Towards Disentangling Information Paths with Coded ResNeXt
Authors: Apostolos Avranas, Marios Kountouris
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present experimental results to assess the performance of the proposed Coded Res Ne Xt. First, we show that our algorithm achieves sub NN specialization. To demonstrate this we show that when the sub NNs specialized on the class of interest are removed, the performance degrades, whereas it remains the same or even improves when the sub NNs removed are not specialized for that class. |
| Researcher Affiliation | Collaboration | Apostolos Avranas EURECOM Sophia Antipolis, France avranas@eurecom.fr Marios Kountouris EURECOM Sophia Antipolis, France kountour@eurecom.fr Currently working at Amadeus, Nice |
| Pseudocode | No | The paper references an 'heuristic algorithm' in Appendix A but does not provide a structured pseudocode or algorithm block within the main document. |
| Open Source Code | Yes | The datasets are publicly available and the code is included in the supplementary material. |
| Open Datasets | Yes | For CIFAR-10 (C10), CIFAR-100 (C100) [32], and Image Net-1k (IN) [61] classification datasets. |
| Dataset Splits | Yes | In Fig. 3 we pick the first class of CIFAR-10/100 and Image Net ( airplane , apple , and tench , respectively) and remove all inactive sub NNs for that class. ...Figure 3 depicts with blue the output distribution when inputting samples of the validation set belonging to the first class of the dataset (i.e., in-distribution positives), and with red when the samples belong to some other class (i.e., in-distribution negatives). |
| Hardware Specification | Yes | We reduce the resolution from 224 to 160, since on the TPU-v2 of Google Colab (the platform used for our experiments) the training would take more than three weeks. |
| Software Dependencies | No | The paper mentions following 'the training process proposed by timm library [1]' but does not specify version numbers for the timm library or any other software dependencies. |
| Experiment Setup | Yes | The epochs are 250 (first 5 as warmup [20] and last 10 cooling down), the batch size is 1536, and the learning rate 0.6. Rand Augment [10] of 2 layers and magnitude 7 (varied with a standard deviation of 0.5) is used and also random erasing augmentation [91] with probability 0.4 and 3 recounts. |