The next era of artificial intelligence (AI), or artificial general intelligence (AGI), will rest on neurophysiology models that emphasize neuronal group dynamics, rather than the behaviors of single neurons.
This post addresses three questions that underlie the next-era neurophysiological underpinnings supporting neural networks – the NEXT generation of neural networks modeling:
- What are the neuronal group dynamics that form the essential basis for next-gen AI?
- What are the key intra-neuron elements (e.g., soma models) that are likewise essential?
- What elements of inter-neuron message passing need to be modeled immediately for next-gen AI, and which can be safely addressed in a follow-on wave?
Neurophysiology: Important Early Works
Early neural networks – the entire fifty-year history (1974 until now, 2023 heading into 2024) – were entirely based on the McCulloch-Pitts neural model (1943) and the Rosenblatt Perceptron (1957).
Hebbian Learning
Most AI-oriented neural networks researchers and developers know about the McCulloch-Pitts neural model and the Rosenblatt Perceptron. However, there are other early neurophysiology works that are as significant.
Chief among these is the work by Donald Hebb, who is most known for his groundbreaking book, On the Organization of Behavior (1949). (While this book is available, it is a bit pricey, historical reviews by Brown (2020) and Brown et al. (2021) give very good overviews of how Hebb’s ideas evolved.)
Quoting from Brown et al. (2021), who conveniently extracted a summary from Hebb’s original work (Hebb, 1949, p. 70), “‘The general idea is an old one, that any two cells or systems of cells that are repeatedly active at the same time will tend to become ‘associated’ so that activity in one facilitates activity in the other.’”
Hebb’s work is generally regarded as one of the major influences in neurophysiology, deemed on a par with Darwin’s On the Origin of the Species (Brown 2020). Within neural networks circles, though, the notion of Hebbian learning has largely fallen out of favor. (To this author’s best knowledge, the only system to seriously employ Hebbian learning was Fukushima’s Neocognitron (1980).)
The reason that Hebbian learning has not taken greater hold within the neural networks community is that Hebbian learning gives a reason to grow connections (and increase connection strength) between two (simultaneously) active nodes, but it does not provide a means to limit or stop the growth.
That “limiting” capability comes about when we have some mechanism to tell us, “Stop! That’s enough.”
This kind of limitation is provided, for example, with stochastic gradient descent methods (e.g., backpropagation). Similarly, energy-based learning methods (e.g., contrastive divergence) provide a limiting bound on node connection weights.
These two limit-providing mechanisms, though, have already had their day. There’s been growing awareness that supervised learning methods (those that require a labeled training data set, such as backpropagation and other stochastic gradient descent methods) are limited in their scope and usefulness. Similarly, we’re seeing that energy-based methods have largely reached their fruition.
Thus, if we can find some means to provide the limits on neural connection growth, Hebb’s original ideas might still prove useful.
The trick is to know when to say “Stop!”
Edelman and Mountcastle: Neuronal Group Selection
In a nutshell, the important work by two leading thinkers, Gerald M. Edelman and Vernon B. Mountcastle, emphasizes how groups of neurons interact. This is a shift from the perspective offered by the McCulloch-Pitts neuron model and the notion of Perceptron architectures. As summarized in the Amazon review of Edelman and Mountcastle’s joint publication, The Mindful Brain (1978), “There are strong conceptual parallels between Mountcastle’s idea of cortical columns and their functional subunits and Edelman’s concept of populations of neurons functioning as processors in a brain system based on selectional rather than instructional principles.”
As a sidenote, I was just reading an MIT Technology Review article about the possibility of AI systems becoming conscious. See the postscript to this blogpost – at the very end, after the Resources and References section.
The interesting thing is that the leading thoughts on how to define an AI system as “conscious” have a lot to do with feedback loops, or as Edelman and Mountcastle describe it, “reentrant signaling.” The ideas first generated in the 1970’s are still important today.
