Event

Professor Joseph Jaja (Chair)

Professor Shuvra Bhattacharyya
Professor Sanghamitra Dutta

 

Date/time: Wednesday, July 23, 2025 at 2:00 PM

 

Title: Towards Preparing AI for Dynamic Environments

 

Abstract: 

 

Modern deep learning models have achieved impressive performance with better architectures, training strategies and stronger augmentations. However, the deployment of these models in real-world, dynamic environments reveals critical gaps between their performance in static training settings and the challenges encountered in practice. Models often exhibit performance degradation over time, and lack robustness to novel and diverse inputs. This thesis proposes a framework to enable the models to address some of these challenges, such as growing capacity in the presence of incoming data, ensuring performance reliability for diverse and novel inputs. Collectively, these contributions enable the development of more resilient, efficient, and trustworthy systems designed for dynamic environments.

In the first part of this proposal, we introduce a novel efficient architecture expansion strategy that enables models to expand their capacity with minimal computational overhead, preparing them for evolving data complexity. We analyze this problem in a federated setting, where the model encounters a growing stream of data from a large number of distributed clients. We study this in the context of two significant bottlenecks: communication requirements and impact on the clients' resources. As the performance of a model saturates, the model capacity is increased with the help of function-preserving transformations, and the overall process continues until the desired performance is achieved. The proposed approach can substantially reduce communication and client computation at the majority portion of training while achieving comparable accuracy compared to existing strategies.

Next, we establish methodologies for assessing performance reliability to ensure that the deep models function robustly across diverse populations, a crucial prerequisite for trustworthy AI in critical domains. The current models, developed on some training datasets, may exhibit performance degradation and biases when deployed in real-world settings. We develop and analyze high-performing AI models on diverse mammography datasets. Specifically, we evaluate how these models perform across different subgroups, defined by six key attributes, to detect potential biases using a range of classification metrics. Our analysis identifies certain subgroups that demonstrate notable under-performance, highlighting the need for ongoing monitoring of these subgroups' performance. To address this, we adopt a monitoring method designed to detect performance drifts over time. Upon identifying a drift, this method issues an alert, which can enable timely interventions.

In the third part, we develop a low-cost enhancement technique for identifying test samples from novel semantic classes that are not part of the training classes; this problem is also known as the Open Set Recognition (OSR), a task that is crucial in many practical scenarios. Existing OSR methods use a constant temperature to the logits before applying a loss function, which hinders the model from exploring both ends of the spectrum in representation learning -- from instance-level to semantic-level features. We address this problem by enabling temperature-modulated representation learning using our novel negative cosine scheduling scheme. Our scheduling lets the model form a coarse decision boundary at the beginning of training by focusing on fewer neighbors, and gradually prioritizes more neighbors to smooth out rough edges. We implement the proposed scheme on top of a number of baselines, using both cross-entropy and contrastive loss functions as well as a few other OSR methods, and find that our scheme boosts both the OSR performance and the closed set performance in most cases, especially on the tougher semantic shift benchmarks.

Building on the insight that better OSR methods are fundamentally dependent on the quality of learned representations, we plan to explore techniques for extracting a richer set of representations. The objective is to design learning frameworks that aim to extract diverse representations, thereby improving the model's ability to separate the known classes from the unknown. While OSR typically addresses unseen semantic classes, deployed models may face a broader spectrum of out-of-distribution (OOD) data, for example because of co-variate shifts, corruptions or unrelated examples, which also were not part of the training distribution. For these inputs, the model's outputs are unreliable. Therefore, future work will focus on adapting and rigorously evaluating our methods on the broader and diverse notion of OOD detection beyond unseen semantic classes.

Finally, We will address a continual test time adaptation problem designed to handle incoming batches of distributionally shifted data during inference, a common real-world scenario where a model's operational environment changes. This method involves training a lightweight adaptation module, allowing it to learn just the necessary changes in features due to shift. By integrating adaptation into a continual learning loop, we aim for a scalable and effective solution to long-term AI, enabling models to adapt and continually generalize to changing data distributions.