# Deep Learning on Small Datasets without Pre-Training using Cosine Loss

## TL;DR

This paper demonstrates that using cosine loss instead of cross-entropy enables effective training of CNN classifiers from scratch on small datasets, achieving significantly higher accuracy and allowing easy integration of prior knowledge.

## Contribution

The paper introduces cosine loss as a superior alternative to cross-entropy for small datasets, enabling training from scratch and incorporation of class hierarchies.

## Key findings

- Cosine loss outperforms cross-entropy on small datasets.
- Training from scratch with cosine loss yields 30% higher accuracy on CUB-200-2011.
- Incorporating class hierarchies with cosine loss improves performance.

## Abstract

Two things seem to be indisputable in the contemporary deep learning discourse: 1. The categorical cross-entropy loss after softmax activation is the method of choice for classification. 2. Training a CNN classifier from scratch on small datasets does not work well. In contrast to this, we show that the cosine loss function provides significantly better performance than cross-entropy on datasets with only a handful of samples per class. For example, the accuracy achieved on the CUB-200-2011 dataset without pre-training is by 30% higher than with the cross-entropy loss. Further experiments on other popular datasets confirm our findings. Moreover, we demonstrate that integrating prior knowledge in the form of class hierarchies is straightforward with the cosine loss and improves classification performance further.