# Seven Myths in Machine Learning Research

## TL;DR

This paper debunks seven common misconceptions in machine learning research, clarifying misunderstandings about tools, data, validation practices, training methods, model architectures, interpretability, and the robustness of popular techniques.

## Contribution

It systematically challenges seven prevalent myths in machine learning, providing clarifications and insights to improve research practices and understanding.

## Key findings

- TensorFlow is often misunderstood as a tensor manipulation library.
- Image datasets are not fully representative of real-world images.
- Test sets are sometimes misused for validation.

## Abstract

We present seven myths commonly believed to be true in machine learning research, circa Feb 2019. This is an archival copy of the blog post at https://crazyoscarchang.github.io/2019/02/16/seven-myths-in-machine-learning-research/ Myth 1: TensorFlow is a Tensor manipulation library Myth 2: Image datasets are representative of real images found in the wild Myth 3: Machine Learning researchers do not use the test set for validation Myth 4: Every datapoint is used in training a neural network Myth 5: We need (batch) normalization to train very deep residual networks Myth 6: Attention $>$ Convolution Myth 7: Saliency maps are robust ways to interpret neural networks