We are changing the way the world computes - PDF document

��We are changing the way the world computes 1 The NeuroMem computing alternativeThe NeuroMem conceptToday’s common platforms A multicore processor surrounded by DMA and SDRAM controllers; �1 GHz, �10 WNeuromorphic memory with 1024 identical cells; 16 Mhz, 0.5 W Memory and processing logic combined in a same cellIdentical cells working in parallelFixed number of I/Os independent of the number of cellsZero Instruction Set Computing architectureless than 1 M$ development cost 130 nm 8 Metal layers 8x8 mm dies sizeMulticore processorsMemory misery bottleneckLimited by the Amdahl’s lawHigh frequencies (GHz), Power demanding (10W)Complex programming��2&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes Broadcast Mode : query/stimulus is broadcasted to all the neurons of a “context group” simultaneouslyDeterministic search time: does not increase with the scaling up of the network Winner takes all: Inhibit the weak responders autonomously in the same deterministic timeUncertain response: when there is multiple/conflicting neurons responding or “lesser quality” (degenerated) neurons spiking.Unknown response: Enable the dynamic addition of new knowledgeBack propagation of error: Self & parallel inhibition of erroneous spiking neurons, no software involvedNo fetch and decode of program instruction: Software is definitively contrary to the biological model, else it’s simulation, not neuromorphic…Beyond biology: Fast upload download enabling knowledge proliferation (some dream of itThe 8 pillars of euromorphic Pantheon ��3&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes The merger of two old, but still infashion concepts Nonlinear classifier , Compound classifier invented by Bruce Batchelor Hardwired parallel architectureCERN’s UA1 experiment lead by Nobel prize winner Carlo RubbiaGuy Paillet DataSud Systems design a parallel architecture (60 CPU’s/Memory on same VME bus)Big Data before time250 Gigabytes/second ZISC (Zero Instruction Set Computer with 36 and later 78 neurons) designed by IBM France and Guy Paillet CM1K (Cognitive Memory with 1024 neurons) designed by Anne Menendez and Guy Paillet DARPA Neural Network Study“The technology is not mature for widespread practical applications, since computer simulation are the primary methods of implementation.”Restricted Coulomb Energy classifier, derived by Leon Cooper (Nobel prize for supraconductivity ��4&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes The NeuroMem InnovationReactive Memories: Deliver answers voluntarily when an input data matches their content. Instead of having a CPU screen sequentially what every memory ‘knows’, the memories volunteer their response in an orderly manner.Associative MemoriesReact not only when an input data matches their content, but also when it is similar “enough”. This feature extends the usage model to a non linear classifier.TrainableemoriesKnow when an input data represents novelty. Can learn/add new models in realtime. If applicable, the memory cells recognizing the new example erroneously correct their influence field autonomously . This feature extends the usage model to a trainable neural network.Massively parallelemoriesImplement a natively parallel architecture which allows the sizing of any bank of cognitive memories WITHOUT impacting the I/O counts nor the existing interfaces to external hosts. The memories connected in parallel are operated at low frequency and therefore require lowpower. Reactive Memories Associative Memories Trainable Memories Massively parallel Memories ��5&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes A stack of 100,000 neurons fully operational XP2LATTICEbuffering CM1K MRAM CM1K CM1K CM1K XP2LATTICEbuffering CM1K MRAM CM1K CM1K CM1K “Spine” UP TO 25 NeuroMem Boards409,600 operations per clock cycle @ 10 MHzTotal ~4 TeraOps/sec 25 WattsNEXT STEPReduce to a single chip in 2014��6&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes 1002003004005006000.20.40.60.81.2WattsNeuronsMillions Demonstrated high Scalability 4096 neuronsWatts 1 Million neurons250 Watts (in progress) 1024 neurons0.5 Watts 40960 neurons10 Watts 10 µsec Recognition latency��7&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes CM1K neuron architecture? A CM1K neuron react (spike) which when incoming pattern is “similar” to its “learned” a memory. Similarity domain is adapted by excitation/inhibition during the teaching process. CM1K neurons are also able to make unsupervised leaning (e.g. clustering, etc…) Category Influence Field * Distance * = Active Context StimulusOr query“Fires” Context Model * Read Only (updated by the neuron itself) Identifier * ��8&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes The CM1K neural networkAll neurons process the input vector in parallelBest match of all neurons, clock cycles later 1 2 3 4 n inputsbit262,144Synapses/chip(256 * 1024) 32,766outputsn=1024 hidden nodesCM1K seen as a 3layer network 1 2 3 n A neural network is a bank of neurons operating in parallel and interacting together to learn autonomously and recognize so that the winnertakesallAll neurons have the same behavior and execute the instructions in parallel with no need for a controller or supervisor Functions performed : IdentificationClassificationClusteringAnomaly DetectionNovelty Detection��9&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes From Data to Insight… A signature or feature vectoris extracted from a data source to model a given context of the data The vectoris broadcasted to all the neuronsin parallel The neuronsknowledgeable about the selected context evaluate their similarity to the vector The “firing” neuronsautonomously queue what they know about the vector in decreasing order of similarity Their responses can be read AS IS or formatted into a new vectorassigned to a different context Text, DNA sequencesAudio, Voice, sensor signalImages Videos ��10&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes FE FE FE Images Videos Voice Signal Data Classification Anomalydetection Noveltdetection Clustering Template matching Tracking Identification Vectors Neurons’ response Feature Extraction Usage modelUsage Models, Level1 NeuroMem ��11&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes Images Videos Voice Signal Data Usage Models, Level2NeuroMem Images Videos Voice Signal Data Data PresentationData Transformation /Reconstruction Combinatorial and/or hierarchical feedback loop��12&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes Ex1: Hierarchical decision making Objects identification and generation of an output list of categoriesData mining with Contextual localization Heron or Stork beakHeron eyeHeron Aigrette neck Transform imageSource imageRecognized objects with categoriesFinal identification Heron Knowledge#1Knowledge#2Knowledge#3 Correction of primitive blocks and generation of a transform image with enhanced blocks��13&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes ApplicationsA companion processor to offload the Van Neumann machines from pattern recognition tasksCognitive SensingConvert sensor signals into insight data at the source and in realtime, selective transmissionCognitive storageIn storage search engines , scalable content access memory banks, intrinsic deduplication, no need for indexingCognitive networkingMatch content of high speed data stream in realtime, selective pull data forward or exclusion��14&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes « DEVELOPMENT OF A MINIATURISED WAVEFRONT SENSOR BASED ON A SILICON NEURAL NETWORK »Engineering Diploma of the ENSPSMaster IRIV Nanophotonics SupervisorsMarc EichhornAlexander Pichler Thibaud MAGOUROUXUniversity year 2010 01/03/2010 31/08/2010 ��15&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes Already on the market ��16&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes Artificial Intelligence and true parallel computing embedded system Big Artificial Brainfrom the computer to the ‘brainputer’ Biometric DataDNA fingerprint iris voice Identification Goal : 10decision cells in parallel for ultra fast data base analysis by pattern matching ��17&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes Droppable and disposable Miniature Visual Event Detector (in progress) ��18&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes Near Sensor attern ecognition ��19&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;&#x/Att;¬he; [/;ott;&#xom ];&#x/BBo;&#xx [2;7.1;# 1;.37;) 4;….9;б ;3.5; ];&#x/Typ; /P; gin; tio;&#xn 00;We are changing the way the world computes