Every theory should lead to testable predictions, for experimental testing is the only sure way to determine the validity of a new idea. Fortunately the memory-prediction framework is grounded in biology and leads to several specific and novel predictions that can be tested. In this appendix I list predictions that can falsify and/or support the proposals made in this book. This material is somewhat more advanced than the material in chapter 6 and is definitely not required to understand the rest of the book. Several of the predictions can only be performed on awake animals or awake human subjects because the tests involve expectation and prediction of the onset of a stimulus. The predictions are not ranked in terms of importance.

We should find cells in all areas of cortex, including primary sensory cortex, that show enhanced activity in anticipation of a sensory event, as opposed to in reaction to a sensory event.

For example, Tony Zador’s lab at Cold Spring Harbor Laboratory has found cells in rat primary auditory cortex that fire precisely when the rat expects to hear a sound even when there is no sound (private correspondence). This should be a general property of cortex. We should find similar anticipatory activity in visual cortex and somatosensory cortex. Cells that fire in anticipation of a sensory input are the definition of prediction, a basic premise of the memory-prediction framework.

The more spatially specific a prediction can be, the closer to primary sensory cortex we should find cells that become active in anticipation of an event.

If a monkey were trained on sequences of visual patterns such that it could anticipate a particular visual pattern at a precise time, we should find cells that show enhanced activity precisely when the anticipated pattern is expected (a restatement of prediction 1). If the monkey learned to expect to see a face but it didn’t know exactly what face or how the face would appear, then we should expect to find anticipatory cells in face recognition areas but not lower visual areas. However, if the monkey fixates on a target and has learned to expect a particular pattern at a precise location in its visual field then we should find anticipatory cells in V1 or close to V1. Activity representing prediction flows down the cortical hierarchy as far as it can, depending on the specificity of the prediction. Sometimes it can go all the way to primary sensory areas and other times it stops in higher regions. Similar results should exist in other sensory modalities.

Cells that exhibit enhanced activity in anticipation of sensory input should be preferentially located in cortical layers 2, 3, and 6 and the prediction should stop moving down the hierarchy in layers 2 and 3.

Predictions that travel down the cortical hierarchy do so via cells in layers 2 and 3, which then project to layer 6. These layer 6 cells project broadly across layer 1 in the region below in the hierarchy, activating another set of layer 2 and layer 3 cells and so on. Therefore cells in these layers (2, 3, and 6) are where we should find anticipatory activity. Recall that active cells in layers 2 and 3 represent a set of possible active columns; they are possible predictions. Active cells in layer 6 represent a smaller number of columns; these are the specific predictions from a region of cortex. As a prediction travels down the hierarchy the activity will eventually stop in layers 2 and 3. For example, say a rat has learned to anticipate one of two different audio tones. Based on an external cue, the rat knows when it will hear one of these two tones but it can’t predict which tone. In this scenario we should expect to see anticipatory activity in layers 2 or 3, in columns that represent both tones. There should not be activity in layer 6 of the same region because the animal can’t predict which specific tone will be heard. If on another trial the animal can predict the exact tone it will hear, then we should see activity in layer 6, in columns that respond to that specific tone.

We can’t completely rule out the possibility of finding anticipatory cells in layers 4 and 5. For example, it is likely there are several classes of cells in these layers with unknown function. Therefore, this prediction is relatively weak, but I still feel it is worth mentioning.

One class of cells in layers 2 and 3 should preferentially receive input from layer 6 cells in higher cortical regions.

Part of the memory-prediction model is that learned sequences of patterns that occur together develop a temporally constant invariant representation, what I call a “name.” I propose that this name is a set of cells in layers 2 or 3 across a region of cortex in different columns. The set of cells remains active as long as events that are members of the sequence are occurring (e.g., a set of cells that remains active as long as any note in a melody is being heard).