Karl Pribram: More than the “Holographic Brain”
Prior to his death, he and his wife Katherine Neville (a well-known novelist in her own right), worked together to create a more “accessible” version of Karl’s works. It with this book, The Form Within: My Point of View, that I suggest we start reading.
Neuronal Activation Decay
Before we go on to our primary theme of neuronal group dynamics, let’s take a quick look at an emergent topic – one that is important for CORTECONs and other emerging architectures, as well as its stand-alone implementation in Liquid Time-constant Networks (LTNs). This topic is, of course, neuronal activation decay.
The important reason to look at newer approaches (based on rather old neurophysiology) for neuronal activation decay is that we need a mechanism to go beyond what I’ve referred to as the “sausage-making” nature of current neural networks.
Within the past few years, LTNs have started to gain a lot of attention (Hasani et al., 2020). These are neural networks that have a more complex internal model (model of the soma, or neural body) than is used in most neural networks. The Hasani et al. work builds work by Funahashi and Nakamura (1992, 1993), who investigated continuous-time recurrent neural networks (CT-RNNs).
Even prior to Funahashi and Nakamura’s (1992, 1993) studies, there was extensive interest in the neurophysiology of single neurons and small neural systems, as well as larger neuronal constructs. The notion of a more complex internal structure for the computational neuron is not new. Koch and Segev (1998) edited an extensive book on the subject, and their review (2000) expertly captures the highlights on neural information processing up to that point.
Early work on CORTECONs (COntent-Retentive, TEmporally-CONnected neural networks) used a simple exponential activation decay (Maren, 1993). The progression of activation decay made it possible to turn a given neuron “on” or “off,” depending on the (epsilon0, epsilon1) parameters governing the enthalpy in the CVM (cluster variation method) equation driving the overall activation patterns in the computational layer.
As expressed in Maren (1993): “The computational layer is composed of processing units which receive inputs and Gaussian noise and which also experience activation decay. The state of each processing unit is governed by a function of both its activation (due to the previously mentioned factors) and the overall drive to minimize the free energy. The process of minimizing the free energy can alter a unit’s state. To do this, the absolute value of the unit’s activation … “
Karl Friston and Active Inference
We’ve put a lot of attention on Karl Friston’s work on active inference. Rather than provide details here, we invite you to go to the Resource Compendium that summarizes a thirteen-post series on the Kullback-Leibler divergence, free energy, and variational (and active) inference.
If you want to start at the beginning of this thirteen-post series, please go to this Starting Point.
Neuronal Avalanches
Neuronal avalanches have been identified and studied as early as the 1990’s, with two important initial works by Beggs and Plenz (2003) and by Pascal Fries (2005).
Castro et al. (2020) note that “We found that the strength of global excitation needed to first trigger ignition in a subset of regions is substantially smaller for the model embedding the empirical human connectome. Furthermore, when increasing the strength of excitation, the propagation of ignition outside of this initial core–which is able to self-sustain its high activity–is way more gradual than for any of the randomised connectomes, allowing for graded control of the number of ignited regions.”
More recently, Corsi et al. (2023) in “Measuring Neuronal Avalanches to inform Brain-Computer Interfaces” (Corsi et al., 2020, updated 2023) have created a more detailed understanding. They note that “… the spreading of neuronal avalanches might be a correlate of the functional interactions among brain areas and, as such, we hypothesize that they could spread differently according to the task at hand, thereby providing a powerful and original marker to differentiate between behaviors.”
Neuronal Coherence
The whole issue of neuronal avalanches is subsumed by the broader question of how different groups of neurons interact with each other. One of the leading thoughts is the notion of neural coherence.
As described by Pascal Fries (2005), “I hypothesize that neuronal communication is mechanistically subserved by neuronal coherence. Activated neuronal groups oscillate and thereby undergo rhythmic excitability fluctuations that produce temporal windows for communication. Only coherently oscillating neuronal groups can interact effectively, because their communication windows for input and for output are open at the same times.”
