Home
People
Projects
Research
topics
Publications
Contact |
Cognitively Inspired
Pattern Discovery methods for latent and Associative model Learning (CIPDAL)
The goals:
Develop new, cognitively motivated pattern discovery and learning methods.
Create new knowledge and theory regarding computational learning, focusing especially on aspects of language acquisition.
Patterns in sequences
We study different discrete representations and transformations of continuous speech into sequences of limited alphabet.
Pattern discovery and learning is then focused on search of frequent, regular subsequences.
We also study multimodal representations and associations between multiple sequences.
Data compression and model learning
Pattern discovery can be seen as a search of regular subsequences, which can be used to compress the sequence.
Efficient compression of sequential information means that important regular subsequences have been found, modeled and removed.
This means that pattern discovery and model learning are very close to data compression.
 |
Character patterns automatically discovered from a noisy text document image.
Characters are discovered
spontaneously, extracted, and clustered with the HML algorithm without any prior information or external help. Two
different discovered cluster sets are shown: o and e.
|
Model learning
Once the models of the discovered patterns are learned, the models work as an active memory, storing the necessary structures for:
- later compression of similar patters
- recognition of similar patterns
- classification of similar patterns
- full reconstruction of discovered and modeled patterns
Among others, we will apply, study and develop further the new HML method.
Learning environment
Study of multimodal pattern discovery related to language acquisition requires interaction between a learner and its environment.
In CIPDAL we can directly control the complexity of the environment by using an entirely simulated platform, where digital audio and visual
input are introduced to the learner according to its actions in the environment.
 |
| A sketch of interaction platform where pattern discovery and learning takes place.
|
|