Tilt and Average : Geometric Adjustment of the Last Layer for Recalibration
Authors: Gyusang Cho, Chan-Hyun Youn
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through this, we demonstrate that our approach, in addition to the existing calibration-map-based techniques, can yield improved calibration performance. Code available : URL. ... The classifier weights newly derived through TNA do induce a very slight change in accuracy, albeit very marginal (less than 0.3%). In terms of calibration performance, the application of weights by TNA surpasses the performance of the original weight. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering, KAIST, Daejeon, Republic of Korea. |
| Pseudocode | Yes | Algorithm 1 TILT AND AVERAGE (TNA) |
| Open Source Code | Yes | Code available : URL. ... The code will be available, URL. |
| Open Datasets | Yes | CIFAR10, CIFAR100 (Krizhevsky et al., 2010): Tiny images from web. ... Image Net-1k (Deng et al., 2009) : Image of natural objects... SVHN (Netzer et al., 2011): Street View House Numbers... |
| Dataset Splits | Yes | CIFAR10, CIFAR100 (Krizhevsky et al., 2010): ... 50,000/5,000/5,000 ... Image Net-1k (Deng et al., 2009) : ... 1.2M/12,500/37,500 ... SVHN (Netzer et al., 2011): ... 73,257/5,000/19,032. |
| Hardware Specification | Yes | Computationally, this process takes less than 30 minutes on a single GPU of RTX3080. |
| Software Dependencies | No | For the Image Net trained networks, the models are borrowed from torchvision (Marcel & Rodriguez, 2010) and timm(Wightman, 2019). The paper mentions these libraries but does not provide specific version numbers. |
| Experiment Setup | Yes | Hyperparameters. As proposed in the algorithm, we search over θ [0, 90]. Over datasets and models, ne is set to 10, α = 5, β = 1. The search interval of rotation number nt is set to 50. Unless noted, the maximum rotated angle θs is set to 0.9rad. The displayed value reflects an average of 5 repeated experiments, with the standard deviation depicted as a subscripted value. The number of bins B is set to 15. ... Specifically, we used Momentum-SGD with batch size of 128, momentum 0.9. The batch size is fixed to 256. The learning rate scheduler follows the work of (Garipov et al., 2018) and is specified as... |