Data Distribution Search: Deep Reinforcement Learning To Improvise Input Datasets

schedule Aug 8th 01:45 - 02:30 PM place Grand Ball Room 1 people 142 Interested

Beyond computer games and neural architecture search; practical applications of Deep Reinforcement Learning to improve classical classification or detection tasks are few and far between. In this talk, I will share a technique and our experiences of applying D-RL on improving the distribution input datasets to achieve state of the art performance, specifically on object detection tasks.

Beyond open source datasets, when it comes to building neural networks for real-world problems, dataset matters, which is often small and skewed.The talk presents a few fresh perspectives on how to artificially increase the size of datasets while balancing the data distribution. We show that these ideas result in 2% to 3% increase in accuracy on popular object detection tasks, whereas small and skewed datasets yield up to 22% increase in model accuracies.

 
 

Outline/Structure of the Talk

update on 19th April:

1. Introduction to the problem: less data, imbalanced data while solving custom classification or detection tasks. 5 mins

2. Introduction to possible solutions: weighted loss functions, data augmentations etc. 3 mins

3. Introduction to data augmentations: state of the art in selecting data augmentation, visual experimental evidence of effects of different augmentations on different types of problems or datasets 7 mins

4. Choosing the right set of augmentation for a given dataset: brief introduction to DRL. 5 mins

5. DRL problem formulation to select data augmentation policies: set of all policies, definition of reward, definition of a state, selection of policies. 5 mins

6. Details of training a DRL model to select a subset of policies: object detection dataset, choice of the base detection architecture, loss function, evaluation function. 5 mins

7. Summary of results: choosing a set of random policies vs policies discovered by DRL model. 5 mins

8. A few key take-aways: the time and compute expenses of training DRL to select the right set of augmentations, visual evidence of the effects of different policies on different problems or datasets 5 mins.

9. Applicability of this work to different applications: classification, detection, healthcare, retail, transportation 5 mins

Beyond computer games and neural architecture search; practical applications of Deep Reinforcement Learning to improve classical classification or detection tasks are few and far between. In this talk, I will share a technique and our experiences of applying D-RL on improving the distribution input datasets to achieve state of the art performance, specifically on object detection tasks.

The key take away of my talk would be an effective application of deep reinforcement learning to enhance computer vision datasets via reward driven data augmentations. In particular, by showing a few examples in retail industry, I will first highlight common issues faced by deep learning practitioners in solving real-world computer vision problems: less data and imbalanced data. Less and imbalanced data, often, doesn't spin the engine of deep learning enough to learn both discriminative as well as generative features to solve classification, detection, or generation tasks with enterprise-grade accuracies.

However, it is a well-known fact that appropriate data distribution, either via data generation or data augmentation, significantly improves the pattern learning performance of deep networks. How do we, then, introduce new data samples for different classes or objects such that, given a network architecture with pretrained weights, we achieve optimal performance? Can deep reinforcement learning help us to navigate the space of sample augmentation and sample generation strategies?

To answers these questions, in this talk, I will present a deep reinforcement learning framework that systematically chooses classes and operations to perform on those classes, be it generation or augmentation, be it a classification problem or a detection problem. We model the network performance improvement as a reward towards choosing the right set of generation and augmentation techniques.

Via extensive performance testing, we showcase that our framework is far more inexpensive to achieve the same or better result than doing a neural architecture search for a given problem. Thus, we establish, with quantitative evidence that, a well-formed data distribution trumps custom discovered neural architecture. Furthermore, we are able to achieve state of the art results on PASCAL VOC and MS COCO datasets, in using Single Shot Multibox Detection (SSD) technique, proposed in 2016, by applying our framework on the underlying datasets. In showcasing these results, I will present a few examples of a custom dataset, with its visualization, where we are able to achieve 22% performance improvement using SSD as compared focal-loss driven RetinaNet architecture or its recent variants.

In a nutshell, this talk drives a key message that cleverly scripted data can significantly boost accuracy of even a reasonably powerful deep network, across different problem regimes such as classification, detection, and generation.

Learning Outcome

questions that this talk poses:

how do we introduce new data samples for different classes or objects such that, given a network architecture with pretrained weights, we achieve optimal performance? can deep reinforcement learning help us to navigate the space of sample augmentation and sample generation strategies?

towards answering these questions,

we establish that cleverly scripted data can significantly boost accuracy of even a reasonably powerful deep network, across different problem regimes such as classification, detection, and generation

we also show that a well-formed data distribution trumps custom discovered neural architecture

we present a deep reinforcement learning framework that systematically chooses classes and operations to perform on those classes, be it generation or augmentation, be it a classification problem or a detection problem to solve less data and class imbalanced data problems

Target Audience

Deep Learning Engineers, Deep Learning Practitioners

Prerequisites for Attendees

basics of deep learning, preliminary knowledge of what is object detection, preliminary knowledge of reinforcement learning, preliminary knowledge of problems caused by data imbalance

schedule Submitted 7 months ago

Public Feedback

comment Suggest improvements to the Speaker
  • Anoop Kulkarni
    By Anoop Kulkarni  ~  6 months ago
    reply Reply

    Thanks for your proposal. In the "outline" section, is it possible for you to include bulletized list of your talk organization? Also, how much time you are scheduling for each sub-topic. This will help us in our review.

    Thanks

    ~anoop

    • Vijay Gabale
      By Vijay Gabale  ~  6 months ago
      reply Reply

      Hi Anoop,

       

      Thanks for response.

      I have updated the outline section with a brief agenda of the talk along with potential time to be spent on each sub-topic. Thanks.