Home

Problems with current AI and inspirations from biology

Here's an attempt at an overview of what's relevant/missing in the quest to unlock machine intelligence.

We need a simple architecture that covers all these bases! (our own brains do!)

What properties do bio-brains have that AI is lacking?

# Considerations
What qualities does a neural engine need to have?

🔸 Inference and training together!
   🤖 In LLM, in-context-learning isn't actually _learning_ (model doesn't get updated)


🔸 Homogeneity
    🌱 Neocortical tissue is homogenous.
    We want a basic data+computation mechanism that's abstracted away from input modality
    So image / audio / text / sensory / etc. data goes thru some initial layers
      INTO a generic neural engine.

🔸 Universal representation
    🌱 Bio-nets use SDRs (Sparse Distributed Representations).
      Neocortical tissue is a (folded up) 'cloth' of columns.
      Underneath this 'cloth' are 'optical-fiber--bundles' (axons) -- 'white matter' under the 'grey matter'.
      Taking a slice through a bundle, you can imagine a pattern of activity.
      It might look like TV static, but there's information encoded.
      And it's an entirely generic protocol. 

🔸 Scale-Free / pondering
    🤖 Current Transformer has fixed nFlops per generated token.
    Want a model that is able to "ponder" a hard token.

    letter <-> word, word <-> phrase, phrase <-> sentence, etc.
    Need a mechanism that handles arbitrary abstraction
    It's ugly to separate into fixed "N layers" (won't scale)

    🌱 Bio-brains have a thalamo-cortical cycle.
    Not every cycle generates an action.
    Watch a bird on a tree branch. See how it twitches sporadically.

🔸 Energy efficiency / sparsity
    🤖 Transformers are fugly huge / inefficient

    🌱 bio-brains run at low watts. Only a small fraction of the brain NN is used for a computation,
    evolutionary forces have pushed us towards efficient computation. fastest reaction wins. (Genetic Algos + multimodal AI gym setups???)

🔸 Agency
    At what level is agency?
    🤖 It wraps around the policy model (e.g. LLM)?
    🌱 Maybe we can see it as an emergent property from neuronal behaviour (competition/inhibition)

    - Recursion / looping
        🌱 Our brains loop. We learn to loop. e.g. socratic dialogue, problem-solving
        🤖 We have dynamical systems (RNN -> SSM; mamba/griffin/rwkv/...)
            yes next-token-prediction is a recursion / state-evolution, but fugly. It leaks memory (accretes state).
    RL is clumsy (takes many tries)

🔸 Introspection / self-modelling
    🌱 Max Bennett (https://www.abriefhistoryofintelligence.com/) observes 5 stages of intelligence:
        - basic organisms like corals have brains (sensory->brain->motor)
        - reptiles have RL systems (basal ganglia)
        - mice are able to simulate
        - monkeys are able to simulate THEIR OWN behaviour
        - humans can do language
    🤖 Should be possible to train/finetune/align an LLM to do something like this (📍 we're working on it)

🔸 Learning
    🔹 [TODO] Intrinsic learning

    🔸 Local Learning
        🤖 weights-update via backpropagation of error-gradient requires global-forward then global-backward.
            The greater the error-delta on a synapse, the more it gets tweaked in the learning process.
            This encourages catastrophic forgetting.
            If a large part of the network is not contributing to this computation, LEAVE IT ALONE!

        🌱 bio-nets act locally; <2% of the brain is active at a given moment.
            The areas of the network that are MOST ENGAGING with the current task will get the strongest learning signal.

    🔹 One-shot learning
        🌱 We don't need to get run over at 1000 red lights to learn to wait for the green when we cross the road
        🌱 Our bio-system releases chemicals to strongly reinforce certain patterns

    🔹 Solidify early-learning (Critical Periods of Plasticity)
        🌱 Observe how our auditory cortex learns. Low-level is baked in before puberty.
            e.g. Dynamically add neurons. A neuron's learning rate decreases over time.

    🔹 Intelligent teacher-student training (curriculum learning)
        🤖 It's ludicrous to train an LLM on a million kids' stories
        📍 Keep feeding the student challenges where it can solve say 70% of them

    🔹 Continual learning
        🌱 A bio-brain can repeatedly contemplate, think/distill thoughts and self-update
        📍 What if the engine can extract targeted information from its environment (e.g. read a new arxiv paper/blogpost)

    🔹 Wake/Sleep / Dreaming
        🌱 Experience replay in REM-sleep walks recent trajectories (fast-brain: hippocampus) training slow-brain (neocortex)

🔸 Maslow's hierarchy of drives
    🌱 There's a firmware in bio-brains; hypothalamus / amygdala provide (+/-) valence/reward signal

🔸 Evolutionary learning (neuroevolution?)
    🌱 Agents with traits operate in environment. Successful ones breed & combine winning traits.