In 2022, Blumberg explicitly described a neural cell learning and predicting its own responses. The 2024 preprint and 2026 published version by Premakumar, Graziano and coauthors implement prediction of internal network activity. These cards show the shared roles, the dated wording, and the limits of each partial comparison.
Source-review edition: These qualified partial comparisons were checked through an independent AI-assisted source review. They do not establish whole-theory equivalence, first-publication priority or influence.
Card 1 of 32
SAN proposes learning and predicting the cell’s own responses
Micah Blumberg / SAN (2022-09-18) → Premakumar, Graziano and coauthors (2024-07-14). The shared role is learning a model of the neural system’s own activity, stated explicitly in the 2022 source before the 2024/2026 papers.
moderate role-level partial
Micah Blumberg / SAN
2022-09-18 · historical Git wording record
My intuition is that the apical is learning its own cell's response, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about its own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
A qualitative neural self-model is not the target’s numerical prediction of selected current hidden activations in one forward pass. This does not establish the complete estimator, future-state semantics or consciousness.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN proposes learning and predicting the cell’s own responses
Micah Blumberg / SAN (2022-08-30) → Premakumar, Graziano and coauthors (2024-07-14). The shared role is learning a model of the neural system’s own activity, stated explicitly in the 2022 source before the 2024/2026 papers.
moderate role-level partial
Micah Blumberg / SAN
2022-08-30 · historical Git wording record
My intuition is that the apical is learning it's responds, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about it's own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
A qualitative neural self-model is not the target’s numerical prediction of selected current hidden activations in one forward pass. This does not establish the complete estimator, future-state semantics or consciousness.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
Micah Blumberg / SAN (2022-06-12) → Premakumar, Graziano and coauthors (2024-07-14). The note models the system’s own thought activity; the later network predicts its own internal activations.
weak-to-moderate conceptual partial
Micah Blumberg / SAN
2022-06-12 · historical Git wording record
the observer is the thought pattern making a thought pattern of the thought pattern
The2012header is owner-reported, not independently authenticated publication. The2022Git object fixes this wording. Metaphysical, meditation and machine-consciousness assertions elsewhere in the note are not adopted as findings. The Numenta reference remains visible.
Reported note date: 2012-09-07. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
In addition to performing a primary task, the networks were responsible for predicting the values of subsets of their own hidden activations.
The network predicts subsets of its own hidden activations.
A qualitative neural self-model is not the target’s numerical prediction of selected current hidden activations in one forward pass. This does not establish the complete estimator, future-state semantics or consciousness.
The owner-reported 2012 header is separate from the 2022 wording record.
Micah Blumberg / SAN (2022-06-08) → Premakumar, Graziano and coauthors (2024-07-14). Self-generated activity returns to the neural system; the target predicts selected internally generated activations.
weak-to-moderate partial
Micah Blumberg / SAN
2022-06-08 · historical Git wording record
Who listens to the song and understands the meaning? the song is a loop, the neurons that play it also listen to it
The2012header is owner-reported. The2022retrospective rendering comment at line3 is not backdated into2012. Hawkins is explicitly recommended and the Pinker quotation is separately attributed. The instrument/song description is an analogy, not proof that every neuron hears a literal sound or of the source's other numerical/physical claims.
Reported note date: 2012-04-19. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
In addition to performing a primary task, the networks were responsible for predicting the values of subsets of their own hidden activations.
The network predicts subsets of its own hidden activations.
A qualitative neural self-model is not the target’s numerical prediction of selected current hidden activations in one forward pass. This does not establish the complete estimator, future-state semantics or consciousness.
The owner-reported 2012 header is separate from the 2022 wording record.
SAN proposes learning from the cell’s own activity
Micah Blumberg / SAN (2022-09-18) → Premakumar, Graziano and coauthors (2024-07-14). Learning about internally generated activity accompanies ordinary signal processing; the target implements this additional focus as an auxiliary task.
weak-to-moderate component partial
Micah Blumberg / SAN
2022-09-18 · historical Git wording record
An interesting thought about the cyclic nature of Layer 5 pyramidal cells, is the idea that the process of the dendrite being a sensor, like an eyeball or an ear appears to apply to both sides of the neuron, on the way in, and on the way out. So the Apical dendrite, as a neural network, is learning from incoming sensory stimulus, but the Soma burst is one of the incoming sensory stimuli, with models of soma burst waveforms becoming part of what the Apical Dendrite has to contend with.
