Deep Forest: Towards An Alternative to Deep Neural Networks
Authors: Zhi-Hua Zhou, Ji Feng
IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiments, gc Forest achieves highly competitive performance to deep neural networks, whereas the training time cost of gc Forest is smaller than that of deep neural networks. In this section we compare gc Forest with deep neural networks and several other popular learning algorithms. |
| Researcher Affiliation | Academia | Zhi-Hua Zhou and Ji Feng National Key Lab for Novel Software Technology, Nanjing University, Nanjing 210023, China {zhouzh, fengj}@lamda.nju.edu.cn |
| Pseudocode | No | The paper does not include any sections or figures explicitly labeled as 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | The code of gc Forest is available 2. 2http://lamda.nju.edu.cn/code/gcForest.ashx |
| Open Datasets | Yes | The MNIST dataset [Le Cun et al., 1998] contains 60,000 images of size 28 by 28 for training (and validating), and 10,000 images for testing. The ORL dataset [Samaria and Harter, 1994] contains 400 gray-scale facial images taken from 40 persons. The GTZAN dataset [Tzanetakis and Cook, 2002] contains 10 genres of music clips. The s EMG dataset [Sapsanis et al., 2013] consists of 1,800 records. The IMDB dataset [Maas et al., 2011] contains 25,000 movie reviews. UCI-datasets [Lichman, 2013] with relatively small number of features: LETTER, ADULT, and YEAST. |
| Dataset Splits | Yes | In detail, we split the training set into two parts, i.e., growing set and estimating set1; then we use the growing set to grow the cascade, and the estimating set to estimate the performance. If growing a new level does not improve the performance, the growth of the cascade terminates and the estimated number of levels is obtained. For all experiments we take 80% of the training data for growing set and 20% for estimating set. |
| Hardware Specification | Yes | Our experiments use a PC with 2 Intel E5 2695 v4 CPUs (18 cores), and the running efficiency of gc Forest is good. |
| Software Dependencies | No | The paper mentions software components like 'Re LU for activation function, cross-entropy for loss function, adadelta for optimization', but does not provide specific version numbers for these or any underlying libraries/frameworks used. |
| Experiment Setup | Yes | Thus, in all experiments gc Forest is using the same cascade structure: each level consists of 4 completely-random tree forests and 4 random forests, each containing 500 trees, as described in Section 2.1. Three-fold CV is used for class vector generation. For multi-grained scanning, three window sizes are used. For d raw features, we use feature windows with sizes of d/16 , d/8 , d/4. Table 1 summarizes the hyper-parameters of deep neural networks and gc Forest, where the default values used in our experiments are given. |