Musing on Brain Architecture

I need to find more time to look deeper into neuroscience again. A lot of the emphasis in Deep Learning has been on models of neurons starting from the McCulloch–Pitts (MCP) neuron in 1943 over Frank Rosenblatt’s Perceptron in 1958 to the Hodgkins-Huxley model which models action potentials in neurons and led to a Nobel Prize in Medicine in 1963. It is interesting that Spiking Neural Nets are more biologically plausible, but have not yet defeated their artificial counterparts and that there are also biologically more plausible alternatives to backprop like zero-divergence inference learning (Oxford) [regarding the learning algorithm also see this] as well as comparisons of DNNs to fMRIs such as Cichy et al.. Interestingly, Beniaguev et al. determined that it takes 5-8 DNN layers to model one cortical neuron [they also proposed a method to determine the computational complexity of any neuron type].

However, it should be beneficial for AI Architecture to take the brain’s compositionality and functional areas into account rather than exclusively focusing on individual neurons. One aspect of this are cognitive architectures targeted at biologically accurate properties such as ACT-R, but it also makes sense to look deeper into neuroanatomy itself. Let’s first take a look at the big picture:

For instance, the rough brain area correspondences are fascinating:

• Cortex: Long-Term Memory
• Basal Ganglia: Program Sequencer
• Thalamus: Input and Output
• Hippocampus: Program Modifications
• Amygdala: Rating What’s Important

The following illustrates the functional areas in more detail [image licensed from Alamy]:

Brain signalling is also inspiring. There are primarily two categories of neuro-transmitters (please see this as reference):

• Small-molecule transmitters (e.g. dopamine and glutamate) which typically act directly on neighboring cells
• Neuropeptides (e.g. insulin and oxytocin) which modify how cells communicate at the synapse

Among these categories there are a few particularly prominent types (see same source):

• Acetylcholine (Ach) – helps translate intentions into signals from neurons to muscle fibres, also helps direct attention and facilitating neuroplasticity (forming new neural connections to adapt to environment)
• Dopamine (DA) – motivation, decision making, movement, rewards, attention, working memory, learning
• Glutamate (GLU) – most excitatory neurotransmitter, helps learning and memory; LTP (molecular process to form memories) occurs in glutamatergic neurons
• Serotonin (5HT) – manage appetite, sleep, memory and decision making; “calming chemical”; lack can lead to depression
• Norepinephrine (NE) – mood, arousal, vigilance, memory, stress [NE is both hormon and neurotransmitter]
• Gamma-Aminobutyric Acid (GABA) – inhibits neural signaling; too much can lead to seizure; helps early brain development, “learning chemical”
• Many others

I feel like the different signals have been underexplored. While Sutton & Barto dedicated chapter 15 of their book to the biological plausibility of RL approaches, it seems like too many approaches just look at dopamine and thus run the risk of simplifying the picture too much. Furthermore, Smith et al.’s “The fine structure of surprise in intuitive physics: when, why, and how much?” (2020) introduces a surprise signal to facilitate learning for violation of expectation (VoE) experiments in intuitive physics – if a property like object permanence is violated, the surprise signal spikes. There is a question of how many of these signals increase performance and to which degree the brain can be regarded as an event-driven architecture (EDA).

I also feel like hardware platforms play a major role. Similar to how GPUs supported the deep learning revolution, both neural accelerators, but also neuromorphic systems like TrueNorth and NorthPole out of Dharmendra Modha‘s lab in Almaden could significantly impact the development of AI architecture. One major difference is that the brain’s compute substrate seems to be simultaneously its storage (afaict), so technologies like Memristors might see a revival (speculative, of course – they have also been discussed since the 70s).

Regarding neuroplasticity, it seems that there are only four types (see this source):

• Homologous area adaptation – assumption of a particular cognitive process by a homologous region in the opposite hemisphere.
• Cross-modal reassignment occurs when structures previously devoted to processing a particular kind of sensory input now accept input from a new sensory modality.
• Map expansion is the enlargement of a functional brain region on the basis of performance.
• Compensatory masquerade is a novel allocation of a particular cognitive process to perform a task.

Since one of the major bets of my career is that AI Architecture has to be regarded as its own field, I keep looking into the neuroanatomical aspects as well. A good summary of circuit architectures can be found here as well as in the excellent Handbook of Brain Microcircuits by Shepherd et al.