1 of 1

Energy Efficient Battery Image Segmentation

Scientific Achievement

Address the over-parameterized deep learning architectures for large scientific datasets through memory, energy-efficient architecture for solving scientific tasks through computer vision problems.

Significance and Impact

Our Proposed PaSeR yields a minimum performance improvement of 174% over efficient SOTA Efficient-VIT on the Intersection-over-Union (IoU)/GigaFlop metric
Demonstration of accuracy vs memory efficiency on Li-ion battery
It can be adapted for multi-resolution datasets like pathological pyramid data, flow-battery packs, etc.

Technical Approach

Developing a framework called PaSeR by leveraging reinforcement learning policy for decision-making to select the different data patches/resolutions to send to the most effective model for achieving energy efficiency.
Demonstration of RL policy effectiveness to learn from unseen noisy contexts with complementary model strengths.
Demonstration of PaSeR computational efficiency on battery material phase segmentation, multi-resolution

PI(s)/Facility Lead(s): Person Name; Ramakrishnan Kannan

Collaborating Institutions: Stevens Institute of Technology

ASCR PM: Kalyan Perumalla and Hal Finkel

Publication(s) for this work: Bharat Srikishan, Anika Tabassum, Ramakrishnan Kannan, Srikanth Allu, Nikhil Muralidhar. Reinforcement Learning Prediction Cascades: A Case Study for Image Segmentation. AAAI 2024

(a) Memory Efficiency: Peformance w.r.t SOTA in terms of IU/GigaFLops. Our model: PaSeR

(b) Comparison : Battery material phase segmentation comparison b/w PaSER and SOTA

Talking Points:

Large Deep learning architectures for image segmentation on scientific images lead to increased accuracy, this is usually accompanied by a larger increase in computation and memory requirements during inference.
While this is a non-issue in traditional machine learning (ML) pipelines, the recent confluence of machine learning and fields like the energy-efficient architectures has rendered such large architectures infeasible for execution in low-resource settings.
Our proposed decision cascades where inputs are passed through models of increasing complexity until desired performance is achieved. We propose PaSeR (Parsimonious Segmentation with Reinforcement Learning) a non-cascading, cost-aware learning pipeline as an efficient alternative to cascaded decision architectures to overcome sub-optimal throughput and increased computational cost due to wasteful intermediate computations.
Through experimental evaluation on both real-world and standard datasets, we demonstrate that PaSeR achieves better accuracy while minimizing computational cost relative to cascaded models.
On the real-world task of battery material phase segmentation, PaSeR yields a minimum performance improvement of 174% over efficient SOTA Efficient-VIT on the Intersection-over-Union(IoU)/GigaFlop metric – that measures the segmentation performance obtained per GigaFlop of computation with respect to baselines.

METADATA:

CITATIONS:

The notes section should also include full citations (including the DOI) to any important and associated publications, datasets or code developed as part this work.

For publications, please make sure full-text copies of the accepted manuscripts have been submitted to OSTI.
Please ensure that there is at least one publication associated with the highlight. This citation should be included in the slide.

AWARDS:

If this work has received any awards at conference or otherwise, please note them here.

REPRODUCIBILITY:

Code and data are available at https://github.com/scailab/paser.

BACKGROUND AND CONTEXT INFORMATION:

You may include further background or other information that would be helpful in understanding and delivering the full impact of this research.