Introduction: This issue of AI Briefing 20210115 will bring you 9 pieces of news, interesting and informative~

This article has a total of 3500 words, and it will take about 7 to 12 minutes to read through.

1. EdgeBERT: Extreme Compression, 13 Times Lighter than ALBERT! Will BERT Run on Raspberry Pi Soon?

There are two types of engineering that are extremely difficult in this world: the first is to maximize something very common, such as expanding a language model to be able to write poetry, articles, and code like GPT-3; the other is to minimize something very common. For NLPers, the most urgent target for this “small engineering” is none other than BERT.

From the 109M parameters BERT in 2018, to the 52M parameters distilled DistilBERT, to the 14.5M parameters distilled TinyBERT, and finally to the 12M parameters ALBERT with shared layers, the once cumbersome BERT that was difficult to load parameters on clusters can now even run on mobile platforms. As we cheer for the lightweight BERT, a group of people has stepped up – just the mobile side is not enough!

Their dream is to run BERT on IoT devices, on low-power chips, on every electronic device we can reach!

This group of software and hardware geeks from Harvard/Tufts/HuggingFace/Cornell has now donned their robes, transforming into alchemists for extreme weight loss of BERT, adding many unexpected ingredients towards this seemingly impossible goal…

• Paper Title: EdgeBERT: Optimizing On-Chip Inference for Multi-Task NLP

• Paper Link:

https://arxiv.org/pdf/2011.14203.pdf

2. Taming Human Cubs with AI: This Dad Found the Hardcore Fun of Parenting

What tech-savvy dads have done to enjoy a few episodes of Netflix in peace…

For a long time, the combination of “dad” + “cute kid” has been viewed skeptically, with some even saying, “A father’s love is like a landslide.”

Of course, not all dads are so unreliable; some can manage parenting quite normally, and Agustinus Nalwan is one of them.

Agustinus Nalwan is a blogger on Medium, who has worked in computer vision, 3D/animation, and game development, and is currently employed by Australia’s largest car trading platform, carsale.com.au.

He has a two-and-a-half-year-old son named Dexie. Dexie is very lively, loves animals, especially eagles, and often pretends to fly around the house like an eagle.

So Nalwan decided to develop a new toy based on Jetson AGX Xavier to realize his dream of “soaring high”, and most importantly, to have more time to watch Netflix.

The new toy is called Griffin (a mythological griffin), and the final effect is as follows:

Griffin’s four main components:

• 3D Game Engine: Generates a 3D magical world with mountains, sky, and Griffin using a flight simulator written in OpenGL.

• Human Pose Estimation: Uses the OpenPose pose estimation model and SSD object detection model to continuously detect the player’s body posture as input to control Griffin.

• Action Mapping and Gesture Recognition: Converts body posture into meaningful actions and gestures, such as raising left/right wings, rolling body left/right, taking off, etc.

• Communication System: Uses sockets to send pose input to the 3D game engine.

3. Detailed Explanation of YOLOv5 Deployment and Code Demonstration in the Latest OpenVINO 2021R02 Version

This article analyzes the original three output layers to implement critical C++ code outputs such as boxes, classes, nms, etc., and demonstrates the code for inference of the YOLOv5s model using pure OpenVINO + OpenCV. The original text has detailed system environment explanations and code implementations with demonstration images. It is divided into the following steps:

• YOLOv5 Download and Test Run

• Model Conversion, ONNX and IR Formats

• OpenVINO SDK

• Read Model

• Set Input and Output Formats

• Set Input Image Data and Implement Inference Prediction

• Parse Output Results and Display Output.

4. This Superconducting Chip Can Run with Only 1.25% Energy Consumption of a 7nm Chip, Including Cooling Costs

A 7nm process chip can actually run 80 superconducting chips???

That’s right, the world’s first adiabatic superconducting microchip MANA has now been released. The MANA (Monolithic Adiabatic Integration Architecture) superconducting chip is made of a superconducting metal called niobium (Nb). Superconductivity occurs when the temperature drops below 10 Kelvin (-263°C).

The key component of this chip is a power-saving superconducting digital electronic structure called AQFP (Adiabatic Quantum Flux Parametron).

The AQFP structure can handle all computational problems, including data processing and data storage. Its energy consumption is only a fraction of CMOS materials and computations.

This superconducting chip can achieve an operating efficiency of 2.5GHz when processing data, matching the current required computing technology.

If further developed, the processing speed can be further optimized.

5. ICPR 2020 | Championship Technology Sharing for Large-Scale Product Image Recognition Challenge

Recently, the International Conference on Pattern Recognition (ICPR 2020) kicked off, and various workshops announced the results of their challenges. The DeepBlueAI team from China won the championship of the large-scale product image recognition challenge jointly organized by ICPR 2020, Kaggle, and JDAI.

[Challenge Introduction]

With the rapid development of internet technology and e-commerce, people’s shopping methods have gradually shifted from traditional physical store shopping to online shopping. To fully meet customers’ massive and diverse online shopping needs, AI retail systems need to quickly identify product stock-keeping units (SKU) categories from images and videos automatically. However, many SKU-level products are fine-grained and visually similar.