The work on measuring neuronal coherence, has continued in the two decades following Fries’ original work. Myers et al. (2021) estimated the spatial reach of neuronal coherence and spike-field coherence (SFC) across three different brain regions, and found that “the strongest coherence within a 2-3 cm distance from the microelectrode arrays, potentially defining an effective range for local communication. This range was relatively consistent across brain regions, spectral frequencies, and cognitive tasks. The magnitude of coherence showed power law decay …”
Myers et al. (2021) provide an excellent literature review, and note that “Coherent FPs [field potentials] can facilitate the transfer of information across shorter and longer distances … [and that] [n]euronal coherence has been shown to support higher
order cognitive processes.” Their work with three different human subjects (undergoing surgery for epilepsy), in a different brain region for each subject, allowed them to record the extent to which the field potentials were correlated.
Effective vs. Structural Connectivity
One of the most important topics within brain research is: How does the effective (or functional) connectivity between brain regions compare with the actual (or structural) connectivity? This important question has been asked for well over two decades. Some of the earliest work was done by Olaf Sporns and colleagues, and my 2016 Brain Sciences paper overviewed work by Sporns and others up to that time (Maren, 2016).
I hadn’t found this interesting – and important – paper by Battaglia et al. (2012) when I wrote my own 2016 work. One of the reasons that this work stands out is because they show that “dynamic effective connectivity can emerge from transitions in the collective organization of coherent neural activity“
They summarize by writing: “Our results provide thus novel theoretical support to the hypothesis that dynamic effective connectivity stems from the self organization of brain rhythmic activity.“
What is particularly useful about the Battaglia et al. (2012) study is that they examine how the same kind of structural connectivity between different neural groups can yield different kinds of “effective connectivity” patterns. Specifically, they state: “Through a causal analysis of time-series of model neural activity, we show that different dynamical states generated by a same structural connectivity motif correspond to distinct effective connectivity motifs.”
They “model the oscillatory dynamics of generic brain circuits involving a small number of interacting areas (structural connectivity motifs at the mesoscopic scale).” This notion of structural (or effective / functional) motifs refers to how many specific areas are being modeled in terms of their interaction. Battaglia et al. studied motifs of size N=2 and N=3.
What makes this relevant to our own work is that, by using the Kikuchi 2-D cluster variation method (2D CVM) constructed with a series of overlapping zigzag chains, we create a system where our own “structural connectivity” is described at the same scale, i.e. N=2 (nearest-neighbor, next-nearest-neighbor) and N=3 (triplet configurations).
What this tells us is that using these very local connectivity patterns, we are able to construct a system that is potentially useful in modeling brain connectivity patterns. Just as Battaglia et al. found that they could identify a range of interesting behaviors with N=2, 3, we similarly do not need to employ more long-range connections; very local ones will suffice.
Coherence Is Not the Same as Communication
A surprising result published in 2021 reveals “an alternative mechanism where coherence is not the cause but the consequence of communication and naturally emerges because spiking activity in a sending area causes post-synaptic potentials both in the same and in other areas.”
As author Vinck summarizes, “We developed an elegant mathematical formula (shown above) to prove that coherence is a direct consequence of the anatomical connection between two brain areas and signal power. Coherence is not a prerequisite for communication.”
In their current article, they developed a new mathematical theory of coherence and communication. They show that coherence simply occurs because individual neurons are active in one area and generate (synaptic) inputs into other areas. As a result, the electric signals in separate areas become coherent in a predictable way.
They also found out that coherence is mainly determined by the sending brain area, not the receiving one, and that coherence provides little information on whether communication is actually taking place.
A Quick Digression
The question that we’re addressing here is: How far out do we need to go in order to create useful models?
That is, is it sufficient to work with nearest-neighbors, next-nearest-neighbors, and triplets? Or do we need to go further?
Fortunately, Kikuchi and Brush addressed this question in their 1967 follow up to the original Kikuchi (1952) work. They investigated a range of different basic clusters, including five different motifs based on a square lattice, and several motifs based on the triangular, or X lattice.
