2020 World Finalists
2020 Asia Final
Japan
'Syrinx' is a medical device that restores the ability to speak for those who had lost their voice box for medical reasons, such as laryngeal cancer. Around the globe, over 300 thousand people lose the ability to speak each year. One of the common methods used to solve this problem is by using a hand-held medical device called 'electrolarynx'. The device consists of a single vibrator that is pressed against the throat and produces vibrations to allow speech. However, 'electrolarynx' can only produce a monotonic robot-like voice, and have the problem of occupying one hand while using. This causes social issues, such as difficulty in communication while in noisy environments or the robotic voice stands out in the public when it is heard. To solve this problem, first, we tried producing more complicated vibration waves calculated from actual human speech, such as LPC residual waves. With this method, we succeeded in producing a human-like hoarse voice, but the volume of the voice was very small that bulky audio amp was needed to make it audible. To produce quality voice efficiently, we turned to the voice production of talking birds, in our case hill mynahs. According to Japanese researchers, hill mynahs have two voice boxes and they mimic human voice by mixing the sound produced by each voice box. Syrinx was inspired by this unique characteristic of hill mynahs double voice box and utilizes two vibrators to produce quality human voice by mixing the sound produced by each vibrator. To produce user-like voice on NUTONE, 30 or more one-sentence sized voice sample taken before the voice box removal surgery is needed. Then machine learning (GMM) based voice conversion library was run on Azure Virtual Machine to produce the exact vibration pattern to produce user-like voice.
Hong Kong SAR
Hollo envisions a future of tech-based, accessible, and comprehensive personal health management tools. As a Smart AI Powered Preventative platform, Hollo, aims at improving the mental health of individuals by integrating Machine Learning with suggestive diagnosis, therapy, and continual monitoring, facilitating self-help and professional therapeutic services. By reducing the need for additional technology (wearables), we make this new tracking technology accessible for less privileged communities in the world. We hold two value propositions close to our heart. We equip those finding stress points or symtoms of mental illness in their lives with a quantitative solution to help track their mental health through different metrics like sleep quality, heart rate, stress and social media sentiment. We aim to continue to improve our research in collaboration with research departments from universities around the world. As a platform, Hollo also equips therapists (and their NGOs), who don’t have consistent, quality patient feedback to aid retentive treatment, with a back-end webapp equipped with empirical and actionable findings on their patients; using the data analysis to enhance time efficiency, and construct predictive models. Streamlining admin services and improving the screening processes for institutions to reach out to at-risk individuals.
India
In the last few years, the steep rise in Counterfeit Drug production in India has led to uncertainty in the minds of patients when they are buying drugs. There is no efficient way for patients, nurses, pharmacists, and other end users to check if a drug is authentic or not. Additionally, if a counterfeit drug is found, it is difficult to find out at which point in the supply chain the fault took place. The information printed on medicine strips can be torn, cut into, etc., thus rendering the patient with little to no information to go on when buying medicines. Therefore, we are introducing Seguro Droga - Drug Counterfeit Checking & Providing Drug Information to Patients using Blockchain. We have used the Azure Hyperledger Fabric VM to setup a supply chain management system, implemented with smart contracts. Every medicine is given a unique RFID tag at the time of manufacturing, and its progress through the supply chain can be tracked using the same RFID tag. We have also developed a desktop web app that will be able to check if a medicine is counterfeit or not. Additionally, patients can use the mobile app to select their allergen filters, and when they purchase any medicine, the app alerts them if any substance they are allergic to is in the medicines they purchased.
Indonesia
Tulibot is an integrated assistive device, created to help the communication between deaf and mainstream society. We see that the number of deaf people will increasing rapidly throughout the year. For now, there are 466 million people in the world that suffers from hearing loss. WHO also estimated that by 2050, the number of people with hearing loss will increase up to 900 million people. Commonly, people with hearing loss are located in middle and low income countries. While assistive device has a very high price, so only 5-15% people with hearing loss that able to reach assistive technology. This brings the idea to us, how can we help people with hearing loss, that located in low and middle income countries, to have a better communication and better understanding with society. How can we build some affordable technology that can help 40 million people with hearing loss in Indonesia and other country? Then, Tulibot came as the answer for all of those question. We try to build something that simple yet helpful enough for deaf people, so they can have a better communication and understanding with mainstream society. We hope that all the inequality and discrimination against deaf people will be gone. We also hope, that people in Indonesia will be able to care more to people with hearing loss. So everyone can have the same chance and opportunity, and much more equality.
