Hierarchical temporal memory (HTM networks)

HTM networks에 대해 논한다.

 

hidden layer 안의 neuron 끼리도 정보를 공유하는 networks..

일반 nn 보다 더 많은 edge를 갖는 networks로 사람의 복잡한 뇌처럼 구조화하려 한 모델이다.

 

딥러닝과 큰 차이점은 back propagation 방식을 쓰지 않고 feed foward 방식으로 학습을 한다는 것이다.

 

online learning이 가능하고, sequential 모델로 볼 수 있어 시계열 분석을 이용한 이상탐지 모델에 주로 사용되고 있는 추세이다.

 

 

출처 : https://github.com/llSourcell/numenta_explained

llSourcell/numenta_explained

 

Online learning with Numenta¶

What is Numenta?¶

• Jeff Hawkin's published On Intelligence in 2004.
• It was a good layman pop neurosci book for the time, and indirectly led people to deep learning
• Andrew Ng for one mentioned it as a strong influence

How is their HTM system different from deep learning?¶

• HTM uses online unsupervised data (temporal streams), so can't compare to CIFAR results
• LSTM is the closest comparison candidate, and it cannot do online learning without scaling very poorly (but this is crucial for AGI)
• Lets keep thinking of entirely new architectures. DeepMind'sDQN + moar processing != AGI (backrop isn't the end!)
• HTM is directly based on neocortex, DL is inspired by neuroscience
• DL systems are very good at doing low-level perceptual tasks, but require huge amounts of examples to learn anything well.
• HTM can learn complex temporal structure with several orders of magnitude fewer examples (dozens rather than millions) and can also be easily configured to do reliable one-shot learning

• HTM networks are not like ConvNets at all. Every aspect is different. Each neuron is more like a network in its own right (using active dendrites), and a layer of neurons uses local k-winner-takes all (inhibition) and not max pooling.
• In addition, there is complex structure in using columnar inhibition to represent arbitrary-order sequence memory and to encode prediction errors.
• HTM is only really like a convolutional neural network in that it is a tree-like architecture with locally connected (sparse) weights.
• Unlike convnets, it doesn't use shared weights. It also serves a different purpose. It acts like a really large predictive autoencoder that predicts its own input one step ahead of time. - From there you can do reinforcement learning and other interesting things. HTM's spatial pooling isn't really like max pooling, it is almost exactly like k-sparse autoencoders though.
• HTM is not feed-forward. Information goes both up the hierarchy, to form abstractions and extract features, as well as back down to form prediction

TL;DR: HTM is like a really beefed up autoencoder.¶

The main thing that interests me in HTM is its online nature. I don't know of any other algorithm that can learn in real-time from data streams without using some form of experience replay. HTM can, and it does so quite well (you can learn a sheet of music, for instance, after as little as 26 iterations).¶

Apps that use HTM?¶

Hierarchial Temporal Memory explained¶

The Neocortex¶

-reptiles dont have one -mammals do -size of the dinner napkin -older parts of the brain involved in basic functions of life -It makes you "you"

-Regions linked together in hierarchy -Ideas become more abstract and permanent up the chain

• HTM prinicple - Common algorithm everywhere
• HTM principle - Sequential memory
• HTM princople - Online learning
• I/O of neocortex (sensory input and motor commands)
• sensory encoder- How to turn input values into SDRs
• GPS encoders have no biological counterpart

Sparse Distributed Representations¶

• Using 1s and 0s to represents neurons on and off (basic principle of HTM)
• The data structure of the brain
• Everything is SDRs. used for every aspect of cognitive function
• Neurons receive them from other neurons and everywhere
• They can overlap i.e sets and unions (combining sdr's')

Encoders¶

• like sensory organs like retina or chochlea
• outermost system of HTM system
• principle 1- Similar data should be highly overlapping
• 2- same input should create same output (determinsitic)
• 3 output should have same dimensionality as input
• 4 output should have similar sparsity as input

Spatial Pooling¶

• Accepts input vector and outputs a vector
• Important when talking about sequence memory
• Maintaining fixed sparsity is a goal
• Maintaing overlap properties of inputs for outputs

Temporal Memory¶

• Learns sequences
• Predicts outcomes

Towards AGI¶

• Weak AI wont produce intelligence
• We need to incorporate movement i.e interacting with environment
• Better neurons like the pyramidal neuron which has Layers and columns
• so the Columns contain layers and layers contain neurons
• It has both an ACtivate and inactive state. And predictive state
• feedforward, lateral input, and apical (higher level)

is HTM similar to Hinton's Capsule Network?¶

Well for both systems

• objects are defined by the relative locations of features
• a voting process figures out the most consistent interpretation of sensory data,

But the big difference is

• HTM models movement. HTM explicitly models how information changes as we move our sensors (e.g. as we move our eyes around), and how to integrate information to quickly recognize objects.