They found that the simplest three-element “angle” motif performed identically to the more complex complete “square” motif. In their words, “The remarkable fact is that B2, whose basic cluster is the angle, gives the same results as A4, which is based on the square.”
Neuronal Interactions Involve Different Regions of the Brain (Cerebrum and Cerebellum)
This work by McAfee et al. is interesting because it presents some insights into neuronal group behavior that are NOT necessarily associated with neuronal avalanches.
From their abstract: “We propose that the cerebellum participates in cognitive function by modulating the coherence of neuronal oscillations to optimize communications between multiple cortical structures in a task specific manner.”
In their Introduction, they state: “Experimental findings have since provided substantial support for the concept of “communication through coherence” (CTC), showing that coherence changes do indeed correlate with changes in the effectiveness of neuronal transmission, and that coherence changes occur in a task-specific manner” and “Importantly, the CTC theory describes a mechanism for flexibility in communication between neuronal groups that allows for selective information flow but does not explain the neuronal mechanism for this selectivity itself.”
Leading in to their Experimental section, they state: “In the following sections, we will review evidence that the cerebellum is essential for the coherence of cerebral gamma oscillations within a well-defined functional network, and that the cerebellar activity reflects information about cerebral oscillations within a broad range of frequencies.”
Further, within their Experimental section, they state: “The mechanism we propose here does not require synchronized oscillations between the cerebral and cerebellar cortex. We predict that that cerebellum continually monitors the phase differences between oscillations in two cerebral cortical structures to detect and correct deviations from the optimal phase difference, based on the specific task and the structures involved.”
Resources and References
Important Historical References: McCulloch-Pitts, Rosenblatt, Hebb, Fukushima
[AJM’s Note: The Edelman and Mountcastle resources are listed further down under the Reentrant Signal Processing section.]
- Brown, Richard E. 2020. “Donald O. Hebb and the Organization of Behavior: 17 years in the writing.” Molecular Brain 13:55. doi:10.1186/s13041-020-00567-8. (Accessed Nov. 18, 2023; available at https://doi.org/10.1186/s13041-020-00567-8.)
- Brown, Richard E., Thaddeus W. B. Bligh, and Jessica F. Garden. 2021. “The Hebb Synapse Before Hebb: Theories of Synaptic Function in Learning and Memory Before Hebb (1949), With a Discussion of the Long-Lost Synaptic Theory of William McDougall.” Front. Behav. Neurosci.: Review (21 Oct. 2021). doi:10.3389/fnbeh.2021.732195. (Accessed Nov. 19, 2023; available from ResearchGate.)
- Fukushima, Kunihiko. 1980. “Neocognitron: A Self-organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position.” Biol. Cybernetics 36: 193 202 (1980) (Accessed May 22, 2023; available online as https://www.rctn.org/bruno/public/papers/Fukushima1980.pdf.)
- Hebb, Donald O. 1949. The Organization of Behavior: A Neuropsychological Theory. (New York: Wiley; reprinted 2002 by Lawrence Erlbaum associates, Mahwah, New Jersey, reprinted again by Psychology Press: London; 1st edition, May 1, 2012.)
- McCulloch, Warren S. and Walter Pitts. 1943. “A Logical Calculus of the Ideas Immanent in Nervous Activity.” Bull. Mathematical Biol. 52 (l/2): 99-115. (Accessed Aug. 2, 2023; available online at https://www.cs.cmu.edu/~./epxing/Class/10715/reading/McCulloch.and.Pitts.pdf.)
- Rosenblatt, Frank. 1957. “The Perceptron: A Perceiving and Recognizing Automata (Project PARA).” Cornell Aeronautical Laboratory, Inc. Report Number 85-460-1 (January, 1957). (Accessed Aug. 2, 2023; available online at https://blogs.umass.edu/brain-wars/files/2016/03/rosenblatt-1957.pdf.)
