AI vs CC (Cortical-Columns-Algo) : tasks

Engineering
, ,
1

I approximately 'm aware of even intricate points of HTM, the things I lack is understanding how to apply it.

Let’s take the THREE major problems/techniques in AI and try to make them work with HTM

  1. Classification : this should be the bread and butter of HTM i.e. recognizing objects

  • a possible approach for dumb-classification is to use just the TM as a two-step sequence prediction. of course there is the problem of start-and-stop of the sequence. Not very feasible cause TM is not suited for artificial separation of sequences.

  • second approach is to use full CC, but then you need and SDR to eventually apply Apically on-top of Temporal Pooler to force somehow TM to sync the sequence with a static SDR. First where will you get this SDR from, second the whole forced approach seems iffy.

  1. Unsupervised learning : this seem somewhat feasible you present different objects at the bottom and TM records the sequence of interactions with an object and TP summarize to give final SDP representation. It is not clear how will TP generate INVARIANT representations of the same object under different interactions.

  2. Reinforcement learning : TM afaik is a recording device, there is no notion of State:Actions and the idea of Agent in the current CC mechanism.
    There is the notion of Agent in the theory but not in the CC algorithm. Also no rewards mechanism.

All THREE problems can be described/represented in a grid world toy scenarios :

  1. Minst number DB
  2. Grid with X/O representing objects
  3. Searching a path

There was one paper that dealt with object recognition, but the description was of the HTM overall design and the results, not of the actual setup and design of the tests, which is expected of course. Things like this are omitted in papers in general in the interest of space.

Help me shoehorn those problems for CC algorithm that uses SDR ?
What is the representation ?
How do you translate from (X,Y,Sign) in the grid to SDR and how does this make sense compared to the final object SDR ?
What are the performance METRICS and on what representation are they applied ?
How do you solve the “impedance”-mismatch between the input data and SDR (dont mean encoders) i.e. the input and output is not only different in representation, but also they are categorically different ?

ARE AI problems not suitable for CC algorithm ?

2

What does “CC” stands for?

3

Cortical Columns algorithm. updated the title

4

Ah, ok thanks. Makes more sense now.

My (too general) guess on this issue is:

  • whatever the CC algorithm is, it should be equally useful in solving any of the three kinds of problems - unsupervised learning, reinforcement learning and classification. And I would include the fourth category generation capability. e.g. we are able not only to associate concept of “six” with a pixelated image of digit “6” but also to internally generate an image of “how does a six looks like?”
  • The algorithm should provide automatic enrollment of “gangs” of limited complexity agents to collaborate/learn together for solving incrementally complex problems. Problems which given the limited resolution/inference capacity of a single agent could not be solved by any of them individually.
  • And the above should optimize for minimizing resources needed to solve problems, both in space and time. I mean in computer metrics to minimize amount of cycles and memory lookups and in brain’s metric to minimize both number of neurons spiking at the same time and specially, since spikes are quite time consuming, the number consecutive spikes needed to reach an answer/conclusion.