Reading or learning from internal activity is not an explicitly specified additional output head, numerical auxiliary regression loss or training schedule. The shared role is preserved without identifying the full implementation.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN proposes learning from the cell’s own activity
Micah Blumberg / SAN (2022-08-30) → Premakumar, Graziano and coauthors (2024-07-14). Learning about internally generated activity accompanies ordinary signal processing; the target implements this additional focus as an auxiliary task.
weak-to-moderate component partial
Micah Blumberg / SAN
2022-08-30 · historical Git wording record
An interesting thought about the cyclic nature of Layer 5 pyramidal cells, is the idea that the process of the dendrite being a sensor, like an eyeball or an ear appears to apply to both sides of the neuron, on the way in, and on the way out. So the Apical dendrite, as a neural network, is learning from incoming sensor stimulus, but the Soma burst is one of the incoming sensor stimulai, with models of soma burst waveforms becoming part of what the Apical Dendrite has to contend with.
Reading or learning from internal activity is not an explicitly specified additional output head, numerical auxiliary regression loss or training schedule. The shared role is preserved without identifying the full implementation.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN: self-representation within sensing and action
Micah Blumberg / SAN (2022-06-12) → Premakumar, Graziano and coauthors (2024-07-14). Both accounts include an internally directed modelling role inside the functioning network.
weak conceptual/architecture partial
Micah Blumberg / SAN
2022-06-12 · historical Git wording record
the observer is the thought pattern making a thought pattern of the thought pattern
The2012header is owner-reported, not independently authenticated publication. The2022Git object fixes this wording. Metaphysical, meditation and machine-consciousness assertions elsewhere in the note are not adopted as findings. The Numenta reference remains visible.
Reported note date: 2012-09-07. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
we augmented the output layer to predict both the class probabilities and the activations of specific target layers within the network.
Self-modeling is implemented as an auxiliary regression task.
Reading or learning from internal activity is not an explicitly specified additional output head, numerical auxiliary regression loss or training schedule. The shared role is preserved without identifying the full implementation.
The owner-reported 2012 header is separate from the 2022 wording record.
SAN: a neural representation is read by another layer
Micah Blumberg / SAN (2022-06-08) → Premakumar, Graziano and coauthors (2024-07-14). An internal representation becomes input to further neural processing; the target adds a specific internal regression task.
weak component partial
Micah Blumberg / SAN
2022-06-08 · historical Git wording record
Comment on the note: What I think I was missing at this point in time (2012) was the part about how a memory-prediction was actually a rendering that a subsequent array or layer of nerve cells would read & interpret.
The2012header is owner-reported. The2022retrospective rendering comment at line3 is not backdated into2012. Hawkins is explicitly recommended and the Pinker quotation is separately attributed. The instrument/song description is an analogy, not proof that every neuron hears a literal sound or of the source's other numerical/physical claims.
Reported note date: 2012-04-19. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
we augmented the output layer to predict both the class probabilities and the activations of specific target layers within the network.
Self-modeling is implemented as an auxiliary regression task.
Reading or learning from internal activity is not an explicitly specified additional output head, numerical auxiliary regression loss or training schedule. The shared role is preserved without identifying the full implementation.
Line 3 is a retrospective comment fixed in 2022, not a verified 2012 statement.
SAN links response learning with learning about responses
Micah Blumberg / SAN (2022-09-18) → Premakumar, Graziano and coauthors (2024-07-14). The earlier source joins two learning roles; the later system trains a primary task and an internal-prediction task together.
weak-to-moderate architecture partial
Micah Blumberg / SAN
2022-09-18 · historical Git wording record
My intuition is that the apical is learning its own cell's response, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about its own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
The source says "My intuition" and describes feedback-linked cell responses. Anatomical speculation is not stripped from the source context.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
The predictions and targets were passed to their respective loss functions and models were trained jointly on these two tasks.
