CR4-DL

Lessons from the Brain

By Brian Pho
February 20, 2020 ⋅ 9 min read ⋅ Other

This page is an incomplete collection of brain findings and principles; some that may help us build AGI. Every time I run across an idea that triggers my sense of importance and utility, I will add it here. The table below maps neuroscience ideas to how we might implement it in AGI. I will organize the table better at some point but for now, I use structure to refer to the static aspects of the brain and function to refer to the dynamic aspects of the brain. Static as in if there were no electrical activity, the brain would have these principles.

Structure

Neuroscience	Source	Description
We known more about the brain than we believe	None	A common saying is that “we don’t know much about the brain” but this isn’t true. Two centuries of neuroscience have lead to the development of a general understanding of the brain and how it works, and we can use these general findings to create AGI.
The brain is highly organized	Neuroscience	The brain isn’t a giant blob of cells but is organized into regions with specialized functions. Like any other organ, the design of the brain follows from selective pressures in evolution and we can exploit this organization to organize our knowledge of the brain. Textbooks have common and well defined table of contents that divide the brain into functional components such as emotion and memory.
The brain has 86 billion neurons and is matched by none other in terms of the number of neurons in the cerebral cortex	The Human Advantage	We need to build devices of equivalent computing power. There’s a low chance of having a device with fewer units of computation but of equivalent intelligence to us. Neuromorphic chips may be the answer.
Our best framework for cognitive development is interactive specialization, the idea that we start out with a general structure that becomes more refined over time and as we mature and develop	Developmental Cognitive Neuroscience / Connectome	AGI should start out with some general priors but not any specific ones, the general priors should enable further refinement and specialization of the network
Intelligence is dependent on consciousness	The Feeling of What Happens	As evidence by cases of epileptic automatism, AGI must be conscious to be intelligent. There needs to be some sense of self and identity to attribute feelings and knowledge to. Also related to Friston’s Free energy principle.
Sparsity	After Digital: Computation as Done by Brains and Machines	Computation in the brain doesn’t use all neurons nor is it distributed. Computation appears to be sparse, requiring only a select fraction of neurons. The same principle appears for memories where it only activates a certain portion of neurons (engram). AGI should use sparse representations to be efficient and only use what’s needed.
Representation	Cognitive Science: An Introduction to the Science of the Mind	The idea of representation has shown up throughout most of the neuroscience textbooks that I’ve read and in ML. It’s because the brain must make some copy of reality which we call representation, it represents reality. AGI will have to implement some form of representation that’s efficient.
Representations serve the brain and not reality	Magenta Is All In Your Head	Surprisingly, the brain can create colors that don’t exist. For example, magenta is a color we perceive but doesn’t map to any wavelength on the electromagnetic spectrum. This shows that the brain’s representations aren’t trying to accurately model reality, but that representations serve survival and the brain.
Sensory organs model the world	Vestibular system and all sensory organs	For the brain to have a meaningful impact on the world, it must match its environment, thereby allowing it to exploit the maximum amount of information regarding the world. E.g. The vestibular organs have three semicircular canals that are all perpendicular to each other, matching the 3D space that we inhabit. Another example is that eyes evolved to see in a certain wavelength because that range of electromagnetic frequency lead our ancestors to survive and reproduce.
Refinement and synaptic pruning to make neural pathways shorter, faster, and more efficient	SciShowPsyc video / Neuroscience: Exploring the Brain	Pathways are the name of the game for neurons as neurons use the path that data travels to process it. A similar idea applies to neural networks where the follow of data along certain paths dictates the outcome of that data. Refinement and synaptic pruning provides one mechanism for the modification of pathways.
The meaning of neural signals is determined by the pathway	Janet Casgard video	What I believe to be the most neglected idea in neuroscience. When the brain receives an action potential from the body, how does it know it came from the arm or leg? The pathway of the action potential defines the meaning since all action potentials are stereotyped/similar. This also results in the somatotopic organization of the brain and spine. This translate into AGI as a new paradigm of programming. To program not the signal but the pathway the signal travels.
Causality and the idea of counterfactual	The Book of Why	Humans are unique in their ability to imagine the unreal. Counterfactuals, or the opposite of facts, deals with our ability to imagine reality as it can’t be, to say that
System 1 (fast and intuitive) and System 2 (slow and rational)	Thinking, Fast and Slow / Yoshua Bengio - From System 1 Deep Learning to System 2 Deep Learning	At a high level, the brain seems to have two systems to process events in the world. These two systems arise from different brain regions and is especially important for how we manage day-to-day events. We shouldn’t expect to build this into an AGI but rather the system should exhibit system-1-and-2 properties.
Development of intelligence through evolution (phylogeny)	Developmental Cognitive Neuroscience	Billions of years of evolution has shown us that intelligence is a rare and difficult trait for an organism to develop. This is further emphasized by the difficulty of finding life on other planets as life seems to prerequisite intelligence. Only through an iterative approach has evolution forged us. This suggests that if we develop AI in an iterative approach, we should expect to move from simpler, older, organisms to us. Evolution provides a measure of progress for AI.
Development of intelligence through maturation (ontogeny)	Developmental Cognitive Neuroscience	Similar to the point above, our maturation provides a measure of progress for AI. If we develop an AI using the approach stated by Alan Turing, we should expect a similar developmental pathway.
Adaptation	No source	The brain developed in many different organisms, ranging from dinosaurs, fruit flies, and humans. This suggests that the brain is generalizable to different environments and times. Another important, but overlooked feature, is that the brain can handle modern times. We didn’t evolve with chess, driving, typing, and digital screens and yet our brain can handle it all. The brain hasn’t changed in the last tens of thousands of years so it’s remarkable that the brain has adapted to whatever time it’s in.
Superintelligence may result from improvements in humans	Superintelligence: Paths, Dangers, Strategies	I’ve been so focused on imbuing machines with intelligence that I forget that we can make ourselves more intelligent through genetic engineering, more brains thus higher collective intelligence, drugs, and improved education. We must remember that the goal is a greater intelligence, not creating machines with greater intelligence.
Embodiment and the idea that the brain evolved to guide action	The Wiley Handbook of Evolutionary Neuroscience	An AGI system must have some way of influencing the world. This means to build some input and output system for AGI such as allowing it to take in byte data and output byte data, or to create cameras for AGI and robotic bodies for it to manipulate the world.
Speed, volume, noise, and energy are the four basic constraints on the design of neural circuits	The Wiley Handbook of Evolutionary Neuroscience	These four limiting factors affect the creation of neuromorphic chips for AGI and also influence our hardware design of AGI.
Social learning and collective intelligence	The Wiley Handbook of Evolutionary Neuroscience and this Reddit post	More evidence and thinking leads us to believe that AGI won’t appear from a single smart device but rather from a collection of devices acting together. Humans have exceptionally high-fidelity copying which allows us to pass on and cumulatively collect a massive knowledge base for current humans to build off of. Our culture has played a massive role in bootstrapping our development and the role of teaching has often been overlooked in AGI.
The brain is both general and specialized	General neuroscience	The brain is able to take in a variety of different sensory input such as light waves and sound waves and process them. It has specialized processing for each sense such as auditory stream segregation for sound and object recognition for light. However, different species have different sensory organs and different bodies, suggesting that the brain is also built to handle a variety of sensors and actuators.
The cortex is organized into cortical columns	Neuroanatomy through Clinical Cases	The cortex isn’t randomly organized but uses units of vertical columns to manage information. It appears that each column manages one representation, such as a line orientation, and that neighboring columns have related representations.
Energy was a big limiting factor in the evolution of the brain	Neuroscience Paper	There’s evidence that gut size was traded for brain size, which likely resulted from having a high-quality diet.

