Some Improvements on Deep Convolutional Neural Network Based Image Classification
Authors: Andrew G. Howard
ICLR 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our system achieved a top 5 classification error rate of 13.55% using no external data which is over a 20% relative improvement on the previous year s winner. |
| Researcher Affiliation | Industry | Andrew G. Howard Andrew Howard Consulting Ventura, CA 93003 andrewgeraldhoward@gmail.com |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper states 'Our models are based on the 2012 winning system [6] and use the code provided at http://code.google.com/p/cuda-convnet as a starting point.' This indicates they used existing open-source code, not that they made their own modifications or implementations open source. |
| Open Datasets | Yes | The Imagenet Large Scale Visual Recognition Challenge (ILSVRC) [8] is a venue for evaluating what the current state of the art for image classification and recognition is. It is large dataset of 1.2 million images with 1000 classes that are a subset of the Imagenet dataset [3]. |
| Dataset Splits | No | The paper mentions training until 'the validation error plateaus' and tables include 'Val Top-1' and 'Val Top-5' results, indicating the use of a validation set. However, it does not provide specific split percentages or sample counts for the training, validation, and test sets. |
| Hardware Specification | No | The paper mentions 'GPU computing' as an enabler for larger models but does not provide any specific details about the GPU models, CPU, or other hardware used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'the code provided at http://code.google.com/p/cuda-convnet as a starting point' and 'the python image library (PIL)', but it does not specify version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | Our model structure is identical with the exception that the fully connected layers are twice as big. We use the same training methodology of training the net until the validation error plateaus and reducing the step size by 10 at each plateau. We start with a step size of 0.001 and divide the step size by 10 once the validation score plateaus. We reduce the step size twice before convergence. In practice we use drop out for the initial training and the first step size reduction to 0.0001, we then turn off drop out to finish the training at 0.0001. Then the model is finished training with drop out turned off and the step size reduced to 0.00001. |