decoding neural operations

The human mind is arguably the most complex and awe-inspiring entity among the totality of entities in the world. The most ambitious scientific/technological projects of the present era lies around understanding and emulating the human brain which I have covered to some extent through a previous article - ‘mapping connectome’. The human brain can be defined as the most efficient intelligent information processing system in the world which consumes about 20 watts of power to perform a wide array of complex operations as compared to artificial intelligence systems which would require a staggeringly high ~5+ megawatt-level power to match all capabilities of a human brain. Artificial intelligence systems which rely on computers follow Von-Neumann architecture where memory storage and information processing operations are siloed in separate physical spaces requiring transfer of information to-and-from memory into a separate processing space to actually perform useful operations whereas biological intelligence system follows a very different architecture where memory storage and information processing can occur in the same physical space which is considered to be the leading cause of the energy efficiency difference between the two systems. New computing paradigms like neuromorphic modelling are being explored which move away from traditional Von-Neumann architecture of computing and instead tries to closely emulate the architecture of the human brain requiring extremely challenging algorithmic breakthroughs to successfully unlock energy efficient artificial intelligence. A high-level mapping of human brain architecture is extremely useful to help understand end-to-end information flow within the brain and how exactly do disparate pieces connect together to produce what we see in our day-to-day life as a fully functional human-being. Human brain can be crudely divided into 3 segments from an evolutionary neurobiology standpoint (which I have covered previously in another article ‘prisoners to old brain’) i.e. neocortex, limbic brain and reptilian brain. Neocortex is the most advanced element of the human brain responsible for the very complex intelligent operations and real-time conscious processing of information whereas the limbic and reptilian brain are primitive elements of the human brain which are responsible for our unconscious instincts and all necessary backend survival operations which run on autopilot. Each of the three elements can be further broken down into sub-elements based on their individual functions. The reptilian brain consists of brain stem (controls automatic, life-sustaining autopilot functions like breathing, heart rate, temperature, and basic reflexes), basal ganglia (coordinates routine procedural habits and motor control ) and cerebellum (coordinates motor skills and physical balance). The basal ganglia acts as the gatekeeper which stores habit programs and decides which external motor controls should get executed whereas cerebellum possesses the information of each motor skill and focuses on specifically executing the required external motor controls. The limbic brain consists of amygdala (assigns emotional significance to sensory inputs, specifically identifying potential threats and triggering the physiological fight-or-flight response), hippocampus (encodes temporary, conscious experiences into stable, long-term neural representations), hypothalamus (synthesizes hormones and regulates the pituitary gland to control metabolic drives like hunger, fluid balance and circadian rhythms), thalamus (filters and pre-processes raw sensory data (excluding olfaction i.e. smell) before dispatching it to the neocortex for final processing), olfactory bulb (bypasses the thalamic filter to directly map environmental chemical stimuli into the emotional and memory centers of the brain), cingulate cortex (monitors cognitive conflict, error detection, and matches emotional distress or pleasure to physical behavioral reactions). Neocortex consists of frontal lobe (exerts top-down inhibition over lower brain drives, orchestrating working memory, future planning, and cognitive flexibility), parietal lobe (computes incoming tactile i.e. touch-sense data and constructs an objective, three-dimensional coordinate map of the body and surrounding space), temporal lobe (assigns meaning to sensory information, processes acoustic frequencies, decodes the syntax of language) and occipital lobe (deconstructs raw neural signals from the retina into basic visual components like orientation, edges, color, and motion). The reptilian brain focuses on modelling the external world in present time in terms of geometric space i.e. an environment where external entities can be computed as threats or not to ensure survival whereas limbic brain focuses on modelling the external world in present time in terms of social space i.e. an environment where elements of social cooperation can be computed to aid in survival. Neocortex focuses on modelling the future to ensure long-term survival as compared to limbic and reptilian brain which focuses on short-term survival. The difference in goals served by new brain i.e. neocortex and old brain i.e. limbic and reptilian brain results in the internal tug of war between seeking delayed and instant gratification which is usually dominated by the old brain which seeks short-term rewards as it is unable to simulate future long-term rewards. The information flow within in this neural architecture is very interesting to study. Sensory information flowing in from various sensory organs into the thalamus is sparse coded i.e. has a surprisingly