2020 EMEA Final
Kenya
Pesticides have been used in weeding due to the rise in demand for higher production of food because of the growing population all over the world and also the price competitiveness of the market which made farmers move from manual labor in the fight against weeds. Technology has now evolved and there is an opportunity to eradicate these weeds mechanically without the use of herbicides whilst still increasing the efficiency of weed eradication, increasing crop yields, reducing environmental pollution caused by herbicides. The autonomous weeding bot using artificial intelligence to discriminate between weeds and crops achieved through cameras as the sensors for getting input from the environment. The vehicle uses the cameras also to navigate through the farm in between the rows of crops together with a rotary encoder. It encompasses a robotic arm for weeding in between the crop row and a plough-like weeding tool that is dragged by the robot as it passes in between the rows of crops to remove inter-row weeds. A suitable vehicle for the field conditions has been developed. It has implemented a four- wheel drive and a four-wheel steering giving it flexibility to employ different and suitable steering strategies with much ease.
Tunisia
I-Remember is a two application project destinated for the well being of the both the alzheimer’s patient and their caregivers.
Kenya
Saving wildlife for the future of the planet is everyone’s responsibility and we combine the best tools we have in tech to achieve this.
Ukraine
We are Allez and we support personal development through sports experience. Our aim isn’t just to maximize the performance of an athlete, but to help coaches to growth individuals who are mentally ready to fight obstacles.
United Kingdom
1st Place Winners of DurHack 2019 (Durham University Hackathon) Access the live application at: https://vhysio.herokuapp.com/ View our presentation slides at: http://bit.ly/2OEQrWv Our code is available on Github: https://github.com/BenHarries/Durhack2019 Vhysio is a machine learning web app utilising cutting edge in browswer deep learning with, tensorflow.js to enable accessible physiotherapy for the Visually Impaired, talking through exercises by responding to users' postures in real-time. Vhysio makes it easier for users to not only complete but to improve their techniques independently. Technology We built Vhysio in 24hrs and pitched it to over 200 people including leaders in industry. We are in the process of migrating our web app to Azure. Vhysio utilises Machine Learning to say what makes a particular position correct and incorrect. For each pose it has been trained on a dataset of images to predict whether the position is correct, or incorrect - and what makes it so and guides the user using voice. We have used 'TeachableMachine' A web-based AI Machine Learning tool to train our models in the various physiotherapy poses. Google's Speech-to-Text API was also used to enable the application to be accessible by the visually impaired. The user can start their exercises via speech remotely this is more convenient and easier to use for our target audience. The application utilises Windows WebKit Speech Recognition, for text-to-speech. This is useful for the visually impaired as they can hear if they are in the right position as the application will tell them to adjust their posture if incorrect. We also use the webcam to track the user's movement which is fed as input to the machine learning model and outputs a status on the user's posture.
2020 Americas Final
United States
Deeptector.io detects DeepFakes using the power of the same Artifical Intelligence methods that generate state-of-the-art DeepFakes in the first place. These algorithms are commonly known as deep neural networks, and they train themselves to differentiate between real videos and fake ones by looking at thousands of examples of each. Deeptector.io is trained on research-grade DeepFakes to ensure that it can detect the most sophisticated material. The neural network at the core of Deeptector.io functionality can pick out differences between real and fake videos that are invisible to the human eye, and it can do so with an overall accuracy of over 90 percent. When you upload a video to Deeptector.io, the video is passed through a convolutional nerual network one frame at a time. The neural network produces a set of features for each frame and identifies fake sections of video by looking at sequences of features, which it learns by examining over 14 million images that are publicly available in the ImageNet dataset.
United States
Tremor Vision is a web-based tool using Azure custom vision that enables physicians to detect early onset Parkinson’s and quantitatively track patient progress throughout a prescribed treatment plan.
