Скачать или смотреть Interpretability Tools are Feedback Loops

Speaker: Jinen Setpal, Machine Learning Engineer, DagsHub

"Fundamentally - Machine Learning as a field is designed to emulate the way humans think; hence, neural networks. When we train our models, we use optimizers and loss functions to measure their success. While these functions make sense mathematically, they are far from intuitive or explaining what happened behind the scenes. It’s hard to pick the correct functions, and performing huge grid searches to hyperparameter tune at scale is as logical as bruteforcing an SHA-256 hash.
On the other hand - Interpretability techniques can’t really be used in a training context but are intuitive in helping us understand how a given model interprets a set of data.
This talk aims to bridge the gap between the two, connecting them within a single training loop to maximize training effectiveness without disproportionately increasing compute or training time. Making training intuitive to how humans learn should help develop models that actually work, without resorting to “useless” training.
I aim to showcase - with a practical demonstration - learning techniques to build feedback loops wherein interpretability is used to better optimize a training sequence. I also aim to discuss how this carries forward to complex architectures, and a potential approach for their relevant implementation.
Structurally: the talk would provide an overview on machine interpretability, provide a brief overview on optimizers and loss functions before jumping into the implementation walkthrough of a case study. The case study uses TensorFlow, but can generally be applied to any desired framework."