The two tasks are classification and prediction of the network's own selected hidden activations. The paper uses additive weighted classification and regression losses.
The source does not specify the target’s weighted cross-entropy-plus-regression objective, optimizer, targets or gradient computation. Coupled learning roles are not mathematical identity of the loss.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN links response learning with learning about responses
Micah Blumberg / SAN (2022-08-30) → Premakumar, Graziano and coauthors (2024-07-14). The earlier source joins two learning roles; the later system trains a primary task and an internal-prediction task together.
weak-to-moderate architecture partial
Micah Blumberg / SAN
2022-08-30 · historical Git wording record
My intuition is that the apical is learning it's responds, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about it's own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
Tentative basal/apical/exit-terminal account. This is learning a response mapping, not an explicit classification-plus-regression objective.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
The predictions and targets were passed to their respective loss functions and models were trained jointly on these two tasks.
The two tasks are classification and prediction of the network's own selected hidden activations. The paper uses additive weighted classification and regression losses.
The source does not specify the target’s weighted cross-entropy-plus-regression objective, optimizer, targets or gradient computation. Coupled learning roles are not mathematical identity of the loss.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN: self-representation alongside sensing and action
Micah Blumberg / SAN (2022-06-12) → Premakumar, Graziano and coauthors (2024-07-14). Primary engagement and internal self-representation coexist in one system; the target implements an explicit paired learning objective.
weak architecture partial
Micah Blumberg / SAN
2022-06-12 · historical Git wording record
the observer is the thought pattern making a thought pattern of the thought pattern
The2012header is owner-reported, not independently authenticated publication. The2022Git object fixes this wording. Metaphysical, meditation and machine-consciousness assertions elsewhere in the note are not adopted as findings. The Numenta reference remains visible.
Reported note date: 2012-09-07. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
The predictions and targets were passed to their respective loss functions and models were trained jointly on these two tasks.
The two tasks are classification and prediction of the network's own selected hidden activations. The paper uses additive weighted classification and regression losses.
The source does not specify the target’s weighted cross-entropy-plus-regression objective, optimizer, targets or gradient computation. Coupled learning roles are not mathematical identity of the loss.
The owner-reported 2012 header is separate from the 2022 wording record.
SAN: self-generated activity feeds a learning loop
Micah Blumberg / SAN (2022-06-08) → Premakumar, Graziano and coauthors (2024-07-14). The internally generated activity participates in learning alongside the system’s other activity; the target implements primary and self-modelling losses together.
weak architecture partial
Micah Blumberg / SAN
2022-06-08 · historical Git wording record
Being is a learning loop, like a feedback loop that learns
The2012header is owner-reported. The2022retrospective rendering comment at line3 is not backdated into2012. Hawkins is explicitly recommended and the Pinker quotation is separately attributed. The instrument/song description is an analogy, not proof that every neuron hears a literal sound or of the source's other numerical/physical claims.
Reported note date: 2012-04-19. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
The predictions and targets were passed to their respective loss functions and models were trained jointly on these two tasks.
The two tasks are classification and prediction of the network's own selected hidden activations. The paper uses additive weighted classification and regression losses.
The source does not specify the target’s weighted cross-entropy-plus-regression objective, optimizer, targets or gradient computation. Coupled learning roles are not mathematical identity of the loss.
The owner-reported 2012 header is separate from the 2022 wording record.
Micah Blumberg / SAN (2022-09-18) → Premakumar, Graziano and coauthors (2024-07-14). The source gives both response generation and self-response prediction adaptive roles; the target makes both actual and predicted internal activations trainable.
moderate role-level partial
Micah Blumberg / SAN
2022-09-18 · historical Git wording record
My intuition is that the apical is learning its own cell's response, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about its own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
The source says "My intuition" and describes feedback-linked cell responses. Anatomical speculation is not stripped from the source context.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
During training, the two could both be varied to find the best critical points under the joint loss function.