JDAI has built a product recognition dataset called Products-10K[1], which is currently the largest product recognition dataset, containing about 10,000 products frequently purchased by Chinese consumers, covering all categories such as fashion, 3C, food, health, and household items.

This challenge, organized by JDAI and ICPR 2020, Kaggle, etc., requires participants to develop algorithms for fine-grained classification based on the provided product images.

[Team Achievements]

The DeepBlueAI team optimized the algorithm step by step through data analysis, network structure design, and loss improvement, achieving the best single model scores of 0.70918/0.73618 on Public & Private, exceeding the best score of the second place. By using model ensemble methods, they achieved first place in Public & Private, leading the second place by two percentage points.

The DeepBlueAI team designed a simple fine-grained image recognition algorithm for large-scale fine-grained product image recognition tasks through data analysis, data augmentation, network structure design, and loss improvement. For details, please refer to the original text.

6. You Have a Message from Your 2020 ‘Personal Annual Summary’, Please Check

If you don’t know how to write an end-of-year summary, please refer to this message.

This is a summary guide that teaches you to review 2020, analyze experiences and lessons, and set goals for the new year.

Doing a personal annual summary at the end of each year has brought me some surprising changes:

• These summaries help me better understand my weaknesses.

• The annual summary improves my self-reflection, allowing me to see my growth over the past year and accept it with pleasure.

• These summaries clarify what I want, helping me plan and turn dreams into reality.

• These summaries play an important role in setting practical goals for the coming year.

In this article, I will outline a series of step-by-step methods to help you create your own annual summary so that you can fully utilize all the positive changes that occurred throughout the year and turn them into motivation and nourishment for the next year. Although I record my summary in written diary form, the format is flexible; you can also write it on a computer or mobile phone. For those who prefer using electronic devices, you can refer to my concept template link, where you can log in directly and start editing. I also want to outline my summary review steps and discuss the results, hoping this will help those who want to study personal annual summaries and improve their development paths.

7. 80GB Medical Imaging Dataset! OCTA-500 Released

OCTA-500 Dataset Download Link:

https://ieee-dataport.org/open-access/octa-500

Data-driven AI technology is developing rapidly, and many internationally renowned datasets and sample libraries have become indispensable important resources for scientific research, experimental testing, and achievement promotion, and have become an essential part of academic papers.

Optical Coherence Tomography Angiography (OCTA) is a new imaging modality based on Optical Coherence Tomography (OCT) technology, which shows the three-dimensional structure of retinal blood vessels with micron-level resolution, compensating for the lack of blood flow information provided by OCT. Since OCTA technology started late and has not yet been fully popularized, there is currently a lack of publicly available datasets for researchers.

To promote the development of OCTA image processing and analysis technology, Professor Chen Qiang from Nanjing University of Science and Technology and his team, editors of the Journal of Image and Graphics, have released the largest OCTA image dataset, OCTA-500. It contains three-dimensional data of OCT and OCTA modalities from 500 eyes, six projection images, four text labels, and two segmentation labels. Based on this database, they also proposed a three-dimensional to two-dimensional segmentation image projection network.

8. How to Get Started with Linear Algebra? Here’s a Python Linear Algebra Lecture Note

Project Address:

https://github.com/MacroAnalyst/Linear_Algebra_With_Python

This lecture note is designed for beginners, covering basic concepts of linear algebra, special matrices, and their applications, and provides corresponding code and illustrations.

The foundation of artificial intelligence is mathematics, and linear algebra is an important part of it. However, for those with a poor mathematical foundation, “linear algebra” is a very abstract course. How to learn linear algebra? This GitHub project introduces a beginner-level linear algebra course lecture note suitable for university students, programmers, data analysts, algorithm traders, etc., with the code written in Python.

The lecture note is set for beginners, but it is also helpful for those with some knowledge of linear algebra and calculus. Learners should have a basic understanding of Python, NumPy, Matplotlib, and SymPy (3 days of training is enough).

To make it easier for everyone to understand the code, all the code involved in the lecture note is written in an intuitive way, without opting for efficient or professional coding styles.

The project author states: These lecture notes will provide learners with the foundational knowledge most needed for subjects that heavily rely on linear algebra, such as data learning, econometrics, mathematical statistics, and control theory. After patiently studying it, you will better grasp the basic concepts of linear algebra, and then you can learn about special matrices and their applications.

9. LeetCode Practice 1: Traversing Patterns on Arrays

The importance of algorithms is something I don’t need to elaborate on; if you want to go to a big company, you must go through basic knowledge and business logic interviews + algorithm interviews. Therefore, to improve everyone’s algorithm skills, this public account will bring you an algorithm question every day from LeetCode!

No more nonsense, let’s take a look at the question.

Given an array of integers, return indices of the two numbers such that they add up to a specific target.

You may assume that each input would have exactly one solution, and you may not use the same element twice.

https://leetcode.com/problems/two-sum/

To simplify, the brute force solution is:

1# python
2for i in range(len(array)):
3    for j in range(len(array)):
4        if array[i] + array[j] == target:
5            return [i, j]

The advanced version is to use map, everyone can try it out. If you want the staff to provide code, please leave a message in the comments.