Differences
This shows you the differences between two versions of the page.
Both sides previous revision Previous revision Next revision | Previous revision | ||
p:subsmp1 [2020/04/01 07:01] – [Principal component analysis] cjj | p:subsmp1 [2020/04/01 07:53] (current) – [Principal component analysis] cjj | ||
---|---|---|---|
Line 2: | Line 2: | ||
Random subsets of the neurons are chosen for analysis. The results are compared to original system and among different sub-sample sizes. | Random subsets of the neurons are chosen for analysis. The results are compared to original system and among different sub-sample sizes. | ||
=====Principal component analysis===== | =====Principal component analysis===== | ||
- | | {{: | + | | {{ : |
- | The errorbars are calculated from the standard deviation of 8 random subsets for each given size. | + | The errorbars are calculated from the standard deviation of 8 random subsets for each given size. It's seems power law scaling $\lambda_k \propto k^{-\alpha}$ with $\alpha \approx 0.5$ is more accurate for a sub-sampled system of smaller |
=====Thermodynamics of subsamples===== | =====Thermodynamics of subsamples===== | ||
Boltzmann learning is performed for each random sub sample for the model parameters, $h_i$ and $J_{i,j}$. The specific heat curves of the resulted models are calculated and as shown below. | Boltzmann learning is performed for each random sub sample for the model parameters, $h_i$ and $J_{i,j}$. The specific heat curves of the resulted models are calculated and as shown below. | ||
{{ : | {{ : | ||
+ | The criticality condition ($T\approx 1.0$) appears pretty robust to sub sampling where only a portion of the neurons from a large network is observed. |