United States
TuringCerts: UC Berkeley X HKUST Website: https://certs.turingchain.tech/en Slidedeck: http://bit.ly/turingcerts-deck-v4-pdf-regular Demo kit: http://bit.ly/turingcerts-demokit ### INTRO A privacy-first diploma validation on the blockchain that makes verification easier for hiring. ### MAIN IDEA By leveraging the incorruptible, traceable, cost-effective, and borderless characteristics of blockchain technology, TuringCerts is capable of redefining traditional educational certificates and eventually enabling unified and sustainable records tracking for the educational industry. TuringCerts not only ensures full data coverage, control, and maintenance for users but also promises a rigorous privacy policy and upholds the highest trust of the source as well as the dignity of authorized certificates. ### Traction We have beta-tested the certificate issuing service in 12 universities and government agencies: including Stanford, UC Berkeley, UNLV, HKUST, National Taiwan University, National Taipei University of Science and Technology, and dozens more. ### TOP 3 KEYS 1. Personal Milestone Book-keeping 2. W3C DID Protocol for Virtual Identity 3. GDPR Principle of Least Privilege ### Fundamental Phase TuringCerts, which is in its fundamental testing phase, is being executed on the second generation blockchain - Ethereum, and are being supervised and operated by multiple sophisticated Smart Contracts. On top of that, IPFS is utilized to parallel store digital images of certificates in large sizes. ### Advanced Stage TuringCerts has built high-level strategic cooperation in regard to technical infrastructures with National Cheng Kung University and BiiLabs. On the other hand, it is actively collaborating with the world-renowned IOTA Foundation for in-depth global adoptions. In order to achieve low latency and cost-effectiveness, TuringCerts is currently being migrated to the IOTA blockchain and connected to the open-source TangelID acceleration layer. ### (1) How to ensure uniqueness and authenticity? 1. TuringCerts adopts Decentralized Identifier Standard (DID) proposed by the World Wide Web Consortium (W3C). 2. TuringCerts provides each certificate holder with a unique identifier via a hash value to prevent being tampered with. 3. Its anti-counterfeiting feature is the cornerstone that leads to sustainable certificate storage and tracking ecosystem. ### (2) How to safeguard user privacy? TuringCerts adopts the “BlockCerts” developed by MIT Media Lab's early open-sourced certificate infrastructure, which is inspired by the “EU GDPR's Least Privilege Principle” so as to grant the verification party with least yet most sufficient information. In addition, TuringCerts utilizes an asymmetric encryption scheme such as RSA, which guarantees high privacy even during information transfer. ### (3) How to sustain TuringCerts's footprint as a global leader? TuringCerts is affiliated with the Xcelerator from UC Berkeley Blockchain Lab, from whom will continuously on-board and integrate the latest technical breakthroughs in Silicon Valley. This collaboration enables TuringCerts to maintain its industry-leading technology evolutions and security characteristics. Finally, TuringCerts is being supported by its geographical advantage and relevant user pool to conduct social experiments in Hong Kong and Taiwan campuses. ### Why did you pick this idea to work on? Video: [http://bit.ly/turingcerts-elevator-pitch](http://bit.ly/turingcerts-elevator-pitch) When I was applying for graduate schools in the US, it took me several weeks sending my transcripts, graduation diploma, certificates, and a variety types of documents from one place to another: Taiwan to Hong Kong, and Hong Kong to the US. The verification fee and post-office fee are not low, which creates an unnecessary time and financial burden for all the hundreds of millions of international students every year. Right at the time when I got in UC Berkeley, I made up my mind to solve this for all the student generations since this was a troublesome pain to me. ### Do you have domain expertise in this area? I worked on blockchain research several years ago during my study in Hong Kong (HKUST) and published three blockchain papers in ACM MobiSys'18, the CPE Journal 2018, and IEEE DappCon'19. The ACM paper even received the best paper award in Munich, Germany. I invited the co-author of the three papers to co-found the idea and work on a probable solution that has been protected by our patent filing. The father of the co-founder is a senior HR manager in a top-tier company, who advising us on proceeding the product into the HR management system that creates huge cost-down and fewer workloads for the process of candidate background screening. The HRs and the admission offices of universities saved a significant amount of efforts on making calls and sending emails only for background checking.