- Rosenblatt, Frank. 1962. Principles of Neurodynamics. (New York: Spartan).
Neural Activation Decay: Neurophysiology, CORTECONs, LTNs
- Funahashi, Ken-ichi and Yuichi Nakamura. 1992. “Neural Networks, Approximation Theory, and Dynamical Systems.” pdf.)
- Funahashi, Ken-ichi and Yuichi Nakamura. 1993. “Approximation of dynamical systems by continuous time recurrent neural networks.” Neural Networks 6(6): 801-806. (Abstract only available for open-access; abstract.)
- Hasani, Ramin, Mathias Lechner, Alexander Amini, Daniela Rus, and Radu Grosu. 2020. “Liquid Time-constant Networks.” arXiv:2006.04439v4 [cs.LG] 14 Dec 2020. (Accessed Nov. 19, 2023; pdf.)
- Koch, Christof, and Idan Segev, eds. 1998. Methods in Neuronal Modeling: From Ions to Networks. 2nd Edition. (Cambridge: MIT Press.)
- Koch, Christof, and Idan Segev. 2020. “The Role of Single Neurons in Information Processing.” Nat Neurosci 3 (Suppl 11): 1171–1177. doi:10.1038/81444. (Accessed Nov. 19, 2023; available on ResearchGate.)
- Maren, A.J. (1993). Free energy as driving function in neural networks. 1993 Symposium on Nonlinear Theory and Its Applications (Hawaii: Dec. 5-10, 1993). DOI 10.13140/2.1.1621.1529. (Accessed Nov. 19, 2023; pdf.)
Reentrant Signal Processing: Edelman et al., Sporns
- Edelman, Gerald. Neural Darwinism: The Theory Of Neuronal Group Selection. Republished by Basic Books; First Edition (December 6, 1987).
- Edelman, Gerald M. and Vernon B. Mountcastle. 1978. The Mindful Brain: Cortical Organization and the Group-Selective Theory of Higher Brain Function. (Reprinted 1982; Cambridge: MIT Press.)
- Edelman, Gerald M. and Joseph A. Gally. 2013. “Reentry: A Mechanism for Integration of Brain Function.” Front. Integr. Neurosci. (2013 Aug 27;7:63). doi:10.3389/fnint.2013.00063. PMID: 23986665; PMCID: PMC3753453. (Accessed Nov. 22, 2023, pdf.)
- Sporns, Olaf. 2013. “Structure and Function of Complex Brain Networks.” Dialogues Clin Neurosci. 2013 Sep;15(3):247-62. doi: 10.31887/DCNS.2013.15.3/osporns. PMID: 24174898; PMCID: PMC3811098. (Accessed Nov. 22, pdf.)
Karl Pribram: Holographic Brain and More
AJM’s Note: I loved Karl. Totally, totally! I would sit in his classes and be enthralled. His articles and books, however …
Prior to his death, he and his wife Katherine Neville (a well-known novelist in her own right), worked together to create a more “accessible” version of Karl’s works. It is HERE that I suggest we start reading.
- Pribram, Karl. 2013. The Form Within: My Point of View. (Westport, CT: Prospecta Press.) (Available for purchase on Amazon.)
Brain Region Connectivity (A Partial Review)
AJM’s Note: I wrote this paper for a Brain Sciences Special Issue on Brain-Computer Interfaces: Current Trends and Novel Applications, and investigated brain-computer interface studies carried out up to 2016. Key works were done by Olaf Sporns and colleagues, as well as Gustav Deco and colleagues. This 2016 paper gives a fairly decent (although limited) review.
- Maren, Alianna J. 2016. “The Cluster Variation Method: A Primer for Neuroscientists.” Brain Sci. 6(4):44; doi:10.3390/brainsci6040044. (Accessed Nov. 16, 2023; https://doi.org/10.3390/brainsci6040044.)
2-D Cluster Variation Method: The Earliest Works (Historical References; Theory Only)
- Kikuchi, R. (1951). A theory of cooperative phenomena. Phys. Rev. 81, 988-1003, pdf, accessed 2018/09/17.
