1. What is the cortical learning algorithm (CLA) based on?
The cortical learning algorithm (CLA) is based on the hierarchical temporal memory (HTM) model, which is a representation of the human neocortex. HTM is a computational framework that aims to mimic the structure and function of the neocortex, the outer layer of the brain responsible for higher cognitive functions. The CLA utilizes the principles of HTM to forecast time-series data by leveraging columns, cells, and synapses as memory and linking components. Each column and cell can transition between multiple states, representing input values at different time steps and in the time-series data, respectively. The CLA achieves forecast accuracy by adjusting the states of columns and cells based on synapse relationships. However, the conventional CLA struggles with changing trends in time-series data, leading to deteriorated forecast accuracy. To address this limitation, the burst-based multilayered CLA (BM-CLA) was proposed, incorporating a lower and upper layer to improve forecast accuracy in the presence of changing trends.
read more
2. What is the significance of synapse permanence value?
The synapse permanence value (p) determines whether a synapse is connected or disconnected based on a given connection threshold (th). If p is greater than or equal to th, the synapse is connected, represented by solid lines in Figure 1. Conversely, if p is less than th, the synapse is disconnected, shown as dashed lines. This mechanism plays a crucial role in neural network simulations, allowing researchers to study the impact of synaptic connections on neural activity and behavior. By adjusting the permanence value and connection threshold, researchers can investigate the effects of synaptic plasticity and network dynamics on various cognitive processes and neurological disorders.
read more
3. What is the prediction process of Burst-Based Multilayered CLA (BM-CLA) and how does it handle trend changes in time-series data?
In the prediction process of BM-CLA, the lower layer receives an input bit string x and makes A c columns the active state, with cells in the active columns also being active. The lower layer counts the number of burst columns, n b, and informs the upper layer when n b >= the user-defined threshold th b. The upper layer maintains the column states and makes A c columns the active state, with cells in their active columns also being active. BM-CLA makes cells with more than th Y cell synapses connected to active cells the predictive state. When n b >= th b, BM-CLA conducts a state change in the upper layer and easily makes the other cells in the lower layer into the predictive state. This prevents the deterioration of prediction accuracy when previously received time-series data are again received. BM-CLA handles trend changes in time-series data by handling the number of burst columns, n b, greater than or equal to th b as a trend change, conducting a state change in the upper layer, and making appropriate cells in the lower layer into the predictive state.
read more
4. What trends are present in X 1 (t), X 2 (t), and X 3 (t)?
X 1 (t) has two trends: sine and sawtooth waves. X 2 (t) has three trends: sine, sawtooth, and triangular waves. X 3 (t) has three trends: sine, sawtooth, and logistic map with a = 3.875 waves. Each trend takes up 100-time steps in one cycle. The trends in X 1 (t), X 2 (t), and X 3 (t) are cyclic and change randomly every 100 time steps. The total number of time steps is 10,500. These trends are used in the experimental settings to generate artificial time-series data for comparison and analysis.
read more