United States
According to the National Center for Biotechnology Information (NCBI), approximately one in eight women worldwide carry a lifelong risk of contracting breast cancer. Breast Cancer is the most common form of cancer and is the 2nd leading cause of death among women. Breast cancer detection is often a laborious process, requiring either the use of mammograms which requires expensive Radiology instruments, trained radiologists, and a clinic with supporting infrastructure, or the use of Fine Needle Aspirations (FNA’s) to perform a biopsy of breast tissue at regular intervals typically after the age of 30 in females. In the case of performing a biopsy, a trained pathologist then needs to evaluate multiple samples of breast tissue in order to detect instances of breast cancer by using a microscope. This is difficult as even the most skilled pathologists can only spot instances of breast cancer approximately 85% percent of the time requiring them to evaluate many samples. According to a study published in the paper Diagnostic Concordance Among Pathologists Interpreting Breast Biopsy Specimens by Elmore et. al., the accuracy of pathologists analyzing breast biopsies for four different types of breast cancer varies anywhere from 48% for atypical hyperplasia to 96% for invasive carcinoma. The entire process takes an average of five days and up to three weeks depending on the size of the backlog of samples a particular clinic has. The involvement of pathologist analysis of makes current biopsies expensive; each one around five hundred dollars. It also makes them slow; After sample collection, a clinic would have to transport the sample and wait for the pathologist to analyze it before finally receiving the pathology report. Finally, it makes them inaccurate, as humans have to simultaneously consider dozens of factors; skilled pathologists have been shown to have 12% false negative and 16% false positive rates. In the US alone there are approximately 1.6 million breast biopsies performed every year. As the number of occurrences of breast cancer continue to rise, the number of screenings worldwide will continue to grow, requiring either additional resources or tools to address. In many developing countries, the aforementioned problems are only further exacerbated. The cost and lack of skilled labor makes biopsy based breast cancer detection very difficult or impossible in many cases, resulting in women not getting regular preventative screening for breast cancer. In cases where breast cancer screening is available, This creates a situation where breast cancer can metastasize, or develop beyond the point of treatment. In fact, according to NCBI, approximately 60% of deaths due to breast cancer occur in the developing world. Given the problems associated with the requirement of a pathologist for breast tissue analysis, our team proposes to create a lightweight deep learning model hosted on a website where clinics can directly upload microscope images, eliminating the need for a pathologist altogether. Our platform, Pink AI, would enable a dramatic decrease in cost for breast cancer detection and provide a diagnosis in real time instead of taking days. Pink AI is designed to run on any device with a web browser including smart phones and laptops, making it ideal for clinics in developing countries without advanced infrastructure to use our platform for breast cancer prediction. Pink AI has a groundbreaking accuracy of 99%, making it more accurate at detecting four distinct types of breast cancer than even the most skilled of pathologists.
Canada
Taking a drug from research to market is estimated to cost on average 2.6 billion dollars, and more than fifteen years on average. Currently, researchers are faced with the challenge of manually examining and testing thousands of different compounds to arrive at five or less that make it through to clinical trials and demonstrate anti-symptom, drug-like properties. For each compound that the researcher must test, they must facilitate the procurement of the compound, which can be incredibly difficult if the compound is novel. In this case, researchers need to manually create a retrosynthesis pathway for the compound. The Solution: Synbiolic In response to major inefficiencies involved in the process of drug discovery, we developed Synbiolic, a platform that leverages machine learning to accelerate rational drug design by generating novel molecules with user-specified properties and creating the retro-synthesis pathway. Overview of Synbiolic’s Pipeline Specifically, we are deploying a variational auto-encoder, a type of machine learning model that generates new data, trained on two separate dataset, a MUV (maximum unbiased validation) dataset of 74,000 molecules to validate the generalizability of our model and another dataset of 1 million molecules from ChEMBL to generate variations of molecular compounds that exhibit drug-like characteristics. The generated compounds are filtered by computing