Estimation is EXTREMELY slow using one large ELM because of the (implicit) matrix inversion. We can drastically speed this up by implementing an ensemble of ELMs. Since solving the least squares problem scales cubically, it will be much faster to decrease the size of the number or observations for each ELM. This will also mean we can use less neurons for each ELM, further increasing speed. Another benefit is that we don't have to search for an L2 penalty because bagging is already giving us regularization. This will also make getting confidence intervals easy because we can just use the variance in the individual estimates from the ensemble.
dscolby
commented
1 month ago
Estimation is EXTREMELY slow using one large ELM because of the (implicit) matrix inversion. We can drastically speed this up by implementing an ensemble of ELMs. Since solving the least squares problem scales cubically, it will be much faster to decrease the size of the number or observations for each ELM. This will also mean we can use less neurons for each ELM, further increasing speed. Another benefit is that we don't have to search for an L2 penalty because bagging is already giving us regularization. This will also make getting confidence intervals easy because we can just use the variance in the individual estimates from the ensemble.