helixcarinae / leela

LEELA (Low-Exposure Emergent Learning Agent)

Introduction

The LEELA Project seeks to create a model that is capable of producing acceptable results by utilizing only a small number of labelled training examples. "You don't have to tell me twice," is a phrase that doesn't apply to a machine learning agent but hopefully this won't always be the case. There are many limitations of current deep learning methods one of the key ones being the importance of big data. This reliance on massive dataset to produce accurate and reliable results massively stifles both the ability to create unique agents where a large dataset doesn't yet exist and limits the ability of small players and start-ups to be able to generate their own useful models.

Thinking Like a Human

While current machine learning models are loosely based off of biological neural networks they don't currently operate to the same capacity. LEELA is driven by a desire to recreate the ability of a human to perceive and understand an object or concept with only a few exposures (past young childhood).

A Brain System Model

While neural networks are a great start for intelligence often the best way to get to an outcome is model after something that already works. While the brain is comprised of a very large mesh net of neurons they clearly don't operate the same, a look at an fMRI can quickly show this. What I don't believe is known yet is whether these sections of the brain are baked into the genome or whether they emerge naturally just from stimulus inputs. We do know they aren't ironclad they can be repurposed, someone that's blind doesn't just have a section of their brain that doesn't do anything it gets reused for other stimulus that is if they went blind at a young enough age for this to happen.

I believe it will be necessary to create a system that mimics these structures to create an agent that can naturally process and output information in a way similar to a human, not necessarily general intelligence, but general processing and the ability to dynamically integrate and process constantly changing stimuli.

In addition I believe some form of "thinking" or "imagination" will play a critical role. In humans the hippocampus is a sort of memory cache that can be quickly retrieved but also quickly removed. It also seems to be a key piece in how we think or imagine, we can think about the past or try to think about the future or approximate a future response from various stimuli that have already been incorporated into the mind.

In terms of a computer model I believe that it will need a mechanism for generating outputs. What this mechanism would look like is currently an underdeveloped territory.