Neural Graffiti – A Neuroplasticity Drop-In Layer for Transformers Models

Liquid neural networks are awesome - they change how that "neuron black box" connects over time given its past experiences, emulating the human brain in relating concepts and how it changes our perspective.

They are great at time series forecasting like weather and analytics, however the idea is to do it on a transformers model, making it acquire neuroplasticity at token prediction - and as we know its very expensive to train a whole model from scratch.

I figured we could splice in a new neuron layer inside the model's networks right between the transformers layer and the output projection layer that actually predicts the tokens. This way the thought would have "influences" of past experiences for every token generated aka. during the entire line of thinking, making the model acquire a "personality in behavior" over time.

The vector embeddings from the transformers layer are mean-pooled and "sprayed" with past memories changing the way each token is generated, influencing the meaning and therefore choice of words in the vocab space. This neural “Spray Layer” also remembers the paths it took before, blending new input with previous ones and gradually evolving its internal understanding of concepts over time.

It won’t guarantee exact word outputs, but it will make the model lean into certain concepts the more it interacts. For example: Tell it you love dogs, and over time, the model will start leaning toward dog-related kindness, loyalty, and fuzziness in its tone and direction. More teste are yet to be done and I know there is a cold start problem, finding the sweet spot is key.

This is quite fascinating, especially because we don't know exactly what happen at the model's transformer neuron level and how it makes the connections, but hacking it like this is interesting to watch.

I called this technique "Neural Graffiti", and it is free and open for everyone.

Try the demo and give it a star on the github repo!