its quantitative estimation of drug-likeness (QED), and only choosing molecules with a high score (>0.5). To generate molecules with the desired property, Synbiolic leverages a type of reinforcement learning approach called policy approximation to train the generate model to create molecules with specific property in the specified range. Synbiolic accomplishes so by designing effective reward functions that entice the model to generate models within desired property range. To create retro-synthesis pathways to help facilitate the researchers to synthesize the generated compounds, Synbiolic employs the Monte Carlo tree search algorithm. This model is trained on over 1.2 M reactions gathered from a dataset of US Chemical Reaction Patents. Our software is developed using Tensorflow (an open-source machine learning library), RDKit (an open-Source Cheminformatics Software), OpenChem. Using Synbiolic, researchers can greatly reduce inefficiencies associated with drug discovery, being able to create medicine faster and cheaper. Our goal with Synbiolic is to give everyone in the world access to medicine, reducing poverty, disease, and creating a better future for humanity. Unique Value Proposition The use of artificial intelligence in drug discovery is extremely new and none has yet to be implemented in most research institutions for drug discovery. Currently, there are only a few companies that are researching and developing such services like Cyclica which we’ve reached out to, Insilico Medicine, and etc. While companies do exist that use AI, Synbiolic’s approach is much different and creates a huge edge over competitors. The generative models used are unique to our project and can go toe-to-toe with other states of the art models, and often even outperform them. Other approaches to generate molecules such as leveraging only an encoding-decoding method fail to control the property of the generated molecules, in fact, these approaches are unable to control the property of molecules. Another approach that uses recurrent neural networks used by some of Synbiolic’s competitors in the AI and drug discovery space also fails to control the property of molecules well. Synbiolic uses a novel approach leveraging machine learning technologies such as reinforcement learning, variational autoencoders, and memory-augmented recurrent neural networks to construct our generative model which is capable of effectively controlling the property of generated molecules, thus its ability to generate molecules with desired effect. For a more in-depth technical explanation of our project, visit our two medium articles below: https://medium.com/datadriveninvestor/drug-design-made-fun-using-reinforcement-learning-212a4f867f33 --> Explains how we can leverage reinforcement learning to design novel molecules with the desired effect. https://towardsdatascience.com/unlocking-drug-discovery-through-machine-learning-part-1-8b2a64333e07 --> Explains how we can use AI to generate novel compounds. Github: https://github.com/joeym-09/Leveraging-VAE-to-generate-molecules --> Starter Code for Part 1 of Synbiolic's pipeline, generating novel molecules. https://github.com/aryanmisra/synbiolic --> Code for Part 2 of Synbiolic's pipeline, using reinforcement learning to output retrosynthesis pathways for synthesizing molecules. Synbiolic's website: https://synbiolic.com/
2020 China Final
China
项目详细介绍 一、项目概述： Muse人工智能作曲机项目包含了底层算法、视频标签匹配算法、Muse小程序及谱曲辅助智能等四部分： 底层算法部分：输入为权重矩阵和指定参数（包括：节拍、乐器、时长及是否回环），应用算法模型进行谱曲，生成MIDI乐谱，最终通过MIDI2MP3生成音频文件。算法模型的生成方式为：首先使用由单向LSTM网络构成的生成网络进行生成，然后再交给双向LSTM构成的对抗网络进行处理； 视频标签匹配算法：我们首先会从视频中提取关键帧截图，然后利用图像识别技术，获取其中的元素信息（此部分使用Azure认知服务），继而对产生的标签进行匹配处理，将其匹配到128维权重矩阵，继而调用底层算法部分生成音乐； Muse灵感笔记（微信小程序）：做为快速创意记录工具使用，具体功能为：笔记查找、笔记新建和用户设置。其中笔记新建中，当用户持续创作时，将播放音乐，当用户停止创作超过指定时间时，将停止播放背景音乐，其中播放的音乐为底层算法生成并通过HLS协议下发的实时音频流； 谱曲辅助智能：通过简单的几个选项和风格，即可调用底层算法生成多个Demo音乐小节，为作曲家提供可能性的灵感，并可使用部分小节继而编曲。 二、市场应用： 底层算法+视频标签匹配算法：自动视频谱曲，即通过视频识别+音频谱曲的方案，根据视频主题、节奏及色调，快速生成符合主题和情感元素的视频背景乐，高效实现视频配乐功能，并有效降低配乐成本。 底层算法+Muse灵感笔记：
2020 Pakistan National Final
Pakistan
Flowlines is a social media experience that goes beyond the screen and strengthens your connection with world. We want to make every contact more valuable, to meet exciting new people and visit the places they cherish. With every post, we provide you with a new place to discover. How do we do this? A user draws a hologram in AR called a 'Flow'. The Flow is anchored to a specific location of significance in the world using Azure spatial anchors and posted on the newsfeed as a video. Whenever a user comes across an intriguing post, the app guides them to the Flow. Once inside the GPS location, the app asks the user to scan the area to find it. The Flow appears at the exact location it was left tethered to the world, providing the person with a new creation and new environment to explore. Users can also take someone elses AR hologram and add to it in their own post, creating a chain of Flows. This is Flowlines, connecting people with the world.