Function

Neuroscience	Source	Description
The brain has a surprisingly long development time compared to other animals	Developmental Cognitive Neuroscience	AGI should intentionally have a delayed maturation period to improve learning and adaptability. This may come in the form of slower learning rates, increase epochs, or the slow development of neural architecture.
The brain builds specialized areas dedicated to certain functions.	Neuroscience: Exploring the Brain	AGI should have specialized modules to handle functions such as language, vision, and audition. It should then combine them in association areas using a hierarchy organization.
Learning, recall and memory	Make It Stick	Learning and memory are intimately linked and learning depends on memory. The result of learning is a memory. Recall and retrieval is the link between learning and memory as learning is the process of retrieving memories. Evidence in the form of patient H.M. shows that a deficit in memory impacts learning.
The brain is event-based, not clock-based	Artificial Intelligence - Perspectives from Leading Practitioners	This idea is in direct contrast with our current computing paradigm and advocates for a neuromorphic-based computing paradigm. The differences in processing with events instead of predefined chunks leads to massively different paradigms in how we think about computation and how it can be realized.
Interleave between digital and analog signal processing	After Digital: Computation as Done by Brains and Machines	The brain uses analog processing in the form of graded temporal rates of APs to convey information such as faster firing for more muscle stretch and in the graded release of synaptic vesicles. However, the brain also uses digital processing in the form of all-or-nothing signals called action potentials. This is one of the weaker ideas that is too general to be applicable to AGI.
Interleave between memory and computation	Fundamentals of Computational Neuroscience	There aren’t any specific areas in the brain dedicated to computation or memory like in a computer. Both are interleaved in the brain in neurons and synapses. This leads to an efficient design of matching memory to where computation is needed. This also avoids the von Neumann bottleneck.
Efference copy and closing the feedback loop	Livewired	To learn how to use our sensory organs and our muscles, the brain employs a feedback-type of learning. The idea is that the brain will generate an efference/motor copy for every action it performs and sends this copy out to relevant parts of the brain to predict the incoming sensory information. If the copy matches the input, then the prediction is correct and inhibits the sensory information. E.g. When you speak, your brain dampens your own voice heard through your ears as it’s producing what’s expected. If the copy doesn’t match the input, the brain adjusts either it’s efference copy or the sensory area to reduce the error.
Memory works through different temporal layers	Livewired	Memory is divisible into different types but also into different temporal scales. Like a computer with different memory (cache, ram, ssd), the brain employs a layered memory system to only retain what it needs.
Information isn’t coded in single neurons	True Brain Computing	If single neurons encoded a representation (grandma cell theory), then the representations wouldn’t be robust to the destruction of the neuron and the brain wouldn’t be able to storage as much.
Not all information is encoded in APs	True Brain Computing	While APs have been the focus of information coding in the brain, another form of electrical activity, graded potentials, also convey information. It’s also been ignored that even if a neuron doesn’t spike due to the summation of graded potentials not surpassing the threshold, this still conveys information.
Multiple Realizability	Neuroscience Paper	A biological function can be implemented in neurons in multiple ways, such as sound localization and vision. This suggests multiple paths to AGI than just the human path.
Processing starts at the collection of information	Neuroscience Paper	As a general rule, information is altered every time it passes through a synapse and this means the eye is doing more than just sending raw pixels to the brain.
The brain is a prediction machine	Livewired	Multiple lines of evidence converge on this idea that the main function of the brain is to predict the future. Evidence comes from Livewired, Spikes, and The Brain is a Time Machine

Previous
Eloquent JavaScript Next
Developmental Cognitive Neuroscience: An Introduction