[Paper Review] Auto-JacoBin: Auto-encoder Jacobian Binary Hashing

Xiping Fu, Brendan McCane+3 more|arXiv (Cornell University)|Feb 25, 2016

Advanced Image and Video Retrieval Techniques28 references3 citations

TL;DR

This paper proposes Auto-JacoBin, a novel auto-encoder-based hashing method that enhances geometric consistency in Hamming space by leveraging the Jacobian of a noise-removing function to preserve local data manifold structure. The method achieves state-of-the-art performance on large-scale datasets, with 92% Recall@1000 on NUS-WIDE using only 64 bits.

ABSTRACT

Binary codes can be used to speed up nearest neighbor search tasks in large scale data sets as they are efficient for both storage and retrieval. In this paper, we propose a robust auto-encoder model that preserves the geometric relationships of high-dimensional data sets in Hamming space. This is done by considering a noise-removing function in a region surrounding the manifold where the training data points lie. This function is defined with the property that it projects the data points near the manifold into the manifold wisely, and we approximate this function by its first order approximation. Experimental results show that the proposed method achieves better than state-of-the-art results on three large scale high dimensional data sets.

Motivation & Objective

• To address the challenge of preserving geometric relationships in high-dimensional data during binary hashing for efficient nearest neighbor search.
• To improve robustness and feature learning by modeling a noise-removing function that projects points near the data manifold back onto it.
• To develop a data-dependent hashing method that outperforms existing state-of-the-art techniques in recall and computational efficiency.
• To integrate first-order geometric information (via Jacobian approximation) into the auto-encoder training objective for better Hamming space distribution.

Proposed method

• Introduces a novel noise-removing function that projects points near the data manifold back onto it, with a focus on local geometric consistency.
• Approximates the optimal noise-removing function using its first-order Taylor expansion (Jacobian matrix), which is analytically derived and integrated into the loss function.
• Uses a three-layer auto-encoder architecture to learn compact binary codes while preserving the geometric structure of the original data.
• Applies a 1-norm approximation to the quantization error between learned features and ideal binary codes to encourage optimal distribution in Hamming space.
• Imposes constraints on the hidden layer to enforce binary code learning, combining auto-encoder reconstruction with binary code optimization.
• Leverages the Jacobian matrix to approximate the tangent space of the data manifold, enhancing local structure preservation.

Experimental results

Research questions

• RQ1 Can a first-order approximation of a geometrically motivated noise-removing function improve binary hashing performance?
• RQ2 How does incorporating Jacobian-based geometric constraints enhance the preservation of local data relationships in Hamming space?
• RQ3 Can a data-dependent auto-encoder model outperform existing state-of-the-art hashing methods on large-scale, high-dimensional datasets?
• RQ4 To what extent does the proposed method maintain high recall at low bit budgets compared to data-independent and other data-dependent hashing techniques?

Key findings

• On the NUS-WIDE dataset, Auto-JacoBin achieves 92% Recall@1000 with only 64 bits, significantly outperforming the previous best method (NOKMeans at 81%).
• For the GIST1M dataset, Auto-JacoBin achieves 90% Recall@1000 with K=2,000, compared to K>6,000 for other methods, indicating superior computational efficiency.
• The method shows consistent performance gains as bit count increases, with m-Recall improving across all datasets and bit settings.
• Auto-JacoBin outperforms data-independent LSH when bit budgets are low, despite LSH's strong performance at high bit counts.
• The method maintains strong performance on SIFT1M and achieves comparable results to state-of-the-art methods like NOKMeans, OKMeans, and ITQ on local feature datasets.
• The use of Jacobian-based constraints leads to better local structure preservation, as evidenced by improved recall and robustness across varying K and bit settings.