-
Beyond Single-Model Views for Deep Learning: Optimization versus Generalizability of Stochastic Optimization Algorithms
Authors:
Toki Tahmid Inan,
Mingrui Liu,
Amarda Shehu
Abstract:
Despite an extensive body of literature on deep learning optimization, our current understanding of what makes an optimization algorithm effective is fragmented. In particular, we do not understand well whether enhanced optimization translates to improved generalizability. Current research overlooks the inherent stochastic nature of stochastic gradient descent (SGD) and its variants, resulting in…
▽ More
Despite an extensive body of literature on deep learning optimization, our current understanding of what makes an optimization algorithm effective is fragmented. In particular, we do not understand well whether enhanced optimization translates to improved generalizability. Current research overlooks the inherent stochastic nature of stochastic gradient descent (SGD) and its variants, resulting in a lack of comprehensive benchmarking and insight into their statistical performance. This paper aims to address this gap by adopting a novel approach. Rather than solely evaluating the endpoint of individual optimization trajectories, we draw from an ensemble of trajectories to estimate the stationary distribution of stochastic optimizers. Our investigation encompasses a wide array of techniques, including SGD and its variants, flat-minima optimizers, and new algorithms we propose under the Basin Hop** framework. Through our evaluation, which encompasses synthetic functions with known minima and real-world problems in computer vision and natural language processing, we emphasize fair benchmarking under a statistical framework, comparing stationary distributions and establishing statistical significance. Our study uncovers several key findings regarding the relationship between training loss and hold-out accuracy, as well as the comparable performance of SGD, noise-enabled variants, and novel optimizers utilizing the BH framework. Notably, these algorithms demonstrate performance on par with flat-minima optimizers like SAM, albeit with half the gradient evaluations. We anticipate that our work will catalyze further exploration in deep learning optimization, encouraging a shift away from single-model approaches towards methodologies that acknowledge and leverage the stochastic nature of optimizers.
△ Less
Submitted 1 March, 2024;
originally announced March 2024.
-
A Survey on the Use of AI and ML for Fighting the COVID-19 Pandemic
Authors:
Muhammad Nazrul Islam,
Toki Tahmid Inan,
Suzzana Rafi,
Syeda Sabrina Akter,
Iqbal H. Sarker,
A. K. M. Najmul Islam
Abstract:
Artificial intelligence (AI) and machine learning (ML) have made a paradigm shift in health care which, eventually can be used for decision support and forecasting by exploring the medical data. Recent studies showed that AI and ML can be used to fight against the COVID-19 pandemic. Therefore, the objective of this review study is to summarize the recent AI and ML based studies that have focused t…
▽ More
Artificial intelligence (AI) and machine learning (ML) have made a paradigm shift in health care which, eventually can be used for decision support and forecasting by exploring the medical data. Recent studies showed that AI and ML can be used to fight against the COVID-19 pandemic. Therefore, the objective of this review study is to summarize the recent AI and ML based studies that have focused to fight against COVID-19 pandemic. From an initial set of 634 articles, a total of 35 articles were finally selected through an extensive inclusion-exclusion process. In our review, we have explored the objectives/aims of the existing studies (i.e., the role of AI/ML in fighting COVID-19 pandemic); context of the study (i.e., study focused to a specific country-context or with a global perspective); type and volume of dataset; methodology, algorithms or techniques adopted in the prediction or diagnosis processes; and map** the algorithms/techniques with the data type highlighting their prediction/classification accuracy. We particularly focused on the uses of AI/ML in analyzing the pandemic data in order to depict the most recent progress of AI for fighting against COVID-19 and pointed out the potential scope of further research.
△ Less
Submitted 3 August, 2020;
originally announced August 2020.
-
Investigating usability of mobile health applications in Bangladesh
Authors:
Muhammad Nazrul Islam,
Md. Mahboob Karim,
Toki Tahmid Inan,
A. K. M. Najmul Islam
Abstract:
Background: Lack of usability can be a major barrier for the rapid adoption of mobile services. Therefore, the purpose of this paper is to investigate the usability of Mobile Health applications in Bangladesh.
Method: We followed a 3-stage approach in our research. First, we conducted a keyword-based application search in the popular app stores. We followed the affinity diagram approach and clus…
▽ More
Background: Lack of usability can be a major barrier for the rapid adoption of mobile services. Therefore, the purpose of this paper is to investigate the usability of Mobile Health applications in Bangladesh.
Method: We followed a 3-stage approach in our research. First, we conducted a keyword-based application search in the popular app stores. We followed the affinity diagram approach and clustered the found applications into nine groups. Second, we randomly selected four apps from each group (36 apps in total) and conducted a heuristic evaluation. Finally, we selected the highest downloaded app from each group and conducted user studies with 30 participants.
Results: We found 61% usability problems are catastrophe or major in nature from heuristic inspection. The most (21%) violated heuristic is aesthetic and minimalist design. The user studies revealed low System Usability Scale (SUS) scores for those apps that had a high number of usability problems based on the heuristic evaluation. Thus, the results of heuristic evaluation and user studies complement each other.
Conclusion: Overall, the findings suggest that the usability of the mobile health apps in Bangladesh is not satisfactory in general and could be a potential barrier for wider adoption of mobile health services.
△ Less
Submitted 15 April, 2020;
originally announced April 2020.