- Kikuchi, R., & Brush, S.G. (1967), “Improvement of the Cluster‐Variation Method,” J. Chem. Phys. 47, 195; online as: online – for purchase through American Inst. Physics. Costs $30.00 for non-members.
Neuronal Avalanches
- Beggs, John M. and Dietmar Plenz. 2003. “Neuronal Avalanches in Neocortical Circuits.” J Neurosci. (2003 Dec. 3) 23(35):11167-77. doi:10.1523/JNEUROSCI.23-35-11167.2003.
- Fries, Pascal. 2005. “A Mechanism for Cognitive Dynamics: Neuronal Communication through Neuronal Coherence.” Trends in Cognitive Science 9(10), 474-480. doi:10.1016/j.tics.2005.08.011.
- Battaglia, Demian, Annette Witt , Fred Wolf, and Theo Geisel. 2012. “Dynamic Effective Connectivity of Inter-Areal Brain Circuits.” PLoS Comput Biol 8(3): e1002438. doi:10.1371/journal.pcbi.1002438
- Castro Samy, Wael El-Deredy, Demian Battaglia, and Patricio Orio. 2020. “Cortical Ignition Dynamics is Tightly Linked to the Core Organisation of the Human Connectome. PLOS Computational Biology 16(7): e1007686. (Accessed Nov. 22, 2023; journal.)
AJM’s Note: Recent work on “long-range phase synchronization” by Arnulfo et al. (2020) challenges some existing notions on long-range correlations in the brain.
- Arnulfo, G., Wang, S.H., Myrov, V. et al. 2020. “Long-Range Phase Synchronization of High-Frequency Oscillations in Human Cortex.” Nat Commun 11, 5363. doi:10.1038/s41467-020-18975-8. (Accessed Nov. 22, 2023; pdf.)
- Myers, John C., Elliot H. Smith, Marcin Leszczynski, James O’Sullivan, Mark Yates, Guy McKhann II, Nima Mesgarani, Charles Schroeder, Catherine Schevon, and Sameer A. Sheth. 2021. “The Spatial Reach of Neuronal Coherence and Spike-field Coupling across the Human Neocortex.” bioRxiv, December 9, 2021. doi:10.1101/2021.12.07.471617. (Accessed Nov. 22, 2023; pdf.)
AJM’s Note: This is the surprise work that has challenged the notion that coherence equaled communication.
- Schneider, Marius, Ana Clara Broggini , Benjamin Dann, Athanasia Tzanou, Cem Uran, Swathi Sheshadri, Hansjörg Scherberger, and Martin Vinck. 2022. “A Mechanism for Inter-Areal Coherence through Communication Based on Connectivity and Oscillatory Power.” Neuron 109(24), 4050-4067.e12. doi:10.1016/.j.neuron.2021.09.037. (Accessed Nov. 22, 2023; journal.)
- Corsi, Marie-Constance, Pierpaolo Sorrentino, Denis Schwartz, Nathalie George, Leonardo L. Gollo, Sylvain Chevallier, Laurent Hugueville, Ari E. Kahn, Sophie Dupont, Danielle S. Bassett, Viktor Jirsa, and Fabrizio De Vico Fallani. 2020, updated 2023. “Measuring Neuronal Avalanches to inform Brain-Computer Interfaces.” bioRXiv, updated version on June 19, 2023. (Accessed Nov. 15, 2023 at https://doi.org/10.1101/2022.06.14.495887 .)
Communication and Coordination across Major Brain Regions (Cerebellum plus Cerebrum)
AJM’s Note: This McAfee et al. paper describes how the cerebrum is involved in cognitive activities largely conducted within the cerebral cortex. This paper is worth a re-read, possibly several re-reads – it is dense in the neurophysiology of the brain, so will be a bit of a slow go, but worthwhile.
