ML Use Cases at Think Big Analytics with Mo Patel & Laura Frølich - #54

Model Management

Managing deep learning models presents unique challenges, especially when it comes to deploying them in production environments. Laura Frølich highlights the lack of standardized tools for model management, noting that Think Big Analytics developed an internal tool called "Think Deep" to address this gap 1. This tool aims to streamline the process of putting AI models into production, which is crucial for enterprises to realize the value of their AI investments.

All the investment they make on AI or deep learning, data science is useless unless they actually put the models into production.

--- Laura Frølich

The conversation also touches on the importance of monitoring and interpretability in model management, which are essential for maintaining model performance and trustworthiness 2.

   

Deployment Challenges

Deploying deep learning models involves overcoming several hurdles, particularly when scaling across different platforms. Laura Frølich discusses the complexities of serving models at scale, which requires integrating traditional DevOps practices with data science workflows 2. Challenges include compressing models for mobile deployment and ensuring efficient inference across various environments.

The training is definitely challenging, but being able to serve the models at scale brings in all your traditional DevOps and data ops.

--- Laura Frølich

Mo Patel adds insights from a traffic detection project, emphasizing the need for robust object detection and tracking methods to enhance assisted driving systems 3.

   

Feature Extraction

Feature extraction in deep learning models is a nuanced process that requires careful consideration of model layers. Laura Frølich explains their approach of using pre-trained models like ResNet and Inception, focusing on extracting features rather than classification 4. By freezing certain layers, they could capture relevant features for tasks like object tracking, which involves comparing feature vectors to ensure consistency across images.

We try to experiment with like till the very end or somewhere in the middle because these models are also the state of the art, have 50 layers, like many layers.

--- Laura Frølich

Mo Patel5.