Page 126 - AN-3-3

Advanced Neurology Evaluating plausibility of thalamic model

stimulus “HB” consistently generates the same completed inputs through PC extraction to facilitate learning
“b.” This demonstrates an important phenomenon of through orthogonal patterns. If our hypothesis holds, it
generalization, where the network produces the same appears that the thalamic function contributes to cortical
response despite different input pattern configurations. encoding, particularly in high-order nuclei. The presence
This highlights the robustness and consistency of the of an orthogonalize during cortical processing suggests
learning process, even when the input is corrupted. that the neural code may partly rely on encoding by
In Figure 11D, the shape of a “B” emerges when a new PCs, highlighting the thalamus as a critical element in
set of “retinal cells” (red dots) are active while presenting deciphering overall neural function. This view is consistent
pattern “HB” and the remaining cells (in green) are with recent findings in the literature on the role of the
saturated. A “3”-shaped inhibition result is obtained when thalamocortical circuit in the overall functioning of the
34
a new set of “retinal cells” is not saturated (Figure 11E). In brain.
Figure 11F, the last Rs triggers, and therefore, the last PC We deepen our understanding of thalamic computations
overlaps with previous inhibitions, creating the final result: by integrating a probabilistic interpretation with vector
an inhibition in the shape of the number 4. This result mathematics. We demonstrate how thalamic neurons
corresponds to one of the images perceived by a subject could be modeling cognitive phenomena such as pattern
when “HB” is shown (Figure 4A). completion through a geometric approach to neural
In Figure 11G, the pattern “4b,” which was not computation. Neurons in the artificial thalamus were
present in the training set of patterns, is generated. This capable of performing complex probabilistic computations,
demonstrates that a hallucinatory image can be generated not merely transmitting sensory information but actively
even when the thalamus is deprived of most of its retinal participating in the interpretation and response decisions.
afferent neurons. Finally, the final inhibition takes the form This aligns with previous models suggesting that thalamic
of an “E” in Figure 11H. Notice that there is no winning relays enhance the signal-to-noise ratio in sensory
neuron at this time because all winning neurons are in the pathways and contribute to attentional mechanisms by
refractory state, as in Figure 11B, C, and G. selectively filtering sensory input based on probabilistic
assessments. 60
This experiment provides insights into the phenomenon
of stabilized images in the retina. Retina neurons respond The results demonstrated that our thalamic
maximally to luminosity variations but saturate when computational model was able to phenomenologically
the illumination is constant, rendering most of the visual replicate the physiological functioning of the Rs with
field inactive or saturated during steady images. Only significant similarity. Our focus on the activity of this
neurons at the image-background borderline, exhibiting specific neuron arises from its essential role in our
unstable behavior, remain active. The thalamic network thalamus hypothesis. The competitive activity among Rs,
projects available PCs onto this small fraction of the input when stimulated by REs, facilitates the reduction of the
pattern, selecting the PC with the largest projection. The dimensionality of sensory or high-order input by extracting
thalamus reconstructs this component from the tiniest the most essential information through PC analysis.
part of the input pattern, generating a sequence of The incorporation of a “universal basis” for vector
coherent fragments perceived sequentially, varying over transformations in neural computations suggests a
time. A computational model of the thalamus replicates mechanism for rapid learning and memory encoding.
this process, yielding results consistent with those of the Neurons could adjust synaptic strengths based on
subjects in the experiment. the statistical properties of the input, favoring certain
interpretations over others based on past experiences,
4. Discussion essentially learning by adjusting their basis vectors. As an

The Hebbian rule, widely recognized as a fundamental adaptive system, the network learns patterns and enables
mechanism for learning and neuronal plasticity, requires their recovery even when the input is incomplete or
the orthogonalization of neural inputs – a key challenge corrupted. This pattern completion capability resembles
for its effective implementation in ANNs. Assuming that the operation of an auto-associative neural network,
this requirement also applies to the brain networks, the which served as inspiration for our model. Leveraging
theory implies the existence of an instance responsible a familiar system within the domain of artificial neural
for this function. Our hypothesis posits that the thalamus networks allows us to utilize established conceptual
acts as this central orthogonalizer within the brain’s and mathematical frameworks, providing a degree of
complex systems, where it not only regulates sensory flow predictability and understanding of the simulated system’s
to the cortex but also plays a crucial role in reformatting operational limitations.

Volume 3 Issue 3 (2024) 14 doi: 10.36922/an.3188