MaxVal Blog

    Filter by Category
    Filter by Category

    Go Back to Homepage

    Hot Topics in Tech | April 2019 Newsletter

    Hot_Topics_in_Tech_April_2019-1

    In the April issue of the “Hot Topics in Tech” newsletter series, we cover Generative Adversarial Networks (GAN) - a technique that promises to give imagination to artificial Intelligence (AI). This issue also briefly explores cultured meat and explores the potential impact of the recent FDA rulings on market and intellectual property activity.

    Generative Adversarial Networks: a primer

    Generative Adversarial Networks (GAN) has sparked a lot of public interest recently because of its application in generating a deepfake, a computer-generated image which looks indistinguishable from a real image.

    main

    Example of a Deepfake (Source: StarGAN GitHub)

    GAN is a popular class of neural networks used for unsupervised learning by artificial intelligence (AI) and machine learning (ML) researchers. In simple terms, GAN can be understood as an artificial con man that learns the natural features of any given dataset in order to imitate the dataset.  It is a generative model that is capable of generating plausible new data on its own, i.e. the machine-equivalent of imagination. Others popular generative models include Variational Autoencoder (VAE) and Autoregressive Models.

    The invention of GAN is credited to Ian Goodfellow - currently with Apple following his earlier stint with the Google Brain team. To trace its beginnings, we have to go back to Ian Goodfellow’s time at the University of Montreal. Ian and his colleagues were discussing a solution for synthesizing new images. The discussion lead to Ian’s grand idea - what if you pitted two neural networks against each other? GAN involves a model with two deep neural networks contending against each other (Fig. 1). One deep neural network called a generator (G) generates random or fake data and another deep neural network called a discriminator (D) learns from real data to determine if the generated data is real or fake (returns 0 or 1). After every iteration, D learns to classify data as real or fake using real data samples and G learns to generate realistic data by receiving feedback (adversarial loss) from the discriminator.
    Generative_Adversarial_Networks

     Fig. 1: Generative Adversarial Networks

    We analyzed data from Google trends to investigate GAN popularity in general. Google Trends revealed a spike in interest for the year 2017 and 2018 resulting in > 500 hits related to this technology (Fig. 2). The non-patent literature on GAN can be traced back to the first GAN paper published on June 2014 by Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Since its publication, the number of research articles in this area has increased exponentially with about 2930 articles published in 2018 alone (Fig. 3). One of the widely circulated articles was the NVIDIA paper on StyleGAN published in December 2018, which revealed a dataset of highly varied and high-quality images of faces of human beings who have never existed. A similar trend was observed in patent literature, where a 9-fold increase in patent publications was observed in 2018 (Fig. 4). The increase in patent filings indicates a growing interest in monetizing GAN. GAN research is now being actively pursued by many Artificial Intelligence/Machine Learning researchers, technology giants and startups for its applications in diverse areas such as gaming, fashion, social media, healthcare and medicine.  

    Google_Trends_hits_for_Generative_Adversarial_Networks

    Scholarly_Articles_related_to_Generative_Adversarial_Networks

    Fig. 2: Google Trends hits for Generative Adversarial Networks      Fig. 3: Scholarly Articles related to Generative Adversarial Networks

    Patent_Publications_related_to_Generative_Adversarial_Networks

     Fig. 4: Patent Publications related to Generative Adversarial Networks

    Technologies that involve generating images or videos, for instance, are being transformed with GAN. Game development and animation production have started using GAN to auto-generate and colorize characters. The world’s first AI generated news anchor was unveiled in China last year and made a tireless promise to accompany the audience 24 hours a day, 365 days a year. Pose guided person image generation may soon have large implications in the field of modeling and marketing. Cross domain GAN, such as CycleGAN and DiscoGAN - with their ability to transform from one domain to another - are expected to be among the first commercial applications of the technology. Fashion bloggers use tools such as PixelDTGAN for suggesting merchandise based on celebrity images. AttnGAN converts our text to images has varied applications for everyday use. In the field of healthcare, GAN is already causing its initial ripples. Generative adversarial autoencoders have been proposed for use in cancer drug discovery. AnoGAN and f-AnoGAN are models for anomaly detection and disease diagnosis in patients.

    GANs have transformed the way we are using deep learning neural networks.  Despite their tremendous potential, GANs still have limitations. For example, the training process needs to ensure balance and synchronization of two adversarial networks, otherwise it is difficult to obtain good training results. Poor interpretability of neural networks continues to be an issue. Further, although the samples generated by GANs are diverse, there exists a “collapse mode” problem, scenarios in which the generator makes multiple images that contain the same color or texture themes resulting in little differences. These limitations provide researchers with further cause to work towards the goal of GAN - a model akin to human imagination.