Here “the two” refers to the predicted activations and the actual hidden activations. Both depend on trainable model weights; the target activity is not held permanently fixed. This explanation is a paraphrase, not an alteration of the quotation.
The target lets both actual and predicted hidden activations vary under its joint loss. The source supplies only the stated adaptive component or qualitative coupled-role account, not the same shared-parameter optimization or measured complexity reduction.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
Micah Blumberg / SAN (2022-08-30) → Premakumar, Graziano and coauthors (2024-07-14). The source gives both response generation and self-response prediction adaptive roles; the target makes both actual and predicted internal activations trainable.
moderate role-level partial
Micah Blumberg / SAN
2022-08-30 · historical Git wording record
My intuition is that the apical is learning it's responds, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about it's own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
Tentative basal/apical/exit-terminal account. This is learning a response mapping, not an explicit classification-plus-regression objective.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
During training, the two could both be varied to find the best critical points under the joint loss function.
Here “the two” refers to the predicted activations and the actual hidden activations. Both depend on trainable model weights; the target activity is not held permanently fixed. This explanation is a paraphrase, not an alteration of the quotation.
The target lets both actual and predicted hidden activations vary under its joint loss. The source supplies only the stated adaptive component or qualitative coupled-role account, not the same shared-parameter optimization or measured complexity reduction.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
Micah Blumberg / SAN (2022-06-12) → Premakumar, Graziano and coauthors (2024-07-14). The source’s internal self-representation can develop; the target also permits its internal predictor to change during training.
weak conceptual partial
Micah Blumberg / SAN
2022-06-12 · historical Git wording record
We have a prediction of awareness, and it seems aware because this prediction is incorporating it's own development into an even higher level of awareness about itself.
The2012header is owner-reported, not independently authenticated publication. The2022Git object fixes this wording. Metaphysical, meditation and machine-consciousness assertions elsewhere in the note are not adopted as findings. The Numenta reference remains visible.
Reported note date: 2012-09-07. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
During training, the two could both be varied to find the best critical points under the joint loss function.
Here “the two” refers to the predicted activations and the actual hidden activations. Both depend on trainable model weights; the target activity is not held permanently fixed. This explanation is a paraphrase, not an alteration of the quotation.
The target lets both actual and predicted hidden activations vary under its joint loss. The source supplies only the stated adaptive component or qualitative coupled-role account, not the same shared-parameter optimization or measured complexity reduction.
Only adaptive self-representation is mapped. The source is not used to assert a separately trained generator.
Micah Blumberg / SAN (2022-06-08) → Premakumar, Graziano and coauthors (2024-07-14). The source describes adaptable response generation and a learning self-loop; the target allows both actual activations and their predictions to change.
weak-to-moderate partial
Micah Blumberg / SAN
2022-06-08 · historical Git wording record
right so as you sense something, an electric pattern is traveling up the neo cortex, and the statistically novel parts leave a bigger impact, so your memory-prediction is the learn response to that sensory-electrical impact, it's the tissue attempting to be ready for that impact the next time it comes, that's how memory-prediction works on the small scale.
The2012header is owner-reported. The2022retrospective rendering comment at line3 is not backdated into2012. Hawkins is explicitly recommended and the Pinker quotation is separately attributed. The instrument/song description is an analogy, not proof that every neuron hears a literal sound or of the source's other numerical/physical claims.
Reported note date: 2012-04-19. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2024-07-14 · arXiv v1 submission
During training, the two could both be varied to find the best critical points under the joint loss function.
Here “the two” refers to the predicted activations and the actual hidden activations. Both depend on trainable model weights; the target activity is not held permanently fixed. This explanation is a paraphrase, not an alteration of the quotation.
The target lets both actual and predicted hidden activations vary under its joint loss. The source supplies only the stated adaptive component or qualitative coupled-role account, not the same shared-parameter optimization or measured complexity reduction.
The owner-reported 2012 header is separate from the 2022 wording record.
