM HTM
22. Computer Modeling of
 Hand-Centered Visual Representation ✦ Galeazzi el al. (2015) investigated how hand-centered visual representations may develop in a neural network model of the primate visual system (VisNet), when the model is trained on images of the hand seen against natural visual scenes. Such neurons may develop through a biologically plausible process of unsupervised competitive learning and self-organization. ✦ The synaptic connections between the
 successive layers of neurons are updated
 using associative learning (trace learning). ✦ Neurons develop responses that are
 selective for the location of visual targets
 with respect to the hand but invariant to
 the position of the hand-object
 configuration on the retina. 22 Galeazzi et al. (2015)
23. Computer Modeling of
 Hand-Centered Visual Representation 23 Galeazzi et al. (2015)
24. ✦ 2-class classifier, maps inputs to either -1 or +1 classes. ✦ Optimal hyperplane linearly separates two classes of objects. ✦ Does not suffer from curse of dimensionality and local minima. ✦ Non-linear separation using kernel functions and transformation to VC dimension. Support Vector Machine 24 Adankon & Cheriet (2015)
25. Principle Component Analysis ✦ PCA is a technique used to emphasize variation, bring out strong patterns in a dataset, and eliminate a dimension (reduction). 25
26. Artificial Life ✦ Hard artificial life produces actual physical hardware that acts autonomously in the physical world (e.g. Brooks and Lipston’s works, Arcana, Baxter). - Avoids the need for an elaborate and detailed internal
 representation of the external environment, - contrasts with traditional robotics by explicitly and
 extensively exploiting inspiration from all forms of life
 including those that are much simpler than humans, - and the structure of the control system is tightly coupled
 to the agent’s morphology. 26 Frankish & Ramsey (2014) Cheney et al. (2014) Lohmann et al. (2012)
27. Artificial Life ✦ Wet artificial life creates new forms of life in test tubes, using the latest materials and methods from biochemistry and molecular biology. - Create artificial cells out of biochemical raw materials, such as lipids, or DNA and RNA molecules, that are not alive. - Main functionalities of artificial cell are to move through fluids and process chemicals, self-maintenance, autonomous control of internal processes, control of mobility, and to reproduce. - Top-down strategy: synthesize and then redesign the genome of existing simple life forms such as bacteria. - Bottom-up strategy: start from non-living materials and build more and more complex physiochemical systems with more and more life-like properties
 (e.g. vesicles and cell-free-extract method). ✦ Soft artificial life is computer simulations or other purely digital constructions that exhibit life-like behavior (e.g. Tierra, Conway’s Game of Life, and Wolfram’s NKS Cellular Automaton, Sugarspace, and Agent_Zero). 27 Frankish & Ramsey (2014)
28. Artificial Life Cellular Automata Examples 28 Rule Thirty Wolfram’s CA Conway’s Game of Life
29. Evolutionary Algorithms ✦ Guided stochastic search and optimization heuristics derived from the classic evolution theory. They always provide an answer, work well in noisy environment, and essentially distributed. ✦ Simple Genetic Algorithm (GA): 29 “Everything must be made as simple as possible, but not one bit simpler” — Einstein (1962) initialize population. evaluate population (calculate fitness). while (Termination Criteria Is Not Satisfied) select parents for reproduction. perform recombination (crossover) and mutation. evaluate population (calculate fitness).
30. "The consequences of machines thinking would be too dreadful. Let us hope and believe that they cannot do so.” — Turing’s answer to the “Heads in the Sand” objection Agent_Zero ✦ Grow Raskolnikov: agents with more fully developed yet conflicted inner lives. ✦ synthesis of three processes: ✦ Emotional: Pavlov classic conditioning as formulated by Rescorla-Wagner model of conditioning. ✦ Rational (Cognitive): Agents form a probability, based on local sampling. ✦ Social: With dispositional contagion, it is not behavior that is contagious in society, but (solo) dispositions. ✦ Agent acts if its total disposition (D) exceeds action threshold. 30 Epstein (2014)
31. Future of AI A Survey of Expert Opinions on Human-level AI Müller and Bostrom (2016) Cognitive Science Integrated Cognitive Architectures New Computational Neuroscience Algorithms Artificial Neural Networks Faster Hardwares Large-scale Datasets Embodied Systems New Unknown Methods Whole Brain Emulation Evolutionary Systems Logic-based Systems Algorithm Complexity Theory No Method Swarm Intelligence Robotics Bayesian Networks 0% 10% 20% 30% 40% 50% 31
32. References Adankon, M. M., & Cheriet, M. (2015). Support vector machine. Encyclopedia of biometrics, 1504-1511. Black, N. T., & Ertel, W. (2011). Introduction to artificial intelligence. Springer Science & Business Media. Dezfouli, A., Piray, P., Keramati, M. M., Ekhtiari, H., Lucas, C., & Mokri, A. (2009). A neurocomputational model for cocaine addiction. Neural computation, 21(10), 2869-2893. Epstein, J. M. (2014). Agent_Zero: Toward Neurocognitive Foundations for Generative Social Science. Princeton University Press. Frankish, K., & Ramsey, W. M. (Eds.). (2014). The Cambridge Handbook of Artificial Intelligence. Cambridge University Press. Galeazzi, J. M., Minini, L., & Stringer, S. M. (2015). The Development of Hand-Centered Visual Representations in the Primate Brain: A Computer Modeling Study Using Natural Visual Scenes. Frontiers in computational neuroscience, 9. Mihaljevic, B., Bielza, C., Benavides-Piccione, R., DeFelipe, J., & Larrañaga, P. (2014). Multi-dimensional classification of GABAergic interneurons with Bayesian network-modeled label uncertainty. Frontiers in computational neuroscience, 8, 150. Müller, V. C. and Bostrom, N. (2016). Future Progress in Artificial Intelligence: A Survey of Expert Opinion. Fundamental Issues of Artificial Intelligence, 553-571. Wang, P. (2008). What Do You Mean by “AI”?. Artificial General Intelligence, 362-373. Washington, N., & Lewis, S. (2008). Ontologies: Scientific data sharing made easy. Nature Education, 1(3), 5. Woergoetter, F., & Porr, B. (2008). Reinforcement learning. Scholarpedia, 3(3), 1448. Zadeh, L. A. (2008). Fuzzy logic. Scholarpedia, 3(3), 1766. Zarkadakis, G. (2016). In Our Own Image: Savior or Destroyer? The History and Future of Artificial Intelligence. Pegasus Books. 32
34. Thank you :-) Dezfouli et al. (2009)