low data rate and the neocortex plays a larger role in actually constructing the final model of the external world based on sensory information. For example, the input formation captured by the retina is a sparse 2D image which has to be processed by the brain to construct the final 3D model of the world which is an extremely challenging mathematical computation. The different elements of the brain are interconnected through different neuron pipelines so that they can interact with each other to ensure cohesive emergent functioning. Sensory information which flows into the thalamus is assigned an emotional value by the amygdala based on past information stored in the neocortex (long-term memory) and hippocampus (short-term memory) usually before the neocortex can construct a complete model of the input sensory information flowing into the thalamus. If the past information stored in memory signals a threat, the amygdala communicates with the hypothalamus to regulate hormone production accordingly so as to prepare our body to face the threat which is what we experience as stress. Sometimes the threat signalling performed by the amygdala could turn out to be a false alarm as the amygdala acts on sparse pre-processed information and the neocortex after post-processing can intervene by communicating with the amygdala and hypothalamus to correct and re-regulate the hormone production. Post-processed sensory information gets stored in the short term memory i.e. hippocampus for a short period of time after which a transfer process takes place where short-term memory in the hippocampus gets converted to long-term memory in the neocortex. Memory storage and recall is surprisingly an emergent process where storage of a single memory of an event is not a space-wise localized allocation of neurons for that memory as seen in file systems of computers, but a process where bits of information of a particular memory are scattered all around the physical space of the hippocampus during storage and then reconstructed from their scattered positions when they are recalled. The scattering and decentralized storage process of memory optimizes for better storage efficiency as new information gets accommodated according to the foundational information structure of the hippocampus. Direct storage of new information into long-term memory is not biologically feasible as new information is very rich in data and long-term memory is organized in a certain structure optimized for larger storage efficiency as compared to the hippocampus which would thereby cause a mismatch in data flow and compatibility overwhelming the neurons involved in long-term memory storage within the neocortex. This is why we require short-term memory storage via hippocampus which orchestrates the eventual selective transfer of information from short-term memory into the long-term memory over time while we sleep in a format compatible with the information structure of our long-term memory. The short-time memory tries to link and add new information within existing information branches in the long term memory due to which new information which cannot find a good link with existing information branches usually does not end up getting transferred into the long-term memory. This is why new information processed and stored in short-term memory in terms of first principles or foundational concepts gets registered more efficiently and easily into the long-term memory and becomes harder to erase. In a standard sequence of new information, only few prominent bits of information finally gets encoded into the long-term memory due to which our past memories are not a reliable estimator of the actual truth as bulk of the original information is lost during the transfer process into the long-term memory. This is why we usually tend to remember mostly the peak and end instances of our past experiences which I have covered in another article - ’peak-end rule error’. When these memories are recalled, our neocortex often constructs an inaccurate story to fill the gaps and explain these past memories based on the sparse underlying data points stored in the long-term memory. Amygdala plays a significant role in the memory registering process where new information gets assigned an emotional value from the amygdala and new information with high emotional value gets registered more easily and almost permanently into the long-term memory which is why we easily remember happy memories and unfortunately also remember painful memories which is what happens in the case of trauma. The frontal lobe in the neocortex is the seat of consciousness which concurrently takes into account all the different subsets of processed and stored information coming in from different sub-elements of the brain and coalesce them to construct a singular coherent conscious experience which explains external reality and guides decision making for the system as a whole. However, the old brain which includes limbic and reptilian brain can have a dominant influence on our overall functioning as they have a larger evolutionary footprint on us as compared to the neocortex and can easily override conscious control exerted by the frontal lobe which is what happens when we have emotional reactions and undertake impulsive actions. This phenomenon is common in individuals highly susceptible to emotions where the amygdala easily hijacks conscious activity in the frontal lobe making them more prone to stressed-induced reactions like panic attacks and energy-preservation states like depression. However, this works as a two-way bridge where sustained conscious activity in the frontal lobe can help