SAN proposes learning and predicting the cell’s own responses
Micah Blumberg / SAN (2022-09-18) → Premakumar, Graziano and coauthors (2026-05-14). The shared role is learning a model of the neural system’s own activity, stated explicitly in the 2022 source before the 2024/2026 papers.
moderate role-level partial
Micah Blumberg / SAN
2022-09-18 · historical Git wording record
My intuition is that the apical is learning its own cell's response, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about its own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
A qualitative neural self-model is not the target’s numerical prediction of selected current hidden activations in one forward pass. This does not establish the complete estimator, future-state semantics or consciousness.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN proposes learning and predicting the cell’s own responses
Micah Blumberg / SAN (2022-08-30) → Premakumar, Graziano and coauthors (2026-05-14). The shared role is learning a model of the neural system’s own activity, stated explicitly in the 2022 source before the 2024/2026 papers.
moderate role-level partial
Micah Blumberg / SAN
2022-08-30 · historical Git wording record
My intuition is that the apical is learning it's responds, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about it's own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
A qualitative neural self-model is not the target’s numerical prediction of selected current hidden activations in one forward pass. This does not establish the complete estimator, future-state semantics or consciousness.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
Micah Blumberg / SAN (2022-06-12) → Premakumar, Graziano and coauthors (2026-05-14). The note models the system’s own thought activity; the later network predicts its own internal activations.
weak-to-moderate conceptual partial
Micah Blumberg / SAN
2022-06-12 · historical Git wording record
the observer is the thought pattern making a thought pattern of the thought pattern
The2012header is owner-reported, not independently authenticated publication. The2022Git object fixes this wording. Metaphysical, meditation and machine-consciousness assertions elsewhere in the note are not adopted as findings. The Numenta reference remains visible.
Reported note date: 2012-09-07. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2026-05-14 · published article
In addition to performing a primary task, the networks were responsible for predicting the values of subsets of their own hidden activations.
An artificial network predicts its internal states.
A qualitative neural self-model is not the target’s numerical prediction of selected current hidden activations in one forward pass. This does not establish the complete estimator, future-state semantics or consciousness.
The owner-reported 2012 header is separate from the 2022 wording record.
Micah Blumberg / SAN (2022-06-08) → Premakumar, Graziano and coauthors (2026-05-14). Self-generated activity returns to the neural system; the target predicts selected internally generated activations.
weak-to-moderate partial
Micah Blumberg / SAN
2022-06-08 · historical Git wording record
Who listens to the song and understands the meaning? the song is a loop, the neurons that play it also listen to it
The2012header is owner-reported. The2022retrospective rendering comment at line3 is not backdated into2012. Hawkins is explicitly recommended and the Pinker quotation is separately attributed. The instrument/song description is an analogy, not proof that every neuron hears a literal sound or of the source's other numerical/physical claims.
Reported note date: 2012-04-19. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2026-05-14 · published article
In addition to performing a primary task, the networks were responsible for predicting the values of subsets of their own hidden activations.
An artificial network predicts its internal states.
A qualitative neural self-model is not the target’s numerical prediction of selected current hidden activations in one forward pass. This does not establish the complete estimator, future-state semantics or consciousness.
The owner-reported 2012 header is separate from the 2022 wording record.
SAN proposes learning from the cell’s own activity
Micah Blumberg / SAN (2022-09-18) → Premakumar, Graziano and coauthors (2026-05-14). Learning about internally generated activity accompanies ordinary signal processing; the target implements this additional focus as an auxiliary task.
weak-to-moderate component partial
Micah Blumberg / SAN
2022-09-18 · historical Git wording record
An interesting thought about the cyclic nature of Layer 5 pyramidal cells, is the idea that the process of the dendrite being a sensor, like an eyeball or an ear appears to apply to both sides of the neuron, on the way in, and on the way out. So the Apical dendrite, as a neural network, is learning from incoming sensory stimulus, but the Soma burst is one of the incoming sensory stimuli, with models of soma burst waveforms becoming part of what the Apical Dendrite has to contend with.
