Page 123 - AN-3-3

Advanced Neurology Evaluating plausibility of thalamic model

REs, coupled with inhibitory facilitation favoring burst “Relay neurons” graph). The Rs responsible for extracting
mode activation, enables the consistent extraction and the PC then enters a refractory state, as recorded in
retransmission of PCs to higher cortical systems. Although “Refractoriness” graph.
not identical, the similarity of reticular activity, as per In Figure 10C, the network dynamics evolve with
our model, stands as a pivotal element supporting the the activation of a second Rs (5, 6 in the “Winning
hypothesis of the thalamus as a central orthogonalizer in neuron” graph). This neuron’s extracted PCs overlap
the brain, showcasing a plausible bioinspired performance. with the previously inhibited REs, creating an “E” shape

3.2. Pattern completion and stabilized images on (Figure 10D). The Rs 1 and 2 also enters a refractory state
the retina and is recorded at positions 1 and 2 in the “Refractoriness”
graph. Additional changes in the REs occur at positions 3,
3.2.1. Pattern completion by the computational model 6, and 3, 7 in the “Relay neuron” graph.
of the thalamus
By the end of the PC extraction process (Figure 10E), a
The process of pattern completion unfolds through the final winning Rs (9, 3) activates, completing the inhibition
inhibitory feedback connections from the second to the process on the REs and fully forming the letter “B”
first layer described in Section 3.1. In this backward path, (Figure 10F). In total, the activation of three Rs (1, 2; 5,
the inhibitions generated are subtracted from the input 6; 9, 3) was required to extract the PCs (“I,” “E,” and “B”)
pattern when presented to the network. These inhibitions from the corrupted “B” input and achieve its completion.
construct a negative replica of the input pattern during This experiment demonstrates the computational pattern
training, mirroring the input pattern itself. Once the completion capability of our artificial thalamic network,
negative feedback pattern matches the positive input which is essential for the subsequent results.
pattern, further reinforcement ceases in the network (for
detailed insights into the biological processes occurring 3.3. Replicating stabilized images on the retina
in the thalamus, please refer reference ). An animated Section 3.2.1 demonstrated that the artificial thalamus
59
graphical representation and the program’s functionality can complete a damaged or noisy pattern. If only a small
can be watched in Video S1. part of the pattern is presented to the network, the PC that
In our model, the REs gradually calculate the mean best matches this part becomes activated, leading to the
of all input patterns and then project the deviations from triggering of a neuron in the second layer and generating
the mean onto the Rs. The highest deviation corresponds backward inhibition in the network. This sequence of
to the first PC, which is represented by the connection inhibitions reproduces a previously presented pattern or a
weights between the REs and the Rs. The inhibitory relevant part of it.
projections from the Rs subtract this deviation from the In the context of stabilized images on the retina,
40
original input, and their projection is adjusted relative presenting partial patterns to the thalamus from the
to previous axes. Subsequently, the remaining deviations retina closely resembles the previous experiment. The
are sequentially processed in descending order by the Rs, constantly changing parts of the image presented to the
reflecting the importance ranking of the PCs (refer to network, caused by the rapid saturation of photoreceptors
Supplementary File, Section 3).
and the activation of only a small fraction of retinal cells
Figure 10 demonstrates the pattern completion corresponding to the pattern, lead to the images reaching
capability of the thalamic network with 81 relay and the thalamus as random pieces of the original pattern. Each
thalamic neurons when corrupted pattern “B” is input. time one of these pieces arrives at the thalamus, a different
One epoch of pixelated (9 × 9) letters, considering the pattern is recovered, not necessarily corresponding to
entire alphanumeric alphabet, was used for training. The the actual pattern shown to the subject. The small pieces
surface in the central graph represents the evolution of of the input pattern arriving at the thalamus generate
the post-synaptic membrane potentials of each neuron in hallucinatory perceived images.
the second reticular layer. We refer to the neurons in their To simulate the scenario of hallucinatory pattern
respective graphs using matrix coordinates (row, column). completion by the thalamic model, we used the same setup
Initially (Figure 10A), the REs in the first thalamic layer as before, comprising 81 relay and Rs, with 13,122 synaptic
receive input from the corrupted “B” (“Pattern”graph). The weights modified at each iteration of the algorithm. A set
Rs (1,2 in the “Winning neuron” graph) extracts the PC of six patterns representing the letters “4,” “H,” “B,” “3,”
from the corrupted pattern and produces an “I”-pattern “b,” and “E” were presented, requiring approximately 100
inhibition over the first layer (Figure 10B; “Inhibitions” iterations to process each pattern. A total of 1000 iterations
graph), initiating a gradual change in the REs (7.5 in the were completed, and with only one or two presentations of

Volume 3 Issue 3 (2024) 11 doi: 10.36922/an.3188