Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: THEend8_
CSYS5010 Assignment 1
Complexity in a brain-inspired
agent-based model. Neural Networks, 33, 275-290.
In this paper, Joyce et al. aimed to develop an agent based model capable
of representing a real human brain at the scale of 90 anatomical regions of
interest. These regions are represented by network nodes, and connections
between them are built by analysing the correlation in brain activity seen in
functional magnetic resonance imaging (fMRI). The model was tested for its
capability to producing a range of ‘brain like’ states and solve simple
computational tests.
The model was developed under the assumption that connectivity is vital in
the functional brain structure. Each region of the brain can independently be
active or inactive but affects those they are connected to, leading to a
range of global patterns including steady states and oscillations with both
short and long periods.
The rules of the model are heavily influenced by the work on cellular
automata (Wolfram, 2002). The update rule uses three numbers for each
node to determine its state in the next time step. The first number is the
current state of the node, set at 1 or 0 representing active or inactive. The
second and third numbers are determined by the positively and negatively
connected nodes. They are also represented by a 1 or 0 depending on if
they pass some threshold value. A genetic algorithm is used to find the
optimal threshold parameters and determine the best update rule, which
could take one of 256 possibilities.
Joyce et al. showed that their functional brain model was able to pass a
common computational test called the density classification task, while the
test was failed by a null network where the connectivity was randomised.
They also showed that the model was capable of producing a range of
global patterns depending on the specific update rule which is used. Some
of these patterns had periods of several thousand steps, quite unlike those
produced by the null networks.
This model provides a useful conceptual model for confirming that complex
patterns resembling the output of the brain can be generated using simple
local update rules. It also showed that the specific network topology of the
brain is important to its computation. In subsequent work, Joyce et al. used
their model as a reference for the real brain, and went on to perform analysis
on its network properties, such as testing resilience and connectivity (Joyce et
al., 2013).
However, the model appears to be caught between two related
approaches, one being artificial neural networks, where the structure and
connectivity are decided based on the task to be performed, and the
second being a more physical approach to modelling the brain using
electrical signals. Since the update rules are simply a deterministic response
to the activity of each node’s neighbours, this model provides little difference
from a model composed of a network of integrate and fire neurons (Burkitt,
2006).
The use of genetic algorithms to find a suitable update rule out of all possible
combinations of values seems more valid in the case of Wolfram’s abstract
model of cellular automata than for a realistic representation of the brain. It is
also unclear what mechanism would allow the brain to switch between
alternative update rules in order to perform different types of computation or
generate different outputs.
To extend the model, would be interesting to build additional agent based to
study self-organising and emergent phenomena, such as transitions between
different brain states and dynamic patterns of connectivity. Neurons can be
allowed to build and prune links based on local criteria (Schoenharl and
Madey, 2003). The collection of fMRI data for non-resting states may be able
to verify such simulation results.