Reading or learning from internal activity is not an explicitly specified additional output head, numerical auxiliary regression loss or training schedule. The shared role is preserved without identifying the full implementation.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN proposes learning from the cell’s own activity
Micah Blumberg / SAN (2022-08-30) → Premakumar, Graziano and coauthors (2026-05-14). Learning about internally generated activity accompanies ordinary signal processing; the target implements this additional focus as an auxiliary task.
weak-to-moderate component partial
Micah Blumberg / SAN
2022-08-30 · historical Git wording record
An interesting thought about the cyclic nature of Layer 5 pyramidal cells, is the idea that the process of the dendrite being a sensor, like an eyeball or an ear appears to apply to both sides of the neuron, on the way in, and on the way out. So the Apical dendrite, as a neural network, is learning from incoming sensor stimulus, but the Soma burst is one of the incoming sensor stimulai, with models of soma burst waveforms becoming part of what the Apical Dendrite has to contend with.
Reading or learning from internal activity is not an explicitly specified additional output head, numerical auxiliary regression loss or training schedule. The shared role is preserved without identifying the full implementation.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN: self-representation within sensing and action
Micah Blumberg / SAN (2022-06-12) → Premakumar, Graziano and coauthors (2026-05-14). Both accounts include an internally directed modelling role inside the functioning network.
weak conceptual/architecture partial
Micah Blumberg / SAN
2022-06-12 · historical Git wording record
the observer is the thought pattern making a thought pattern of the thought pattern
The2012header is owner-reported, not independently authenticated publication. The2022Git object fixes this wording. Metaphysical, meditation and machine-consciousness assertions elsewhere in the note are not adopted as findings. The Numenta reference remains visible.
Reported note date: 2012-09-07. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2026-05-14 · published article
we augmented the output layer to predict both the class probabilities and the activations of specific target layers within the network.
Internal-state prediction is trained as an auxiliary task.
Reading or learning from internal activity is not an explicitly specified additional output head, numerical auxiliary regression loss or training schedule. The shared role is preserved without identifying the full implementation.
The owner-reported 2012 header is separate from the 2022 wording record.
SAN: a neural representation is read by another layer
Micah Blumberg / SAN (2022-06-08) → Premakumar, Graziano and coauthors (2026-05-14). An internal representation becomes input to further neural processing; the target adds a specific internal regression task.
weak component partial
Micah Blumberg / SAN
2022-06-08 · historical Git wording record
Comment on the note: What I think I was missing at this point in time (2012) was the part about how a memory-prediction was actually a rendering that a subsequent array or layer of nerve cells would read & interpret.
The2012header is owner-reported. The2022retrospective rendering comment at line3 is not backdated into2012. Hawkins is explicitly recommended and the Pinker quotation is separately attributed. The instrument/song description is an analogy, not proof that every neuron hears a literal sound or of the source's other numerical/physical claims.
Reported note date: 2012-04-19. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2026-05-14 · published article
we augmented the output layer to predict both the class probabilities and the activations of specific target layers within the network.
Internal-state prediction is trained as an auxiliary task.
Reading or learning from internal activity is not an explicitly specified additional output head, numerical auxiliary regression loss or training schedule. The shared role is preserved without identifying the full implementation.
Line 3 is a retrospective comment fixed in 2022, not a verified 2012 statement.
SAN links response learning with learning about responses
Micah Blumberg / SAN (2022-09-18) → Premakumar, Graziano and coauthors (2026-05-14). The earlier source joins two learning roles; the later system trains a primary task and an internal-prediction task together.
weak-to-moderate architecture partial
Micah Blumberg / SAN
2022-09-18 · historical Git wording record
My intuition is that the apical is learning its own cell's response, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about its own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
The source says "My intuition" and describes feedback-linked cell responses. Anatomical speculation is not stripped from the source context.
Premakumar, Graziano and coauthors
2026-05-14 · published article
The predictions and targets were passed to their respective loss functions, and models were trained jointly on these two tasks.
The two tasks are classification and prediction of the network's own selected hidden activations. The paper uses additive weighted classification and regression losses.
