Talk:DreamTeam - Revision history

192.195.83.130: date

2017-02-16T06:43:34Z

date

2017-02-16T06:42:53Z

rbm.py 15 Feb 2017

@@ Line 4: / Line 4: @@
 #!/usr/bin/env python3
 import math
 import random
-from loader import MNIST # see https://github.com/sorki/python-mnist
+from loader import MNIST
 print('Hello!')
 num_phase = 3
 num_visible = 28 * 28
-num_hidden = 100
+num_hidden = 28 * 28
 temperature = 42
-learning_rate = 0.0042
+learning_rate = 0.00042
 batch = 1
 ABS_limit = num_phase
 hidden = list(0 for i in range(num_hidden))
@@ Line 71: / Line 73: @@
 def alternate_gibbs():
    for AGS_phase in range(1, num_phase + 1):
      if AGS_phase % 2: #is odd
        compute_hidden_energy()
@@ Line 100: / Line 101: @@
    for i in range(10):
      res.append(random.choice(epoch[i]))
    return res
@@ Line 111: / Line 113: @@
    for v in minibatch:
      visible = v
      alternate_gibbs()
    apply_weight_update()
@@ Line 149: / Line 152: @@
 for every hidden_node :
 weight[i][j] += learning_rate/batch * weight_update[i][
-"""
+j] """
 </nowiki>
 ==Brainduino images==

2017-02-09T07:28:46Z

MNIST github link

← Older revision		Revision as of 00:28, 9 February 2017
Line 6:		Line 6:
	import math		import math
	import random		import random
	from loader import MNIST		from loader import MNIST # see https://github.com/sorki/python-mnist

	print('Hello!')		print('Hello!')

2017-02-09T07:27:50Z

rbm python code - first draft

2016-12-22T16:57:16Z

2016-12-22T03:56:39Z

Upload images of brainduino

← Older revision		Revision as of 20:56, 21 December 2016
Line 1:		Line 1:
			==Brainbuino images==
			[[File:IMG_9574.jpg]]
			[[File:IMG_9573.jpg]]


	==Discussion 14 December 2016==		==Discussion 14 December 2016==
	Restricted Boltzmann Machine		Restricted Boltzmann Machine

2016-12-15T05:13:46Z

rbm

← Older revision		Revision as of 22:13, 14 December 2016
Line 1:		Line 1:
			==Discussion 14 December 2016==
			Restricted Boltzmann Machine
			* is RBM not an RNN ?
			* summary recollection - RBM is a type of NN.
			Has layers. Input, output, hidden. Can have multiple hidden layers.
			Has nodes and weights.

			What is a "Recurrent" Neural Network?
			Why not "Recursive"?

			How is Error Backpropogation different concept from "Recurrent"?

			Network architecture ...
			adjust weights within rather than redefine network topology

			RNN may have generative capacity.

			Recollections of current comprehension:
			* it is restricted. ie connections not allowed within layer

			RBM involves MCMC (Markov chain Monte Carlo) process
			via Gibbs Sampling

			relates to Ising Model
			Hopfield nets

	==Discussion 07 October 2015==		==Discussion 07 October 2015==
	Distributed neural networks		Distributed neural networks

2016-12-01T04:47:31Z

← Older revision		Revision as of 21:47, 30 November 2016
Line 13:		Line 13:
	https://media.blackhat.com/bh-dc-11/Grand/BlackHat_DC_2011_Grand-Workshop.pdf		https://media.blackhat.com/bh-dc-11/Grand/BlackHat_DC_2011_Grand-Workshop.pdf

	Restricted Boltzmann Machine		==Restricted Boltzmann Machine==
	* roots in Hopfield network		* roots in Hopfield network
	* relevance of pruning - restricted connectivity makes training feasible for non-trivial number of neurons		* relevance of pruning - restricted connectivity makes training feasible for non-trivial number of neurons

2016-12-01T04:47:09Z

← Older revision		Revision as of 21:47, 30 November 2016
Line 8:		Line 8:
	* folding@home 40 petaflops - how might this support simulation of something like 10**11 neurons		* folding@home 40 petaflops - how might this support simulation of something like 10**11 neurons
	* compare to spiNNaker - specialized network for thousands of ARM cores, designed for spiking neural networks but can support traditional backpropagation		* compare to spiNNaker - specialized network for thousands of ARM cores, designed for spiking neural networks but can support traditional backpropagation

			==Hacking Hardware==
			Presentation @ Black Hat 2011, discusses JTAG
			https://media.blackhat.com/bh-dc-11/Grand/BlackHat_DC_2011_Grand-Workshop.pdf

	Restricted Boltzmann Machine		Restricted Boltzmann Machine

2015-10-15T05:23:58Z

Discussion 07 October 2015

← Older revision		Revision as of 22:23, 14 October 2015
Line 2:		Line 2:
	Distributed neural networks		Distributed neural networks
	* hadoop - useful to handle reliability when dealing with thousands of servers		* hadoop - useful to handle reliability when dealing with thousands of servers
	* YARN yet another resource ~~manager~~		* YARN yet another resource negotiator
	* apache - also mahout		* apache - also mahout
	* task oriented - wants to solve actual data analysis quickly - not energy efficient		* task oriented - wants to solve actual data analysis quickly - not energy efficient