Top Banner
Hiroshi Yamakawa FUJITSU LABORATORIES LTD. JAPAN Brain-inspired equivalence structure (ES) extraction technique for generating frames
14

Brain-inspired equivalence structure extraction technique for generating frames

May 10, 2015

Download

Technology

This is presentation in the nanosymposium of the Society for Neuro Science, in 2013 at San Diego
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
  • 1.Brain-inspired equivalence structure (ES) extraction technique for generating frames Hiroshi Yamakawa FUJITSU LABORATORIES LTD. JAPAN

2. Outline 1. Human-level intelligence can explore from neocortex learning. Artificial intelligence (AI) lacks flexible sampling function of neocortex Equivalence structure (ES) extraction is key for such a function2. Use local sequences to extract equivalence structures (ESs). Inspired by theta phase precession of hippocampal formation3. Simple simulation of ES extraction 4. Summary Frame generation: Promising way to achieve artificial general intelligence (AGI)2 3. HLDL mostly means human-level artificial intelligence (HLAI)because untrodden machine intelligences D are concentrated on neocortex. Complete intelligenceDeep learning (DL) High-performance unsupervised machine learning technology corresponding to neocortexFeasible intelligence (with limited resources) Human intelligenceArtificial intelligenceControl theory (Cerebellum)DEfficient arithmetic operation and logic inferenceDeep C r e a t i v i t ylearningHuman-level DL (HLDL) Fully simulate neocortex computing and its learning functions. (with help of hippocampus).(DL) I n t u i t P a tn t e r n i o r e c o g n i t i o n Ge n e r a l i n t e l l i g e n c eRetrieval from big dataReinforcement learning Emotion (Basal ganglia) (Amygdala)What is the problem in achieving HLDL? 4. Deep learning lacks flexible sampling Example: Intuitive decision making for chess-like game High-level featuresPrecuneus Caudate Quick generation of best next-moveHippocampusPerception of Supports intuitive decision making board pattern - Cannot be explained by expertsSupport learning of neocortex- Cannot be acquired by deep learningEyesupport learningVisual cortexConvolution SamplingV3Convolution SamplingV2HippocampusMTG /V6Convolution SamplingConvolution layer Well-developed for machine learning: Simple cell, Auto encoder network, SOM, Boltzmann machine, Info-MAX, Manifold learning, ...Sampling/Pooling layer: Human encode structure of hierarchical retinotopy: Complex cell, Max-pooling, ...Chess-like game (Wan,V1Convolution SamplingScience 2011)4 Supports visual invariances Need flexible sampling 5. Equivalence structures for flexible sampling Equivalence structure (ES) indicates portions of subspace that could be regarded as equivalent.Variable set: xOriginal frameA B C D E F G HSubspaceSubspaceInput sequenceX Y ZA B CD E F 1 2 3 4 5 6 71 2 3 4 5 6 7TimeTimeCombined 1 2 3 4 5 6 7 frameInvariance Increased events enhance deductive inference.Time: tInvariance in basic image processingInvariance for face recognitionNeed more flexibility for higher-level sampling5 6. Outline 1. Human-level intelligence can explore from neocortex learning. Artificial intelligence (AI) lacks flexible sampling function of neocortex Equivalence structure (ES) extraction is key for such a function2. Use local sequences to extract equivalence structures (ESs). Inspired by theta phase precession of hippocampal formation3. Simple simulation of ES extraction 4. Summary Frame generation: Promising way to achieve artificial general intelligence (AGI)6 7. Static patterns are poor for ES extraction this could exist in neocortex.If using common static binary patterns as similarity to compare subspaces, Input sequenceSet of N variablesProblem: Variation in static patterns 2d is not enough to categorize thousands of subspaces NCd(~Nd).A B C D E F G HESsX Y Z1 2 3 4 5 6 7 Original frameTime: tSubspace of d variables A B CCombined frameD E FToo many other subspacesNeeds similarity with rich variation. 