downregulate emotional activity in the limbic brain and build resistance to amygdala hijacking. Another very interesting classification of the brain is the left vs right brain dichotomy where due to different neural architectures, the left brain composed of compact nested interneuron connections specializes in sequential processing of information and the right brain composed of long-range diverse interneuron connections specializes in parallel processing of information. This dual information processing mechanism is crucial to efficiently model the external world in real-time and simulate the future. Sequential processing helps in time tracking, language/text processing and algorithmic/logical thinking. Parallel processing helps in developing in broader context, visual/spatial processing and creative/lateral thinking. Visual information like text and images are input streams with high data rate (i.e. bits of information per unit time) as large chunks of information can be consumed in a single unit of time and be parallelly processed to a broader extent depending on the nature of information and past familiarity, whereas input streams like audio and to some extent video are relatively constrained in data rate as only smaller chunks of information can be consumed per unit time due to sequence and bandwidth constraints of information flow. Although text information has a high data rate like images, unfamiliar text largely requires sequential processing for decoding semantic meaning which is energy intensive in comparison to parallel processing which is why reading usually feels harder than viewing an image or watching a video even though we can consume more data per unit time through reading. Visual information like images and video largely undergo parallel processing in our brain which is more energy efficient and involves wide array of interconnected neuron firing which leads to much easier processing and registering of information into our memory. The mechanism of wide interneuron firing is an extremely powerful means of information assimilation which is why storytelling which enables a much wider range interneuron firing as compared to raw recital of facts, is a very useful skill to enable deep communication of messages. The mechanism of wide interneuron firing is also why visualization is an extremely powerful and sometimes underrated mental tool for information processing. The superpower of the neocortex truly lies in imagination i.e. the ability to simulate alternate information-rich realities separate from the outside world purely within our mind. This ability allows us to perform thought experiments i.e. useful experiments on external reality solely within our mind by accurately simulating all external conditions and interactions. The ability to perform thought experiments is what allowed Einstein to formulate the complex and counter-intuitive theories of special and general relativity prior to being able to perform a testable physical experiment in the external world. Deep visualization can allow us to create internal realities which can at times become indistinguishable from external reality which is what happens in the case of dreams. Powerful visualization which fuels creativity requires developing wide a range of very diverse interneuron connections which allows us to have new permutations and combination of interneuron firings thereby allowing us to simulate and create new possibilities which were previously inaccessible or incomprehensible. The capabilities of human mind can be further amplified by external tools like computers and more particularly artificial intelligence to develop an unparalleled and seamless symbiotic connection of broader information flow and processing. It can be argued that we are presently cyborgs to some extent with computing in the form of phones and laptops serving as our tertiary brain but we face limitations in accessing larger capabilities due restrictive flow of information. While input bandwidth is reasonably high owing to the high data rate of visual information, our output bandwidth which exists through speech or typing is relatively very low making it one of the bottlenecks for seamless human-machine symbiosis. Brain-computer interface technologies aims to solve this issue by creating a high data rate channel of information flow between our brain and computers. Solving the bandwidth constraint helps us unlock deeper efficiency while interfacing with external computing and blurs the barrier between mind and machine. The key benefit of human-machine symbiosis lies in the accessibility to larger surface area of parallel processing which is an extremely powerful and sometime underlooked tool. The leap-frog in language generative AI capabilities was a result of discovering the mechanism of attention birthed from Google Research team’s famous paper ‘Attention is All You Need’. The attention mechanism enabled computers to bypass the constraints of sequential processing of text information (which required sequential processing to maintain semantic structure) by creating a mathematical filter which mathematically added inter-text semantic meaning to the input text prior to being fed into the actual artificial intelligence model thereby allowing the model to undergoing faster parallel processing of input text without compromising semantic structure. Through seamless human-machine interfacing we can leverage the wider scale of parallel processing along with the benefits of accessibility to an almost infinite memory bank of existing information to further synthesize new and previously inaccessible information and possibilities in the world.