The source does not specify the target’s weighted cross-entropy-plus-regression objective, optimizer, targets or gradient computation. Coupled learning roles are not mathematical identity of the loss.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN links response learning with learning about responses
Micah Blumberg / SAN (2022-08-30) → Premakumar, Graziano and coauthors (2026-05-14). The earlier source joins two learning roles; the later system trains a primary task and an internal-prediction task together.
weak-to-moderate architecture partial
Micah Blumberg / SAN
2022-08-30 · historical Git wording record
My intuition is that the apical is learning it's responds, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about it's own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
Tentative basal/apical/exit-terminal account. This is learning a response mapping, not an explicit classification-plus-regression objective.
Premakumar, Graziano and coauthors
2026-05-14 · published article
The predictions and targets were passed to their respective loss functions, and models were trained jointly on these two tasks.
The two tasks are classification and prediction of the network's own selected hidden activations. The paper uses additive weighted classification and regression losses.
The source does not specify the target’s weighted cross-entropy-plus-regression objective, optimizer, targets or gradient computation. Coupled learning roles are not mathematical identity of the loss.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
SAN: self-representation alongside sensing and action
Micah Blumberg / SAN (2022-06-12) → Premakumar, Graziano and coauthors (2026-05-14). Primary engagement and internal self-representation coexist in one system; the target implements an explicit paired learning objective.
weak architecture partial
Micah Blumberg / SAN
2022-06-12 · historical Git wording record
the observer is the thought pattern making a thought pattern of the thought pattern
The2012header is owner-reported, not independently authenticated publication. The2022Git object fixes this wording. Metaphysical, meditation and machine-consciousness assertions elsewhere in the note are not adopted as findings. The Numenta reference remains visible.
Reported note date: 2012-09-07. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2026-05-14 · published article
The predictions and targets were passed to their respective loss functions, and models were trained jointly on these two tasks.
The two tasks are classification and prediction of the network's own selected hidden activations. The paper uses additive weighted classification and regression losses.
The source does not specify the target’s weighted cross-entropy-plus-regression objective, optimizer, targets or gradient computation. Coupled learning roles are not mathematical identity of the loss.
The owner-reported 2012 header is separate from the 2022 wording record.
SAN: self-generated activity feeds a learning loop
Micah Blumberg / SAN (2022-06-08) → Premakumar, Graziano and coauthors (2026-05-14). The internally generated activity participates in learning alongside the system’s other activity; the target implements primary and self-modelling losses together.
weak architecture partial
Micah Blumberg / SAN
2022-06-08 · historical Git wording record
Being is a learning loop, like a feedback loop that learns
The2012header is owner-reported. The2022retrospective rendering comment at line3 is not backdated into2012. Hawkins is explicitly recommended and the Pinker quotation is separately attributed. The instrument/song description is an analogy, not proof that every neuron hears a literal sound or of the source's other numerical/physical claims.
Reported note date: 2012-04-19. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2026-05-14 · published article
The predictions and targets were passed to their respective loss functions, and models were trained jointly on these two tasks.
The two tasks are classification and prediction of the network's own selected hidden activations. The paper uses additive weighted classification and regression losses.
The source does not specify the target’s weighted cross-entropy-plus-regression objective, optimizer, targets or gradient computation. Coupled learning roles are not mathematical identity of the loss.
The owner-reported 2012 header is separate from the 2022 wording record.
Micah Blumberg / SAN (2022-09-18) → Premakumar, Graziano and coauthors (2026-05-14). The source gives both response generation and self-response prediction adaptive roles; the target makes both actual and predicted internal activations trainable.
moderate role-level partial
Micah Blumberg / SAN
2022-09-18 · historical Git wording record
My intuition is that the apical is learning its own cell's response, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about its own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
The source says "My intuition" and describes feedback-linked cell responses. Anatomical speculation is not stripped from the source context.
Premakumar, Graziano and coauthors
2026-05-14 · published article
During training, the two could both be varied to find the best critical points under the joint loss function.
Here “the two” refers to the predicted activations and the actual hidden activations. Both depend on trainable model weights; the target activity is not held permanently fixed. This explanation is a paraphrase, not an alteration of the quotation.
