Mid-January update Kalman Filter implementation & - PowerPoint PPT Presentation

01/15/16 Mid-January update Kalman Filter implementation & reconstruction debugging Rebecca Carney Student Instrumentation Weekly 1

Kalman Filter in spiking NN - update RMD Carney 01/15/16 TrueNorth ‘neuromorphic’ chip TrueNorth chip: brief brief intro 64 x 64 neurosynaptic cores in 28nm 1 kHz effective clock (called a tick, or synchronization point) Tiled, distributed computing and memory Fully connected Best for neural-network-like applications 2

Kalman Filter in spiking NN - update RMD Carney 01/15/16 From neuroscience terminology… Neurons, cores, crossbars NB these aren’t gates, it’s neuroscience symbols for an axon …to inter-core connectivity & recursion! …to a splitter.. …to a crossbar… 3

Kalman Filter in spiking NN - update Kalman Filter in spiking NN - update RMD Carney 01/15/16 Learned connection Binary connections in chip Assigned strengths probabilities (weights) based on probability from training Weights in training In software Model file on chip Synapse ‘strength’ (1 or -1) is assigned before training. Usually random or with other initialization patterns (not important for today). During network training a ‘backward’ traversal through the network updates the probability of a connection being useful. In the final step, a stochastic sampling decides which weights are useful and only those connections are carried through in the model file. 4


 Kalman Filter in spiking NN - update RMD Carney 01/15/16 http://ieeexplore.ieee.org/stamp/stamp.jsp? Kalman filters appear to eliminate the need for explicit offline training. arnumber=737502 Rather, the process of passing data Weight updates online through the data updates the various parameters of the process and measurement equations. Kalman Filtering and Neural Networks by Simon Haykin However, one of these update steps appears to change the weights of the neurons. 
 There is no way to update the ‘weights’ of a neuron online in TrueNorth. 
 So my question is: how were you planning to get around this? 
 Is there a different KF implementation that does not require this? 5

Kalman Filter in spiking NN - update RMD Carney 01/15/16 http://ieeexplore.ieee.org/xpls/abs_all.jsp? Reduced version of the algorithm arnumber=5910570&tag=1 Spiking Neural Network Decoder for Brain-Machine Interfaces by Julie Dethier, et. al. Plan for the week: implement this version of the KF in software 6

Offline reconstruction: bug (?) RMD Carney 01/15/16 RDO and transient/persistent conversion Generation A cluster HepMC Simulation G4 Hits Individual pixel hits Digitization Where has the RDO pixel information gone? Atlfast G4 Digits Reconstruction Real Data ESD tot = (*pix_itr)->auxdataConst< std::vector< int > >(" rdo_tot "); if (tot.empty()){ Create AOD std::cout<<" ToT vector is empty!" <<std::endl; exit( 1 ); AOD } Analysis /afs/cern.ch/work/r/rcarney/private/rc_analysis/run/mySample_edited.InDetDxAOD.pool.root ===========================Event #0/5000===================== 7 ToT vector is empty!