This set of cells representing the name of the sequence is made active via feedback from layer 6 cells in higher regions of cortex. I suggest these name cells are layer 2 cells because of their proximity to layer 1. But it could be any class of cell within layers 2 and 3, which have dendrites in layer 1. For the naming system to work, the apical dendrites of these name cells must form synapses preferentially with axons in layer 1 that originated in layer 6 of higher regions. They should avoid forming synapses with axons in layer 1 that originated in the thalamus. Thus the theory suggests we should find a class of cells, within layers 2 and 3, with apical dendrites in layer 1, that have a strong preference for forming synapses with axons from cells in layer 6 in the region above. Other cells with layer 1 synapses should not have this preference. This is a strong and, as far as I know, novel prediction.

A corollary prediction is we should find another class of cells in layers 2 or 3 whose apical dendrites form synapses preferentially with axons originating in nonspecific regions of the thalamus. These cells predict next items in a sequence.

A set of “name” cells described in prediction 4 should remain active during learned sequences.

A set of cells that stays active during a learned sequence is the definition of a “name” for a predictable sequence. Therefore we should find sets of cells that remain active even as the activity of cells in the rest of a column (cells in layers 4, 5, and 6) is changing. Unfortunately we can’t say what the activity of the name cells will look like. For example, the constant activity of a name pattern could be as simple as a single spike in unison across the set of name cells. Therefore, this group of active cells might be hard to detect.

Another class of cells in layers 2 or 3 (different from the name cells referred to in predictions 4 and 5) should be active in response to an unanticipated input, but should be inactive in response to an anticipated input.

The idea behind this prediction is that unanticipated events must be passed up the cortical hierarchy, but when an event is anticipated we don’t want to pass it up the hierarchy precisely because it was predicted locally. Therefore there should be a class of cells in layers 2 and 3, different from the name class described in predictions 4 and 5, that shows activity when an unanticipated event occurs, but doesn’t show activity if the event was anticipated. The axons of these cells should project to higher regions of cortex. I propose one mechanism to achieve this change in activity. Such a cell could be inhibited via an interneuron activated by a name cell, but at this point there is no way to make a solid prediction of the mechanism. All we can say is that some cells should exhibit this differential activity. This is another strong and, as far as I know, novel prediction.

Related to prediction 6, unanticipated events should propagate up the hierarchy. The more novel the event the higher the unanticipated input should flow. Completely novel events should reach the hippocampus.

Heavily learned patterns are predicted lower in the hierarchy, and, conversely, the more novel an input, the higher it should propagate up the hierarchy. It should be possible to design an experiment to capture this difference. For example, a human could listen to an unfamiliar but simple melody. If the subject hears a note that, although unexpected, is consistent with the style of music, the unexpected note should cause changes of activity in auditory cortex, up to some level in the cortical hierarchy. However, if instead of hearing a note consistent with the style of music the subject heard a complete nonsense sound, such as a crash, we would expect changes of activity from this sound to travel higher up the cortical hierarchy. The results should switch if the subject was expecting to hear the crash but instead heard the note. It might be possible to test this prediction with fMRI on human subjects.

Sudden understanding should result in a precise cascading of predictive activity that flows down the cortical hierarchy.

The “aha” moment when a puzzling sensory pattern is finally understood—such as recognizing the dalmatian dog in figure 12—begins when a region of cortex attempts a new memory match of its input. If the match fits the local region, predictions are passed down the cortical hierarchy in quick succession to all lower regions. If this is a correct interpretation of the stimulus, then each region of the hierarchy will settle on a correct prediction in rapid succession. The same effect should occur while viewing an image with two interpretations, such as a silhouette of a vase that can look like two faces or a Necker cube (an image of a cube that alternately appears in two different orientations). Every time the percept of such an image changes we should see a propagation of new predictions flow down the hierarchy. At the lowest levels, say V1, a column representing a line segment of the image should stay active in either perception of the image (assuming the eyes haven’t moved). However, we might see some cells in that column swap active states. That is, the same low-level feature exists in each image but different cells within a column may be active in the different interpretations. The main point is that we should see a propagation of predictions flow down the cortical hierarchy when a high-level percept changes.

A similar propagation of prediction should occur with each saccade over a learned visual object.

