Home
We demonstrate the performance of the three components of the weighted neuronal assignment (namely Regularization by involvement, Normalisation by response strength, and Penalize relevant neurons that did not spike) and their combinations. The figure below shows that component pairs outperform single components and that the combination of all three components leads to the best performing results.
We demonstrate the model performance improvements in post-processing using image sequences via SeqSLAM. The Area Under the Curve (AUC) of the model with 400 places is 39.9% (Figure 6 of paper) using the weighted probability-based assignments. The Figure below demonstrates that the use of SeqSLAM with a sequence length of 5 and 10 frames increased model performance to 88.3% and 98.4% respectively. Using sequenced-based input can be used to encode a larger number of places, which is a desirable future work direction.
To get a better understanding of how the SNN model scales with an increasing number of places, we also explore two more ablation studies using Brian2 on CPU. First, we explore how the computation time scales with increasing the number of output neurons. The figure below displays that the increase in the number of output neurons results in an approximately linear increase (although further analysis can provide a better description) in inference time per query image. In each simulation timestep, the neuronal dynamics of all neurons are computed separately, causing an increase in inference time as the number of output neurons increases. We note that the inference time of the model with 800 output neurons is about 1 second, which is acceptable in many cases.
Next, we investigate how many output neurons are needed to maintain the same high performance for an increasing number of place labels. The figure below demonstrates the SNN model can maintain a similar AUC performance by increasing the number of output neurons for an increasing number of place labels. To achieve an average AUC of
