Akira’s Machine Learning news — #23

Summary for Week 31, 2021 (week of 8/1/2021)

Featured Paper/News in This Week.

Machine Learning in the Real World


— — — — — — — — — — — — — — — — — — –

In the following sections, I will introduce various articles and papers not only on the above contents but also on the following five topics.

  1. Featured Paper/News in This Week
  2. Machine Learning Use Case
  3. Papers
  4. Articles related to machine learning technology
  5. Other Topics

— — — — — — — — — — — — — — — — — — –

1. Featured Paper/News in This Week

[2107.00079] Using AntiPatterns to avoid MLOps Mistakes
A paper on the failures of MLOps. The paper presents a case study that occurred when applying MLOps to the financial industry. The paper discusses tracking data quality, tracking model history, and operating Human-in-the-Loop at multiple levels.

A tracking method that improves accuracy by using all object detection suggestionsarxiv.org

[2006.06664] Quasi-Dense Similarity Learning for Multiple Object Tracking
Normally, tracking uses only the sparse signal of the teacher, but they use the suggestions of all objects from object detection to make dense feature matching. Even without relying on motion prior distribution, our method outperforms all existing methods in MOT, BDD100K, Waymo, and TAO tracking benchmarks.

— — — — — — — — — — — — — — — — — — –

2. Machine Learning Use case

This is an introductory article on NVIDIA’s video conferencing tool, Talking Heads (GANs), which allows you to upload your formal attire in advance and make it look like you are participating in a video conference in formal attire, even if you are in your nightgown. Instead of a complete video stream, only the head position and key points are compressed for communication, reducing the communication volume to 1/10.

The Downside of AI Promotiontheconversation.com

This article argues that AI is being promoted all over the world, but does not focus on the negative aspects, such as the fact that only companies with platforms like Google can benefit from it. (Personally, I think it contains some polemics.)

— — — — — — — — — — — — — — — — — — –

3. Papers

[2011.11778] KeepAugment: A Simple Information-Preserving Data Augmentation Approach
Random data augmentation may degrade the nature of the data, but they proposed KeepAugment, a data augmentation method that maintains the confidence of the image for the label by keeping the parts that the network judges to be important and changing the others. They confirmed its effectiveness in image classification and object detection.

Deep Fake Detection from Landmark Motionarxiv.org

[2104.04480] Improving the Efficiency and Robustness of Deepfakes Detection through Precise Geometric Features
Deep Fake detection based on the movement of face landmarks after calibration. They input the landmark information directly into the RNN and the difference from the previous frame into the RNN. Achieved AUC0.999 with FaceForensics++.

Protection against adversarial attacks applicable to large networksarxiv.org

[2104.00447] Towards Evaluating and Training Verifiably Robust Neural Networks
CROWN can efficiently compute the vulnerability to perturbation by adversarial attacks, but the computational cost is very large for large networks. In this study, they proposed a relaxation of CROWN, LBP (linear bound propagation ), and obtained better performance than the conventional IBP (Interval Bound Propagation) while supporting large networks

Extending CLIP to autioarxiv.org

[2106.13043] AudioCLIP: Extending CLIP to Image, Text and Audio
By extending CLIP to include audio, they propose AudioCLIP, which performs 3-modal learning (audio, image, and language). It can perform zero-shot classification just like the original CLIP, and achieves SotA performance in speech classification.

Validating the Robustness of a Model Using 3D Synthesis Techniquesarxiv.org

[2106.03805] 3DB: A Framework for Debugging Computer Vision Models
They propose 3DB to verify the robustness of trained models by adding noise and applying changes to the background and viewpoint using 3D synthesis techniques,. By making various changes, we can see how the model makes decisions. It is available here: https://github.com/3db/3db

Reinforcement Learning x Transformer to predict the action sequencearxiv.org

[2106.01345] Decision Transformer: Reinforcement Learning via Sequence Modeling
A study of applying Transformer to offline reinforcement learning, where Transformer is used to output action sequences that achieve the desired value, rather than maximizing the value. Unlike usual RL, which does sequential prediction, Self-Attention learns by looking at the entire action sequence.

— — — — — — — — — — — — — — — — — — –

4. Articles related to machine learning technology

An article by OpenAI about the exponential increase in learning costs. Learning costs are increasing by a factor of 10 every year, but since most of the hardware resources are used for inference (actual use), companies can still afford to buy learning hardware, and this trend is likely to continue.

— — — — — — — — — — — — — — — — — — –

5. Other Topics

Books on Interviewing for Machine Learning Jobs

A book on interviewing for machine learning positions. The book is divided into two parts: Part 1 provides an overview of the machine learning interview process, what types of machine learning roles are available, what skills are required for each role, what questions are commonly asked, and how to prepare for them. The second part contains more than 200 knowledge questions covering key concepts and common misconceptions of machine learning, each with a marked difficulty level.

— — — — — — — — — — — — — — — — — — –

🌟I post weekly newsletters! Please subscribe!🌟

— — — — — — — — — — — — — — — — — — –

About Me

Twitter, I post one-sentence paper commentary.