Neurons in the brain work by the direction of their signals—electrical and chemical. Neurons do not function, for all that the brain is said to do, as being in charge, but serve at the pleasure of signals, conceptually
Whatever neurons are described as involved in—firing, wiring, activation, inhibition, hierarchy, and so forth—revolve around the signals. It is established in brain science that neurons are in clusters across the central and peripheral nervous systems. These clusters are theorized to enable the signals to collect into sets or loops.
It is these sets or loops that become how information is organized in the brain. Information includes memory, feelings, emotions and modulation of internal signals. Simply, how a memory is different from the next is because of the way that sets of signals organize one from the other. The same applies to how a memory is different from a feeling or an emotion. It is also within the sets of signals that differences in how they are graded are mechanized—like with attention, awareness, subjectivity, and free will.
There is a new paper in Nature, Abstract representations emerge in human hippocampal neurons during inference, where the authors wrote, “Here we characterized the representational geometry of populations of neurons (single units) recorded in the hippocampus, amygdala, medial frontal cortex and ventral temporal cortex of neurosurgical patients performing an inferential reasoning task. We found that only the neural representations formed in the hippocampus simultaneously encode several task variables in an abstract, or disentangled, format. This representational geometry is uniquely observed after patients learn to perform inference, and consists of disentangled directly observable and discovered latent task variables. What would be the signature of an abstract neural representation that enables this kind of adaptive behaviour? How can a neural or biological network efficiently encode many variables simultaneously? One solution is to encode variables in an abstract format so they can be re-used in new situations to facilitate generalization and compositionality.”
If geometric neural representations in the brain are basis for learning and behavior, how are neurons that are sometimes near static able to pass on their ‘representations’ from one part of the brain to the other—during processes? Simply, if neural representations are basis for learning, how do the representations relay? If the answer is that electrical signals covey these representations, how? Electrical signals have a direct [sequential] relationship with chemical signals, not with their host, neurons. Also, given the close-to-linearity with which electrical signals move, they may not be able to convey the architecture of different solid [state] representations. So, either representations are not the basis, or there is some other way that the representations [or their summaries] are made available elsewhere, for continuity in processing, conceptually.
There are several openings at the synaptic cleft, for the presynaptic and postsynaptic neurons. There are ways that the chemical signals exit the cleft. It is theorized that there is a configuration that has to be achieved for an information to be structured. Summaries of this configuration is what electrical signals convey, not neural representations, conceptually. This is easy since chemical signals are mostly fluid, not solid. Also, if learning were based on neural representations, how much dynamism do solid [state] neurons present, for all that is possible in memory, emotions and others, compared to the variety of chemical signals, and capacity of electrical signals?
Artificial neural networks [ANNs] were developed by the aerial observation of how neurons work, such that firing, wiring, activation, inhibition and others, were taken as what neurons did, which, digitally, were applicable enough, and easily tweaked but biologically, signals did those things, with neurons hosting.
Simply, the stadium is the neurons, the signals are the players and audience. As large language models [LLMs] became excellent, several studies have been exploring how to use artificial neurons to understand biological neurons. This may benefit artificial neurons, but biological neurons work for signals, while artificial neurons have no signals. It is like the possibility to know how to build a stadium from seeing one without the players and audience. However, to get something functional, a stadium that can host people has to be built to similar specifications.
There is a recent paper in APL Machine Learning, Brain-inspired learning in artificial neural networks: A review, stating that, “Backpropagation is a powerful error-driven global learning method that changes the weight of connections between neurons in a neural network to produce a desired target behavior. This is accomplished through the use of a quantitative metric (an objective function) that describes the quality of behavior given sensory information (e.g., visual input, written text, and robotic joint positions). The backpropagation algorithm consists of two phases: the forward pass and the backward pass. In the forward pass, the input is propagated through the network, and the output is calculated. During the backward pass, the error between the predicted output and the “true” output is calculated, and the gradients of the loss function with respect to the weights of the network are calculated by propagating the error backward through the network. These gradients are then used to update the weights of the network using an optimization algorithm such as stochastic gradient descent. This process is repeated for many iterations until the weights converge to a set of values that minimize the loss function.”
There is a recent paper in Nature, Inferring neural activity before plasticity as a foundation for learning beyond backpropagation, stating that, “Backpropagation, as a simple yet effective credit assignment theory, has powered notable advances in artificial intelligence since its inception and has also gained a predominant place in understanding learning in the brain. Due to this success, much recent work has focused on understanding how biological neural networks could learn in a way similar to backpropagation; although many proposed models do not implement backpropagation exactly, they nevertheless try to approximate backpropagation and much emphasis is placed on how close this approximation is. However, learning in the brain is superior to backpropagation in many critical aspects. For example, compared to the brain, backpropagation requires many more exposures to a stimulus to learn15 and suffers from catastrophic interference of newly and previously stored information. This raises the question of whether using backpropagation to understand learning in the brain should be the main focus of the field.”
It is generally accepted that backpropagation does not work in the brain. It is also accepted that predictive coding works in the brain, or that the brain makes predictions.
So, what is the difference between similar seeming backprop and predictive coding?
Electrical signals, in sets, are theorized to split, with some going ahead of others, to interact with chemical signals like they had before. Such that there is always an option for an incoming partition to go in another direction if the initial perception did not match. This second one corrects what is called the prediction error.
Simply, what is called prediction is a split of electrical signals. And while forward is the direction, not every part of the set goes forward simultaneously. Still, signals are the ones making the moves, not neurons.
There are relay versions of what might be described as backpropagation with signals, but it may not be with functions, but with the qualifiers of functions. For example, switches between attention and awareness [or less than attention] are a form of backpropagation, within an array, conceptually.
The signals of neurons, in sets, are the ways the brain works, not neural representations, conceptually. Understanding how the signals mechanize functions and grade them in sets would arrive closer to answering several unknowns about the brain than by using ANNs.
AI safety and alignment would also be defined within how sets of signals work, conceptually.
Human intelligence can be defined with sets of signals and their relays. Human, animal and machine intelligence can be described as memory maximization. Simply, intelligence means how memory is maximized, such that more than capability [language, embodiment, and so forth], what matters to how intelligent anything is, are the relays that take excellently from memory areas.
Several organisms that are close to humans are quite intelligent, even without an established language—like humans. However, several more of what is possible in their memory are not that utilized, indicating relays deficit, otherwise they would have been more formidable, even without language.
LLMs have language, but the relays [predictions] maximize much within the available language memory—that several of their outputs are similar to expert humans across fields. The human mind, which the signals are, with their interactions and graders, hold more answers in understanding the brain than with ANNs.