- McAfee, Samuel S., Yu Liu, Roy V. Sillitoe, and Detlef H. Heck. 2022. “Cerebellar Coordination of Neuronal Communication in Cerebral Cortex.” Front. Syst. Neurosci. (11 January 2022) 15: 2021. doi: 10.3389/fnsys.2021.78527. (Accessed Nov. 15, 2023 at https://doi.org/10.3389/fnsys.2021.781527.)
Parallel Processes in the Brain
AJM’s Note: I’m copying this in from a prior post; one of the most important things in brain processes is the existence of parallel processing pathways; this article gives recent results.
Here’s a layman’s overview:
- Sheedy, Carolyn. 2023. “Neuroscientists Gain New Understanding of Neural Pathway.” Carnegie Mellon Institute Neuroscience News (May 22, 2023). (Accessed May 17, 2023; available online at https://www.cmu.edu/ni/news/neuroscientists-gain-new-understanding-of-neural-pathway.html.)
Here’s the original research article:
- Isett, Brian R., Katrina P. Nguyen, Jenna C. Schwenk, Jeff R. Yurek, Christen N. Snyder, Maxime V. Vounnatsos, , Kendra A. Adegbesan, Ugne Ziausyte, and Aryn H. Gittis. 2023. “The indirect pathway of the basal ganglia promotes transient punishment but not motor suppression.” Neuron (May 16, 2023) S0896-6273(23)00302-1.. doi: 10.1016/j.neuron.2023.04.017. (Accessed May 17, 2023; available online at https://www.biorxiv.org/content/10.1101/2022.05.18.492478v1.full.)
Conscious AI: Current and Very Early Thoughts
Most of the time, when people think about AI systems becoming conscious, their first step is to invoke Turing’s test for true machine intelligence.
Recently, Grace Huckins published an article in the MIT Technology Review on what would be needed for an AI to be seriously considered as “conscious.” She emphasized experimental work done by Liad Mudrik, a neuroscientist at Tel Aviv University. She also notes a requirement for consciousness as proposed by Mudrik.
Mudrik and her colleagues have managed to establish some concrete facts about how consciousness works in the human brain … feedback connections—for example, connections running from the “higher,” cognitive regions of the brain to those involved in more basic sensory processing—seem essential to consciousness. (This, by the way, is one good reason to doubt the consciousness of LLMs: they lack substantial feedback connections.)” [AJM’s note: bold & italics are mine.]
Huckins, Grace. 2023. “Minds of Machines: The Great AI Consciousness Conundrum.” See full citation below.
So, if feedback loops are essential for consciousness, then the progenitor for this line of thinking would be the early works of Edelman and Mountcastle. (See citations in Resources and References, above.)
From Francis Schmitt’s Introduction to Edelman and Mountcastle’s The Mindful Brain comes this perspective on consciousness:
Consciousness, it is hypothesized, may result from reentrant signaling that involves associations between current sensory input and stored patterns of neuronal groups. Details are specified about the stages in which reentrant signals are processed in relation to responses in primary and secondary repertoires.
“It is important to emphasize that the concept of reentry is a critical one. Because of the degenerate nature of the proposed selection process, the absence of reentry would lead to a failure of continuity in the system as well as a failure to form coordinated abstract representation of external signals. In other words, reentry guarantees continuity in a distributed selectional system. Consciousness may be a kind of associative recollective updating by reentrant inputs that continually confirms or alters the theory of the self by parallel sensory or motor inputs and outputs.”
Francis Schmitt’s Introduction to Edelman and Mountcastle’s The Mindful Brain. See full citation in Resources and References (above).
Thus, we’re seeing that the notion of feedback loops – from as far back as the 1970’s – is of crucial importance in how we view the prospect of consciousness in AI systems today.
- Huckins, Grace. 2023. “Minds of Machines: The Great AI Consciousness Conundrum.” MIT Technology Review. (October 16, 2023). (Accessed Nov. 19, 2023; available at MIT Tech Rev.)
2 comments
Comments are closed.