    Innovations In Cultured Meat Industry To Benefit From FDA and USDA oversight

    Cultured meat (or clean meat) refers to meat grown in cell cultures rather than meat obtained from animal slaughter. It promises to provide healthier, safer, and disease-free meat to consumers, and hopes to mitigate the negative effects associated with industrial animal agriculture. Consequently, it has received the support of humane societies, visionaries such as Richard Branson and Bill Gates, established players such as Tyson foods and Cargill. The clean meat industry has been earmarked for its high growth potential and commercial relevance within the next five to ten years by the National Academies of Science, Engineering, and Medicine (NASEM).

    Noteworthy players in the cultured meat market are MosaMeat, Memphis Meats, SuperMeat, Future Meat Technologies, IntegriCulture, Finless Foods, Tyson Foods, General Mills, and Cargill. Several companies have been mimicking meat and meat products using plant and microbial substitutes for several decades now including Quorn, Procter & Gamble, General Foods, General Mills, Cargill, Impossible foods, Solae LLC, Firmenich, and Savage River Inc. The world’s first hamburger made from cells was unveiled by Mosameat’s co-founder Mark Post back in 2013. This lab-grown hamburger cost $325,000. By 2015, the cost for the same burger fell to $11. In 2018, Mosameat got $8.8 million in funding towards its goal of commercializing cultured meat by 2021. Memphis Meats, a startup that produced the world’s first cultured chicken strips in 2016, is also among the notable players in the cultured meat market. Tyson Foods invested $2.2 million in Future Meat Technologies (FMT) in May 2018. FMT focuses on developing manufacturing technology that enables the cost-efficient production of fat and muscle cells, the core building blocks of meat.

    A common method that is used to produce cultured meat is to extract a tissue sample from an animal under anesthesia, collect stem cells and induce their differentiation into primitive fibers that bulk up to form muscle tissue. Although, culturing meat is theoretically considered to be adequately efficient to supply meat products to consumers, in vitro cultured meat production is still in the early stages of development and requires more research and advanced technical skills for optimized production and commercialization.

    The use of animal cell cultures in food industry goes back to several decades. While several of the big players have invested funding and resources into animal cell cultures, few of them have focused on innovations related to cultured meat production for human consumption. Patenting activity related to cultured meat peaked in 2011 and again in 2015 (Fig. 1). Patent filing is higher in China as compared to the rest of the world. The United States is in the second position with respect to the number of patent applications in this industry. Australia, Canada, South Korea and Brazil also have high numbers of patent filings related to the animal cell culture industry. Exemplary patent applications include Micro-Nature LLC publication US20140271994 for meat slurry cultures, MosaMeat’s publication GB201807326D0 for apparatus and methods that enable the production of tissue from cells.

    Cultured_Meat

    Fig. 5: Cultured meat patent dashboard showing (top left) the count of patent families with their earliest priority year since 1998, (top right) some of the top assignees in the industry, and (bottom center) application distribution based on geography.

    The need for careful labeling of food products arises from public demand on distinguishing meat substitutes, lab-grown meat and slaughtered meat. In October, 2018, the US Department of Agriculture (USDA) and the Food and Drug Administration (FDA) discussed the use of cell culture technology to develop products derived from livestock and poultry. The discussion focused on the potential hazards, oversight considerations, and labeling of cell culture based food products, particularly those from livestock and poultry. On March 7, 2019, the two agencies announced a formal agreement to jointly oversee food products derived from cultured cells. Under the agreement, the agencies agreed upon a regulatory framework wherein the FDA oversees cell collection, cell banks, cell growth and differentiation. The oversight transition from the FDA to the USDA’s Food Safety and Inspection Service (FSIS) will occur during the cell harvest stage. The FSIS will oversee the production and labeling of food products for human consumption derived from the cells of livestock and poultry.

    With the growing awareness on the effects of animal farming on climate change, deforestation and scarcity of resources, the demand for clean meat will further drive research on reducing the cost of lab-grown meat and improving its taste, appearance and nutrition. The concerns about safety for consumption among the public are very real and are one of the reasons for the slow rise in popularity among meat consumers. The cultured meat industry is likely to benefit from the recent FDA – FSIS agreement. With FDA and FSIS guidelines on quality and safety, more established companies may be willing to venture into this sector, furthering the availability of cultured meat before the end of this decade.

    Go Back to Homepage

    Featured Patent Applications of the Week: May 2, 2019
    MaxVal Provides Enhanced Automation Capabilities in Latest Release of Max-IDS™ 4.3

    About Author

    MaxVal
    MaxVal

    Related Posts
    Hot Topics in Tech | October 2019 Newsletter
    Hot Topics in Tech | October 2019 Newsletter
    Hot Topics in Tech | September 2019 Newsletter
    Hot Topics in Tech | September 2019 Newsletter
    Hot Topics in Tech | August 2019 Newsletter
    Hot Topics in Tech | August 2019 Newsletter

    Comment

    Subscribe To Blog

    Subscribe to Email Updates