Sparsifying neural networks often suffers from seemingly inevitable performance degradation, and it remains challenging to restore the original performance despite much recent progress. Motivated by recent studies in robust optimization, we aim to tackle this problem by finding subnetworks that are both sparse and flat at the same time. Specifically, we formulate pruning as a sparsity-constrained optimization problem where flatness is encouraged as an objective. We solve it explicitly via an augmented Lagrange dual approach and extend it further by proposing a generalized projection operation, resulting in novel pruning methods called SAFE and its extension, SAFE$^+$. Extensive evaluations on standard image classification and language modeling tasks reveal that SAFE consistently yields sparse networks with improved generalization performance, which compares competitively to well-established baselines. In addition, SAFE demonstrates resilience to noisy data, making it well-suited for real-world conditions.

Large excavation reclaimers are used to dig ore and transfer it to the blast furnaces that refine the ore into pure metal form. The reclaiming job consists of two operations: (1) landing a reclaimer bucket on the surface of a pile and (2) slewing its boom with rotating buckets to scoop the ore. An automatic landing method for choosing where to dig in a pile of raw ore is proposed to achieve autonomous reclaiming. The method comprises detecting the shape of a pile, extracting contour lines of the pile, obtaining the joint angles of the reclaimer and determining an optimal landing point. A 3D range finder was developed with laser radar concepts to detect the shape and height of the ore pile. A series of image processing techniques for extracting the contour line from the 3D range data of a pile is suggested. A height map is obtained from the acquired range data for a pile and a contour map is obtained through image processing steps, including interpolation and edge following. The optimal landing point of the bucket on the contour line is determined so that an overload problem does not occur in the slewing operation and the reclaiming efficiency can be maximized. The algorithm for finding the landing point requires an inverse kinematics solution for the reclaimer. The forward kinematics of the reclaimer is first obtained. A constraint equation based on the geometrical relationship is suggested to solve the inverse kinematics of the reclaimer with redundancy. The proposed method was successfully applied to a working reclaimer. An autonomous reclaimer is now being operated in the yards of Kwangyang Steelworks in Korea.

The huge reclaimers in the raw ore yard are used to dig ore and transfer it to the blast furnaces. In this paper, an automatic landing method of the reclaimer is proposed to develop an autonomous reclaimer. A 3-dimensional range finder which consists of a 2-dimensional scanner and an additional joint was developed with laser radar concepts and is installed on the roof of the reclaimer to detect the ore pile. A height map is obtained from the acquired range data for an ore pile and a contour map for the pile is obtained through some image processing steps. The optimal landing point in the contour line is determined so that an overload problem does not occur in the slewing operation and reclaiming efficiency is maximized. The algorithm for finding the landing point needs an inverse kinematics solution for the reclaimer. The reclaimer has a redundancy due to the rotating buckets at the end of the its boom. A constraint equation for solving the inverse kinematics of the reclaimer is also derived by the geometrical relationship.