Fixed transforms no longer need guessing: the Adaptive Interpolating Transform beats fixed-transform baselines with nine added parameters in vision and three per encoder layer in language models
Synopsis
Fixed transforms such as the discrete Fourier transform and discrete cosine transform are widely used in machine learning to extract features, but choosing among them usually requires knowing the dataset's properties, and the approach weakens when that knowledge is missing. The Adaptive Interpolating Transform (AIT) turns the transform itself into learnable parameters, learning data-driven mappings tailored to the dataset and task. On ImageNet-2012 and the General Language Understanding Evaluation tasks, AIT-enhanced models outperform fixed-transform baselines under both from-scratch and fine-tuning workflows, adding only nine parameters to vision models and three per encoder layer to language models.
Interpretation
The transform is no longer picked in advance from dataset properties; it is learned along with the task inside a parameterized framework, and the same idea serves both vision and natural language processing. The prior practice was to choose one suitable transform from many, including the discrete Fourier transform, discrete cosine transform and wavelets, a choice that depends on understanding the dataset and becomes less effective when that knowledge is unavailable. On ImageNet-2012 and the General Language Understanding Evaluation tasks, AIT-enhanced models outperform fixed-transform baselines under both from-scratch and fine-tuning workflows.
In vision models, AIT learns a mixture of transforms for each channel rather than applying one fixed transform to all channels. AIT is a parameterized transform framework that interpolates among several transforms with learnable parameters, yielding a mapping tailored to the dataset and task. This learning is exercised on ImageNet-2012 vision models, at a cost of only nine added parameters.
In language models, AIT learns data-driven combinations of transforms while remaining drop-in compatible. Fixed transforms in language models are likewise limited by a pre-chosen transform form, whereas AIT lets the data decide the combination. On the General Language Understanding Evaluation tasks, each encoder layer adds only three parameters, an overhead described as negligible.
Perspective
The demonstrated scope is vision tasks on ImageNet-2012 and language tasks in the General Language Understanding Evaluation suite, covering from-scratch and fine-tuning workflows; the added parameters are nine for vision models and three per encoder layer for language models.
Whether tasks beyond these two benchmark families and these two workflows benefit similarly is not given; the abstract reports only the comparison of exceeding fixed-transform baselines, without a specific margin of improvement.