GRAPHIC: KELLIE HOLOSKI/
SCIENCE
BASED ON THOMAS SPLETTSTOESSER
SCIENCE science.orgmay have complementary roles and compet-
ing influences on the associative properties
of pyramidal neurons, with some control-
ling dendritic calcium spikes through direct
dendritic inhibition and others controlling
the firing of inhibitory neurons to release
dendrites from inhibition (disinhibition).
Moreover, inhibition in layer 1 has itself been
found to be plastic and undergo experience-
dependent changes ( 3 , 8 ). Such modulation
may form the basis of selection and activa-
tion of engram pyramidal cortical neurons—
the neurons whose changes in firing encode
new memories ( 12 ).
Conversely, stabilization could also involve
the regulation of postsynaptic dendritic excit-
ability ( 4 ). Modulation of the intrinsic excit-ability of the tuft dendrites could, in princi-
ple, have a large effect on the output of the
neuron because they generate powerful local
dendritic N-methyl-D-aspartate (NMDA) and
calcium spikes. Other open questions include
the extent to which local dendrite-specific
protein synthesis in the apical dendritic
compartment is required for memory stabi-
lization and how the influence of memory-
related projections to layer 1 is mediated,
that is, via specific neurotransmitters or neu-
romodulators that may be spatially targeted.
The hypothesis that associative, context-
related memories are stored in layer 1 raises
interesting questions about the psychology of
learning. Features are thought to be learned
during early development in a critical win-
dow within which the dominant form oflearning involves the extraction of regulari-
ties ( 13 ). For example, children typically ab-
sorb new motor skills and low-level features,
such as the accent of a second language,
with great ease. According to our hypothesis,
learning features would be independent of
layer 1 and the tuft dendrites of pyramidal
neurons. Consistently, the complex dendritic
properties necessary for binding context and
features are unavailable before adolescence
( 14 ). Feature learning would therefore mod-
ify synapses distributed throughout the other
layers of the cortical column.
Thus, our proposal is that there are two
phases of learning. The first phase involves
the grounding of the cortical network in the
statistical features most prevalent in the ex-ternal world, which takes a long time (many
repetitions) to establish but is then relatively
stable over time. The second phase involves
a more dynamic learning of context and
high-level concepts that establish an inter-
nal model using the putative predictive role
of the pyramidal neuron tuft dendrites ( 11 ).
For example, adults can more quickly and ex-
plicitly learn the higher-level structure of a
second language (e.g., grammar), while never
completely acquiring the detailed features of
the accent.
The shape of pyramidal neurons could
facilitate this two-phase model of learn-
ing by allowing plasticity mechanisms to be
segregated and separately modulated in the
different dendritic trees at the top and the
bottom of the pyramidal neuron. The physi-cal segregation of learning within the neu-
ron would also allow independent access to
memory of features and associative memory.
Furthermore, top-down information could
be deployed flexibly without compromising
feed-forward information that should remain
stably grounded in the outside world. The
link between top-down control and semantic
memory might explain how it can be explic-
itly accessed. Accordingly, whether grammat-
ical knowledge is accessed unconsciously ver-
sus explicitly would depend on whether the
knowledge is encoded as features or context.
Information segregation using multicom-
partment neurons may also inform machine
learning models ( 15 ). Despite their successes,
modern artificial networks still perform very
poorly on tasks that require inference from
very little information, which is a distinct fea-
ture of adult learning in intelligent animals.
Identifying the locus of memory in the neo-
cortex allows more biologically inspired mod-
els to be generated. For example, it was found
that after learning, pyramidal neurons emit-
ted bursts of action potentials that were more
salient to memory retrieval ( 7 ). It is therefore
possible that the mammalian brain uses a
ternary rather than binary system of outputs
(i.e., “0,” no output; “1,” low-frequency out-
put; or “2,” burst output). This could signal
the type of information (statistical versus
associative) to downstream neurons and
facilitate credit assignment for learning in
a feedback system ( 15 ). Such observations
can offer inspiration for the design of more
robust and intuitive machine learning prin-
ciples. Therefore, the identification of layer 1
as the locus of semantic memory promises to
accelerate our understanding of learning and
memory in the human brain and provide in-
sights for the development of treatments for
memory disorders as well as the design of ar-
chitectures for artificial intelligence. jREFERENCES AND NOTES- Y. Dudai, A. Karni, J. Born, Neuron 88 , 20 (2015).
- L. E. Williams, A. Holtmaat, Neuron 101 , 91 (2019).
- E. Abs et al., Neuron 100 , 684 (2018).
- J. Cichon, W.-B. Gan, Nature 520 , 180 (2015).
- M. B. Pardi et al., Science 370 , 844 (2020).
- Y. Yang et al., Nat. Neurosci. 19 , 1348 (2016).
- G. Doron et al., Science 370 , eaaz3136 (2020).
- B. Schuman et al., Annu. Rev. Neurosci. 44 , 221 (2021).
- H. Makino, T. Komiyama, Nat. Neurosci. 18 , 1116 (2015).
- S. Manita et al., Neuron 86 , 1304 (2015).
- M. Larkum, Trends Neurosci. 36 , 141 (2013).
- T. Kitamura et al., Science 356 , 73 (2017).
- C. Blakemore, G. F. Cooper, Nature 228 , 477 (1970).
- J. J. Zhu, J. Physiol. 526 , 571 (2000).
- A. Payeur et al., Nat. Neurosci. 24 , 1010 (2021).
ACKNOWLEDGMENTS
Supported by the Deutsche Forschungsgemeinschaft
(EXC 257 NeuroCure and SFB 1315 – 327654276) and EU
Horizon 2020 (Human Brain Project and European Research
Council). We thank B. Rudy, A. Holtmaat, J. Aru, S. Elz, and
R. Naud as well as the reviewers for insightful comments and
SciStyle for artwork advice. G.D. is employed by Bayer AG.10.1126/science.abk1859Neocortical layersNeocortexShapeMotionColorFeatureContext?First
orderHigher
order
Thalamus
Basal ganglia Memory structuresL1L2/3L4L5AmygdalaHippocampusMemory switchNeocortical layer 1—the memory layer?
Neocortical layer 1 (L1) receives a convergence of information from
different memory-related structures (e.g., the hippocampus, amygdala,
and basal ganglia). It is hypothesized that these act as a switch that
gates plasticity in L1. Alternatively, these inputs might directly provide
contextual memory information (dashed green lines) to integrate
feature-specific information with context-related information.29 OCTOBER 2021 • VOL 374 ISSUE 6567 539