Target Aware Network Architecture Search and Compression for Efficient Knowledge Transfer

TL;DR

This work tackles the challenge of transfer learning with over-parameterized pre-trained CNNs by introducing TASCNet, a two-stage framework that first learns a target-specific configuration for the deeper layers via Bayesian Optimization over a defined hyperparameter space, and then prunes redundant filters based on cosine similarity with a regularized optimization objective. The method reduces both trainable parameters and FLOPs while preserving or improving accuracy across multiple CNN backbones (VGG-16, ResNet-50, DenseNet-121) on image datasets (CalTech-101, CalTech-256, Stanford Dogs) and also validates the approach on NLP tasks (Movie Reviews) using cross-domain transfer. Key contributions include an automatic target-aware tuning mechanism using a Gaussian Process surrogate with Expected Improvement, a cosine-similarity–based pruning pipeline with a novel regularizer, and demonstrated efficiency gains without substantial performance loss. The results indicate that resource-efficient knowledge transfer is feasible for both computer vision and NLP, enabling faster inference and lower memory usage in practical deployments.

Abstract

Transfer Learning enables Convolutional Neural Networks (CNN) to acquire knowledge from a source domain and transfer it to a target domain, where collecting large-scale annotated examples is time-consuming and expensive. Conventionally, while transferring the knowledge learned from one task to another task, the deeper layers of a pre-trained CNN are finetuned over the target dataset. However, these layers are originally designed for the source task which may be over-parameterized for the target task. Thus, finetuning these layers over the target dataset may affect the generalization ability of the CNN due to high network complexity. To tackle this problem, we propose a two-stage framework called TASCNet which enables efficient knowledge transfer. In the first stage, the configuration of the deeper layers is learned automatically and finetuned over the target dataset. Later, in the second stage, the redundant filters are pruned from the fine-tuned CNN to decrease the network's complexity for the target task while preserving the performance. This two-stage mechanism finds a compact version of the pre-trained CNN with optimal structure (number of filters in a convolutional layer, number of neurons in a dense layer, and so on) from the hypothesis space. The efficacy of the proposed method is evaluated using VGG-16, ResNet-50, and DenseNet-121 on CalTech-101, CalTech-256, and Stanford Dogs datasets. Similar to computer vision tasks, we have also conducted experiments on Movie Review Sentiment Analysis task. The proposed TASCNet reduces the computational complexity of pre-trained CNNs over the target task by reducing both trainable parameters and FLOPs which enables resource-efficient knowledge transfer. The source code is available at: https://github.com/Debapriya-Tula/TASCNet.

Figures (2)

• Figure 1: For each layer, cosine similarity between filters belonging to the same filter pair is computed after each training cycle. In the diagram, a cuboid, $i$ represents the collection of weights of the $i^{th}$ filter over all the training epochs.
• Figure 2: In each pruning iteration, the top 5% of the filter pairs with high similarity are chosen for optimization. During optimization, we improve the similarity between the filters by re-training the model using the custom regularizer. After optimization, the filter having a lower $\ell_1$-norm compared to the other filter in the same pair will be pruned.