Skip to content

Latest commit

 

History

History
257 lines (206 loc) · 9.41 KB

tsad.org

File metadata and controls

257 lines (206 loc) · 9.41 KB

Research

anomaly

does depth help?

how to define

RNN have promise for ts b/c memory

RNN-RBM Hessian free opt?

rnn or cRBM

what input causes the hidden unit i to be maximally active?

\[ a_i = f\left(∑\limits_j ωij(1)x_j + b_j \right) \] constraint $||x||_2=1||$ \[ x_j = \frac{ ωij(1) } { \sqrt{∑\limits_j\left(ωij(1)\right)^2} } \]

IDEA!: use RNN to guide finding covariance matrix for a GP!

Resources

mendely

Coursera neural net

Experiment

  • nhidden
  • types of ts: frequency-rich, stnarity
  • can it be insensitive to the window?, robustness in general

data sources

  • sleep stage classification
  • astro
  • ted dunniang ekg

code

pyleanr2

  • smaller data was faster on cpu. gpu shuffling?
  • the yaml thing makes it complicated for a beginner. it’s supposed to make your life easier but it’s another thing to learn it’s weird

theano rnn

Algorithm:

need to have invariance wrt time

challenges:

training

contrastive gradient pretraining

choosing the right

if deep we have memory but deep diminishing tangents

hessian free

optimization

ilays sutskever init with echo state nets then train rnn

invariance: 1. scaling: rectifed lin units 2. shift: conv nets.

coursera lec 15 “modeling real-valued data with rbm”

must be as auto as possible. least amount of fidddling

id and attacking the saddle pt prblem in high dim non conv opt

http://arxiv.org/abs/1406.2572

thesis

  • thanks ted dunning, kirk borne, jay kaufold
  • we want to be more general with RNN as opposed to HMM, lin dynamics sys

theano-nets

  • learn about batches
  • can train multiple len sequences?
  • how does it go though training

So Network obj is a base obj.: regressor, autoenc, b/c it has a cost fx

  • see if it uses the latest hf optimizer

only run setup_enc/decoder once

  • check if dataset can do varying length sequences
  • waht does train_batches do?? it’s added in seqdataset
  • i want to predict on just one sequence..see how i can cpy params from trained mdl
  • trn and val gen ds still not working
  • why does the ds beak the time axis when using callable()?
  • the callable is exhausted..what happens?
  • how would classifying work? bc it’s seq->(one cls)
  • also chk the log out of climate. it says logistic but i’m not sure
  • can mod the code to make a ‘monitor’ as a separate proc?

Bayesian Optimization

does hyperopt handle stochastic??

GP

coming fron sci and eng . tricky thnx engstats.com. however it took my analytical function oriented mind to get around appealing to stats for “functions”.

  • REALLY understand joint, conditional, and marginal distributions in

this more abstranct sense how to learn on your own. ive been out of classes for a while:

  1. find intro papers, and videos
  2. build a toy. use alpha blending

[conditionals ovhttp://stats.stackexchange.com/questions/30588/deriving-the-conditional-distributions-of-a-multivariate-normal-distribution REALLY understand the differences b/w marginals, conditionals, and joints., and EV in this more abstract context gptools adds noise?

Inbox

need to consider temporal aspect. the literature if filled with techniques suitable for specific application domains! want to bring review neural net literature anomaly detection

http://www.idsia.ch/~juergen/rnn.html Hyper-opt for a ‘db’ of priors hierarchy AND recurrent aspects HF opt http://pillowlab.wordpress.com/2013/06/11/lab-meeting-6102013-hessian-free-optimization/ great ideas for data examples in anomaly detection of ts simple explanation of rnn http://www.willamette.edu/~gorr/classes/cs449/rnn1.html explain rnn vidoedeo http://techtalks.tv/talks/on-the-difficulty-of-training-recurrent-neural-networks/58134/ basins of attraction lstm tutorial http://techtalks.tv/talks/on-the-difficulty-of-training-recurrent-neural-networks/58134/ exploding and vanishing gradients problem for rnn..rnn hard for long term dependency

