During hyperparameter tuning, I encountered excessive multiple local minima, a symptom of a highly sensitive model performance to small changes in the values of hyperparameters. And it revealed the epistemological limitations of the grid search and an illusion of knowledge created by the interpolated 3D visualization.
The grid search can only capture the model performances over the search-grid, leaving us no information whatsoever regarding the performances off the search-grid.
To make the matter worse, 3D interpolated visualization, by smoothing out the performance landscape, blind-holds our vision and creates an illusion of knowledge regarding the actual model performances off-the-grid.
In a nutshell, both of these defects are embedded on the design of these methods. Therefore, these defects are architecturally inherent and cannot be resolved. That sets an epistemological limitation: we will never attain our perfect knowledge of the ground-truth performance landscape, thus, the ground-truth global minimum.
Ironically, bias-variance trade-off opened up an enigmatic fundamental question: whether the ground-truth global minimum does really matter at all.
In addition, it also left us another question: whether excessive multiple local minima indicates the model instability in the deployment domain.
Climate risks are difficult to model, thus, vulnerable to model risks. Unfortunately, without a solid standard on climate risk models, climate models are subject to abuses for discretionary manipulations (political or/and business). It could increase the risk of GREENWASH.
Any model has its inherent limitations embedded on its own architecture. Therefore, it is imperative for us to understand those inherent limitations of a model in use and the limited implications of its outputs.
I hope that this post serves a basis for checklist for those involved in the production and the use of a model to assess the climate-related physical risks.
Series: Basic Intuitions for Machine Learning & Deep Learning
Throughout this series, I would like to outline basic intuitions and inspirations that I gained from my past Machine Learning journey. At the beginning it was very difficult for me to grasp the high-level conceptual framework of Machine Learning and Deep Learning. Whatever the real reason it might have been, to get right intuitions and inspirations was very helpful for me to grasp the subject.
Now, reflecting my own experience, my intention for this series is to share those intuitions with beginners. In this context, as a general policy throughout this series, I would like to draw a rough sketch rather than to go into details. Fortunately, there are oceans of open-resources that you can dive deeper into for these topics. Instead of creating redundant works, my intention in this series is to present guiding intuitions that help the readers navigate their own ML & DL journeys.
The content of this series is divided into several chapters as below:
In the ocean of blockchain topics in the web space, we frequently encounter blockchain-idealists’ mantras. Some cast a radical notion that a blockchain-base distributed system can eliminate rent-seeking middlemen and, as a result, deter, or even eliminate an abuse of system by monopolist and oligarchs.
However, the reality of blockchain as of today is far from, is even contrary to, such radical notions. The gap between the popular blockchain-idealist’s mantras and reality has caused confusions about blockchain.
Despite the gap, their beautiful mantras continue to echo and have strong influence in shaping our collective perception about and our collective behaviour toward blockchain.
Are they hypnotising us with those deceptive mantras? Or, are we in the middle of a long journey to realise such a radical revolution in the way we organize transactions?
The series shares a part of my journey in understanding causes for the gap. Although it does not aim to cover a comprehensive list of issues, it intends to promote a better understanding about the confusion.