Next Steps, and Dreaming in Relation to Long-Term Memory.

This is an excerpt from my email I wrote regarding next steps:

1. Make a diagram of the conception of the system as it was before, including how I was thinking of links between high level neurons having different weights for each high level feature (which may be made up of a vector of lower level features, eg. histogram). This would be the context in which to situate discussions of machine learning systems.

2. Read more about LIDA to understand better how percepts and knowledge from memory stores are structured, and if a subset of the architecture can be used.

3. Wish list for machine learning / clustering (in no particular order)

• Unsupervised

• Ideally statistical (stores a distribution of patterns, not all the patterns themselves, but this is flexible)

• Output could be manifest in connection weights between neuron-like structures that represent higher level concepts. (concept as distribution of activation) Each item should belong to a concept to a degree, not Boolean.

• If it is deep learning and I can forgo the feature extraction, it must be elegant (simple enough for me to understand) and be feasible for high resolution colour input images.

• If it is deep learning it should be generative: can learn high level abstractions (concepts) from sensory data, and also construct sensory data from those high level abstractions.

Thank is all I can think of off hand.

Read more

Biologically Inspired Cognitive Architectures

On one hand we have the deep machine learning systems discussed in the previous post, which are quite complex (from a mathematical perspective) and only roughly informed by neuro and cognitive sciences. On the other hand are cognitive architectures that are inspired by biological knowledge (BICAs). These are systems that attempt to balance high level (psychological knowledge) with low level (neurological knowledge) in a tractable system that models mind-brains. These systems can be purely symbolic, connectionist (sub-symbolic) or hybrids of the two. While they are biologically inspired they favour a high level abstraction of cognition over fine details of brain (neuron) function.

Read more

Deep Machine Learning

As part of stepping back to see the big picture I’ve turned a second look at deep machine learning systems and biologically inspired cognitive architectures (as suggested by my supervisor), the latter of which will be discussed in another post.

Deep machine learning systems attempt to resolve an issue with “shallow” learning which has increasingly become the following process: Input → Feature Extraction → Machine Learning. An argument against this approach is that the “intelligence” shifts from the machine learning system to the human-centred, and domain specific, art of feature extraction. Deep learning systems excise the middle man, allowing Input → Machine Learning, without the intermediary.

Read more

Stepping back to see the big picture

After writing the previous post I had a meeting with my supervisor yesterday. He suggested that the answers to these questions of feature abstraction should be contextualized by the machine learning method used to organize these perceptual units. So I’m putting further development on the back burner to look at the big picture and do more reading. Following is a figure of the overall structure of the whole system, as I currently imagine it.

Read more

Colour Features and Concepts

Since it’s time to explore some clustering methods for these perceptual patches, I first calculated a RGB histogram for each patch, making use of the mask. In attempting to make sense of the structure of the openCV hist (1 channel, but 3 dimensions) I realized that perhaps the hist is not appropriate for this project. The central question is: what is important about colour for the purpose of associating perceptual units? Perhaps the histogram should be abstracted into a set of concepts (for lack of a better term) representing major colour clusters.

Read more

Segmentation Success!

I’ve finally managed to get the segmentation data into a useful form. Following is a reconstruction of the original image from extracted patches, and a corresponding image that shows the segments, filled in random colours.

Read more