7 8. Subspaces can be compared using local sequences Theta phase precession Several sequential events are packed in each phase (~5 Hz) Inspired by information representation in hippocampus. ( Sato and Yamaguchi : Neural Computation 2003)Subspace of d variablesSet of N variablesESs Input sequence A B C D E F G HX Y ZD E F 1 2 3 4 5 6 71 2 3 4 5 6 7TimeTimeCombined frame 1 2 3 4 5 6 7Original frameA B CTime: tLocal sequences are used to compare subspaces. (Skipping a detailed explanation.) 8 9. Outline 1. Human-level intelligence can explore from neocortex learning. Artificial intelligence (AI) lacks flexible sampling function of neocortex Equivalence structure (ES) extraction is key for such a function2. Use local sequences to extract equivalence structures (ESs). Inspired by theta phase precession of hippocampal formation3. Simple simulation of ES extraction 4. Summary Frame generation: Promising way to achieve artificial general intelligence (AGI)9 10. Simple simulation to validate this idea Input image for experiment: Dot wave sequence A swinging dot image in sequence of one-dimensional spaces, representing an idealized video image of natural scenes (up to 300 frames) Ti e m di .I m D1 A 1 B 2 C 3 D 4 E 5 F 6 G 7 H 80 0 0 1 0 0 0 020 0 1 0 0 0 0 030 0 0 1 0 0 0 040 0 1 0 0 0 0 050 0 0 1 0 0 0 060 0 0 0 1 0 0 070 0 0 0 0 1 0 080 0 0 0 0 0 1 090 0 0 0 0 0 0 110111213141516171819200 0 0 0 0 0 0 00 0 0 0 0 0 0 10 0 0 0 0 0 0 00 0 0 0 0 0 0 10 0 0 0 0 0 1 00 0 0 0 0 1 0 00 0 0 0 1 0 0 00 0 0 1 0 0 0 00 0 1 0 0 0 0 00 0 0 1 0 0 0 00 0 1 0 0 0 0 0300 Expected ES: A cluster of adjacent variable sets Cluster of 12 subspaces, each of which consisting of 3 adjacent variables, is expected to be extracted depending on spatial continuity of input sequence. ABCDEFCDEFGHXBCDEFGBCDEFGYCDEFGHABCDEFZCluster of subspaces 10Combined frame 11. Result: Expected ES is extracted as a clusterSubspacesAll permutation of subspaces (366 patterns)Expected ES containing adjacent variable set is extracted as cluster from a numbers of local sequences clustering.Index of local sequences (Only non-zero elements shown)11Combined frame X Y Z E D C B C D G F E D E F F E D E F G D C B C D E C B A A B C H G F F G H 12. Outline 1. Human-level intelligence can explore from neocortex learning. Artificial intelligence (AI) lacks flexible sampling function of neocortex Equivalence structure (ES) extraction is key for such a function2. Use local sequences to extract equivalence structures (ESs). Inspired by theta phase precession of hippocampal formation3. Simple simulation of ES extraction 4. Summary Frame generation: Promising way to achieve artificial general intelligence (AGI)12 13. Summary Untrodden machine intelligent functions are concentrated on neocortex, so emergence of HLDL mostly means emergence of HLAI. Learning of sampling layer is minimally needed to generate high-level features for HLDL. This learning is assumed to be equivalence structure extraction. Inspired by theta phase precession, I introduced numbers of local sequences for each subspace. Clustering of subspaces by these frequencies enabled extraction of ESs in a simple demonstration. I'd like to specify the sub-region of the hippocampal formation within theta loops that perform ES extraction. Future work includes constructing a neocortex-hippocampus model implementing ES extraction. 13Where is responsible sub-region for ES extraction in theta loop of hippocampal formation?(Buzsaki, 2007) 14. Human-level general AI needs ability to generate frames. General intelligence systems should be able to learn to solve problems that were unknown at time of their creation. Obviously, human brain can generate new frames to solve various new problems using learning ability of neocortex. Neuron Column EventsA 1 2 3 4 5 6Variables B C DECombined new frameEquivalence structure (ES)Values frameDesigning HLDL by referring to neocortex is a promising approach. 14