4 ways to to train RNN: back prop thru time your weights blow up or diminish: has local min problem

  • long short term memory LSTM
  • Hessian free opt: can deal with small grads
  • Echo state net
  • good init with momentum

comparison in hf paper.

talk about hierarchy of ts in training and analysis don’t care about interpretability that much but it may be incl for free. need to ‘model’ ts.. what if we don’t want human involvement to model? comparison with other techniques: in review paper most reserach on training (they are difficult) instead of analysis now go through different

specify a novelty detection mathematical context

Recurrent Neural Networks http://www.cs.bham.ac.uk/~jxb/INC/l12.pdf

better predictor makes a better novelty detector? can you use a nn to understand the data gen proc?

what?:

  • architechture
  • training:

cool thing about ts is easy to viz. ‘watch’ the predicted as it is training by superimposing. maybe we don’t need to go to the most min possible can a neural net show regularity progressively?

RNN display and memorize temporal nature

future work:

  • multivar ts
  • lstm

human intuition => less computation no human => lots of computation

transfer learning: if another ts is similar. then, the work used to train 1st one could help train the next?

if you just do a sliding window, it def wont remember anything beyond size of win. so it requires user intervention.

next task: read on the difficulty of training rnn

do the long term effect captured if they are outside the bptt?

ilyan sketkuver goes through the rnn algos. and makes the case, for me!, that the other algos need specific tweaks

important to get unknown lags

motivations: lots of ts and idk the chars of them. want robustness

how to feed the net??

so is sgd doing backprop?

time series segmentation like anomaly detection?

google: time series segmentation anomaly detection’ to put things in perspective

why is generative model beter for ad??

understand vanishing gradients in ‘advances in traiing rnns’

state of the art coming from cs, data mining, and stats communities

now go through GroundHog and assoc paper

  1. How do neural nets fit in statistical machine translation?

I Feature extraction I Continuous-space representation I Truly data-driven: requires minimal domain knowledge

123

http://www.nehalemlabs.net/prototype/blog/2013/10/10/implementing-a-recurrent-neural-network-in-python/ u could use a simple feedforward neural network and feed all dimensions as inputs. But this comes with downsides: first, the dimensionality of your first layer will be huge, leading to more parameters to learn; second, if the data really is dynamic it could be the case that inputs will have different dimensionality and there is no elegant way to handle that.

theanets trainingRNNs/ might have clipping http://andrew.gibiansky.com/blog/machine-learning/speech-recognition-neural-networks/ We may want to predict things other than the things in the range of the nonlinearity, so instead we do not apply the nonlinearity ^^check the recrnt code in thenaonet.. how exactly does it train? how does it make the trng seq?

for prod. find a way to reduce ts size while keeping important features. reduce and extrapolate to higher def. find the errror

should i just give it various distortions of the full ts? prediction or autoencoder? seems like it predicts better with long seqs. cant fill in short seqs error vs seq len make ‘epoch’ a hyperparam! maybe opt for length of window

machine learning andrew ng

cool uses: compressiion

mcomp package http://yahoolabs.tumblr.com/post/114590420346/a-benchmark-dataset-for-time-series-anomaly?utm_content=buffer2047a&utm_medium=social&utm_source=twitter.com&utm_campaign=buffer http://nikhilbuduma.com/2015/01/11/a-deep-dive-into-recurrent-neural-networks/ three layers of optimization furutre: practical issue: reduce fidelity of ts w/o losing features future: multivariate perfect reproduction. identity function problem: assume validation set has no anomaly? argument for multiple lstm layers https://wiki.inf.ed.ac.uk/twiki/pub/CSTR/ListenTerm1201415/sak2.pdf they beat me to it!!! https://www.elen.ucl.ac.be/Proceedings/esann/esannpdf/es2015-56.pdf different lstms are about the same http://colah.github.io/posts/2015-08-Understanding-LSTMs/