The target lets both actual and predicted hidden activations vary under its joint loss. The source supplies only the stated adaptive component or qualitative coupled-role account, not the same shared-parameter optimization or measured complexity reduction.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
Micah Blumberg / SAN (2022-08-30) → Premakumar, Graziano and coauthors (2026-05-14). The source gives both response generation and self-response prediction adaptive roles; the target makes both actual and predicted internal activations trainable.
moderate role-level partial
Micah Blumberg / SAN
2022-08-30 · historical Git wording record
My intuition is that the apical is learning it's responds, and through feedback loops, with itself and with other pyramidal neurons the cell is going to make predictions about it's own patterns of causation, what patterns it responds & how it reacts or responds. With the Apical Dendrite perhaps symbolizing a more complex response that is connected to more complex learning. (compared to other types of neurons)
Tentative basal/apical/exit-terminal account. This is learning a response mapping, not an explicit classification-plus-regression objective.
Premakumar, Graziano and coauthors
2026-05-14 · published article
During training, the two could both be varied to find the best critical points under the joint loss function.
Here “the two” refers to the predicted activations and the actual hidden activations. Both depend on trainable model weights; the target activity is not held permanently fixed. This explanation is a paraphrase, not an alteration of the quotation.
The target lets both actual and predicted hidden activations vary under its joint loss. The source supplies only the stated adaptive component or qualitative coupled-role account, not the same shared-parameter optimization or measured complexity reduction.
The source is tentative (My intuition) and mixes basal/apical/exit terminology; compare the learning role without endorsing every anatomical implementation claim.
Micah Blumberg / SAN (2022-06-12) → Premakumar, Graziano and coauthors (2026-05-14). The source’s internal self-representation can develop; the target also permits its internal predictor to change during training.
weak conceptual partial
Micah Blumberg / SAN
2022-06-12 · historical Git wording record
We have a prediction of awareness, and it seems aware because this prediction is incorporating it's own development into an even higher level of awareness about itself.
The2012header is owner-reported, not independently authenticated publication. The2022Git object fixes this wording. Metaphysical, meditation and machine-consciousness assertions elsewhere in the note are not adopted as findings. The Numenta reference remains visible.
Reported note date: 2012-09-07. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2026-05-14 · published article
During training, the two could both be varied to find the best critical points under the joint loss function.
Here “the two” refers to the predicted activations and the actual hidden activations. Both depend on trainable model weights; the target activity is not held permanently fixed. This explanation is a paraphrase, not an alteration of the quotation.
The target lets both actual and predicted hidden activations vary under its joint loss. The source supplies only the stated adaptive component or qualitative coupled-role account, not the same shared-parameter optimization or measured complexity reduction.
Only adaptive self-representation is mapped. The source is not used to assert a separately trained generator.
Micah Blumberg / SAN (2022-06-08) → Premakumar, Graziano and coauthors (2026-05-14). The source describes adaptable response generation and a learning self-loop; the target allows both actual activations and their predictions to change.
weak-to-moderate partial
Micah Blumberg / SAN
2022-06-08 · historical Git wording record
right so as you sense something, an electric pattern is traveling up the neo cortex, and the statistically novel parts leave a bigger impact, so your memory-prediction is the learn response to that sensory-electrical impact, it's the tissue attempting to be ready for that impact the next time it comes, that's how memory-prediction works on the small scale.
The2012header is owner-reported. The2022retrospective rendering comment at line3 is not backdated into2012. Hawkins is explicitly recommended and the Pinker quotation is separately attributed. The instrument/song description is an analogy, not proof that every neuron hears a literal sound or of the source's other numerical/physical claims.
Reported note date: 2012-04-19. This earlier public-release date is not independently authenticated here.
Premakumar, Graziano and coauthors
2026-05-14 · published article
During training, the two could both be varied to find the best critical points under the joint loss function.
Here “the two” refers to the predicted activations and the actual hidden activations. Both depend on trainable model weights; the target activity is not held permanently fixed. This explanation is a paraphrase, not an alteration of the quotation.
The target lets both actual and predicted hidden activations vary under its joint loss. The source supplies only the stated adaptive component or qualitative coupled-role account, not the same shared-parameter optimization or measured complexity reduction.
The owner-reported 2012 header is separate from the 2022 wording record.