The memory-prediction framework requires that pyramidal neurons can detect precise coincidences of synaptic input on thin dendrites.

For many years it was thought that neurons might be simple integrators, summing the inputs from all their synapses to determine if the neuron should spike. Within neuroscience today, there is much uncertainty about how neurons behave. Some people still hold to the idea of neurons being simple integrators, and many neural network models are built with neurons that work this way. There are also many neuron models that assume a neuron behaves as if each dendritic section operates independently. The memoryprediction model requires that neurons be able to detect the coincidence of only a few active synapses in a narrow window of time. The model could work with even a single potentiated synapse being sufficient to cause a cell to fire, but more likely it would be two or more active synapses in proximity on a thin dendrite. Thus a neuron with thousands of synapses can learn to fire on many different precise and separate input patterns. This is not a new idea, and there is evidence to support it. It is, however, a radical departure from the standard model used for many years. If it were shown that neurons don’t fire on precise and sparse input patterns, it would be difficult to keep the memory-prediction theory intact. Synapses on thick dendrites on or near the cell body need not work this way, only the many synapses on thin dendrites.

Representations move down the hierarchy with training.

I argue that through repeated training, the cortex would relearn sequences in hierarchically lower regions of cortex. This follows naturally from how the memory of sequences of patterns would change the input pattern passed to the next higher regions of cortex. There are a couple of consequences of this process. One is we should find cells that respond to a complex stimulus lower in the cortex after extensive training and higher in the cortex after minimal training. In a human, for example, I would expect to find cells that respond to printed letters in a region such as IT after training to recognize individual letters. But, after learning to read entire words, I would expect to find cells that respond to letters in different parts of V4 in addition to IT. Similar results should be attainable with other species, other regions, and other stimuli. Another consequence of this learning process is that where recall occurs and where errors are detected should move. That is, sensations of highly learned patterns should propagate less distance up the hierarchy. This might be detectable via imaging techniques. We should also be able to detect a change in reaction times to certain stimuli because inputs will not have to travel as far in the cortex to be recognized and recalled.

Invariant representations should be found in all cortical areas.

It is well known that cells exist that respond to highly selective inputs invariant to many details. Cells that respond to faces, hands, Bill Clinton, etc., have been observed. The memory-prediction model predicts that all regions of cortex should form invariant representations. The invariant representations should reflect all sensory modalities below a region of cortex. For example, if I had a Bill Clinton cell in visual cortex it would fire whenever I see Bill Clinton. If I had a Bill Clinton cell in auditory cortex, it would fire whenever I hear the name “Bill Clinton.” I would then expect to find cells in association areas that receive both visual and auditory input that respond to either the sight or the spoken name of Bill Clinton. We should find invariant representations in all sensory modalities and even motor cortex. In motor cortex, cells would represent complex motor sequences. The higher up the motor hierarchy, the more complex and invariant the representation should be. (Recent studies appear to have found cells that activate complex hand-to-mouth movements in monkeys.) These are not novel predictions. Most researchers believe in the general idea that invariant representations are formed in many locations throughout the cortex. However, even though I discussed this as a fact, it has not been demonstrated everywhere. The memory-prediction model predicts we will see such cells in every region of cortex.

The preceding predictions are some of the ways the model in this book can be tested. I am sure there are others. However, it isn’t possible to prove a theory is correct. It is only possible to prove a theory is incorrect. So even if all the predictions I listed above were shown to be true, that wouldn’t be proof that the memory-prediction hypothesis is correct, but it would be strong evidence in support of the theory. The flip side is also true. If some of the predictions above turn out not to be true, it wouldn’t necessarily invalidate the entire thesis. For some of the predictions there are alternate ways the needed behavior could be achieved. For example, there are other ways names of sequences could be created. This appendix is only intended to show that the model leads to several predictions and, therefore, can be tested. Designing experiments is challenging work and would need much more discussion than is appropriate for this book. It would also be great if we could find ways of testing this theory with imaging techniques such as fMRI. There are many imaging laboratories and these experiments can be performed relatively quickly when compared to recording directly from cells.