Yoojin Oh

Yoojin Oh, Junhyug Noh

Class Activation Mapping (CAM) and its extensions have become indispensable tools for visualizing the evidence behind deep network predictions. However, by relying on a final softmax classifier, these methods suffer from two fundamental distortions: additive logit shifts that arbitrarily bias importance scores, and sign collapse that conflates excitatory and inhibitory features. We propose a simple, architecture-agnostic dual-branch sigmoid head that decouples localization from classification. Given any pretrained model, we clone its classification head into a parallel branch ending in per-class sigmoid outputs, freeze the original softmax head, and fine-tune only the sigmoid branch with class-balanced binary supervision. At inference, softmax retains recognition accuracy, while class evidence maps are generated from the sigmoid branch -- preserving both magnitude and sign of feature contributions. Our method integrates seamlessly with most CAM variants and incurs negligible overhead. Extensive evaluations on fine-grained tasks (CUB-200-2011, Stanford Cars) and WSOL benchmarks (ImageNet-1K, OpenImages30K) show improved explanation fidelity and consistent Top-1 Localization gains -- without any drop in classification accuracy. Code is available at https://github.com/finallyupper/beyond-softmax.

Fatemeh Pesaran Zadeh, Yoojin Oh, Gunhee Kim

Aligning large VLMs with human preferences is a challenging task, as methods like RLHF and DPO often overfit to textual information or exacerbate hallucinations. Although augmenting negative image samples partially addresses these pitfalls, no prior work has employed listwise preference optimization for VLMs, due to the complexity and cost of constructing listwise image samples. In this work, we propose LPOI, the first object-aware listwise preference optimization developed for reducing hallucinations in VLMs. LPOI identifies and masks a critical object in the image, and then interpolates the masked region between the positive and negative images to form a sequence of incrementally more complete images. The model is trained to rank these images in ascending order of object visibility, effectively reducing hallucinations while retaining visual fidelity. LPOI requires no extra annotations beyond standard pairwise preference data, as it automatically constructs the ranked lists through object masking and interpolation. Comprehensive experiments on MMHalBench, AMBER, and Object HalBench confirm that LPOI outperforms existing preference optimization methods in reducing hallucinations and enhancing VLM performance. We make the code available at https://github.com/fatemehpesaran310/lpoi.

Yoojin Oh, Shao-Wen Wu, Marc Toussaint et al.

We propose a formalism for shared control, which is the problem of defining a policy that blends user control and autonomous control. The challenge posed by the shared autonomy system is to maintain user control authority while allowing the robot to support the user. This can be done by enforcing constraints or acting optimally when the intent is clear. Our proposed solution relies on natural gradients emerging from the divergence constraint between the robot and the shared policy. We approximate the Fisher information by sampling a learned robot policy and computing the local gradient to augment the user control when necessary. A user study performed on a manipulation task demonstrates that our approach allows for more efficient task completion while keeping control authority against a number of baseline methods.

Yoojin Oh, Marc Toussaint, Jim Mainprice

In this paper we present a framework for the teleoperation of pick-and-place tasks. We define a shared control policy that allows to blend between direct user control and autonomous control based on user intent inference. One of the main challenges in shared autonomy systems is to define the arbitration function, which decides when to let the autonomous agent take over. In this work, we propose a model and training method to learn the arbitration function. Our model is based on a recurrent neural network that takes as input the state, intent prediction scores and user command to produce an arbitration between user and robot commands. This work extends our previous work on differentiable policies for shared autonomy. Differentiability of the policy is desirable to further train the shared autonomy system end-to-end. In this work we propose training of the arbitration function by using data from user performing the task with shared control. We present initial results by teleoperating a gripper in a virtual environment using pre-trained motion generation and intent prediction. We compare our data aggregation training procedure to a handcrafted arbitration function. Our preliminary results show the efficacy of the approach and shed light on limitations that we believe demonstrate the need for user adaptability in shared autonomy systems.