Posted on

AI Tools Will Help Us Make the Most of Spatial Biology


In spatial biology, we can anticipate that applying AI to cell-by-cell maps of gene or protein activity will pave the way for significant discoveries. (David W. Craig, Ph.D.)

Contributed Commentary by David W. Craig, Ph.D. and Brooke Hjelm, Ph.D.

We have heard a lot about cellular and tissue spatial biology lately, and for good reason. Tissues are heterogeneous mixtures of cells; this is particularly important in disease. Cells are also the foundational unit of life, and they are shaped by those cells proximal to them. Not surprisingly, the research field sought to survey cellular and tissue heterogeneity. The last decade saw massive adoption of single-cell sequencing RNA. This approach requires that we disaggregate cells, leading to accounting and characterization of cell populations, but at the same time losing their spatial context such as their proximity to other cells or where they fit with traditional approaches such as histopathology.

Enter Spatial Genomics

That’s why we have welcomed spatial transcriptomics and a focus on mapping RNA transcripts to their location within a tissue. After all, understanding disease pathology requires that we understand not only the underlying genomics and transcriptomics but also the relationship between cells and their relative locations within a tissue. Along for the ride: new avenues for the study of cancer, immunology, and neurology, among many others. What’s changed is the emergence of new tools for resolving spatial heterogeneity. SeqFISH and MerFISH are novel approaches for mapping gene expression within model systems. Multiple companies such as 10x Genomics and NanoString are now democratizing access to spatial transcriptomics, introducing new technologies and assays. They are opening up the study of disease pathology.

AI & Deep Learning: Adding to Our Vocabulary

New experimental methods often start with historical analysis approaches. Let’s consider the first step in analysis: finding clusters of spots/cells with similar gene expression and then visualizing by reducing dimensions. In single-cell RNA-seq, the tSNE projection and color-coding clustering may be the signature plot, much like the Manhattan plot was to the GWAS.

Yet, critically, we haven’t leveraged the underlying histopathology image—the foundation of diagnosis and study of disease. We haven’t leveraged the fact that two spots are neighboring. What happens when we do? What happens at the edges between two clusters? What happens when cell types intersperse or infiltrate, such as in immune response? Are there image analysis methods we aren’t considering that have a high potential impact?

Indeed, concepts such as convolutional neural networks (CNNs) and generative adversarial networks (GANs) have been instrumental in classifying features and underlying hidden layers. We can go beyond the tSNE in spatial transcriptomics—and the question should be about viewing the latent space (the representation of the data that drives classifying regions and the discovery of hidden biology). These terms and concepts are foundational when it comes to artificial intelligence and need to be front and center in spatial transcriptomics analysis.

Of course, the use of AI and deep learning terminology is ubiquitous. Getting away from the hype, from self-driving cars to the successes in image recognition (ImageNet Challenge), some of the most remarkable achievements leverage spatial and imaging data. Data matters and one then asks: should we consider a single spatial transcriptomics section as one experimental data point, or is it 4,000 images and 4,000 transcriptomes?

In spatial biology, we can anticipate that applying AI to cell-by-cell maps of gene or protein activity will pave the way for significant discoveries that we might never achieve on our own. Incorporating spatially-resolved data could be the next leap forward in our understanding of biology. There will be questions we never even knew to ask that may be answered by combining spatial transcriptomics and spatial proteomics. But to get there, we need to come together and work as a community to build up the training data sets and other resources that will be essential for giving AI the best chance at success.

We have yet to truly make the most of the spatial biology data that has been generated. If we do not address this limitation, we will continue to miss out even as we produce more and more of this information.

David W. Craig, PhD (davidwcr@usc.edu), and Brooke Hjelm, Ph.D. (bhjelm@usc.edu) are faculty within the Department of Translational Genomics, University of Southern California Keck School of Medicine.

Read More

Posted on

High-quality Deepfake Videos Made with AI Seen as a National Security Threat


Deepfake videos so realistic that they cannot be detected as fakes have the FBI concerned about they pose a national security threat. (GETTY IMAGES)

By AI Trends Staff

The FBI is concerned that AI is being used to create deepfake videos that are so convincing they cannot be distinguished from reality.

The alarm was sounded by an FBI executive at a WSJ Pro Cybersecurity Symposium held recently in San Diego. “What we’re concerned with is that, in the digital world we live in now, people will find ways to weaponize deep-learning systems,” stated Chris Piehota, executive assistant director of the FBI’s science and technology division, in an account in WSJPro.

The technology behind deepfakes and other disinformation tactics are enhanced by AI. The FBI is concerned natural security could be compromised by fraudulent videos created to mimic public figures. “As the AI continues to improve and evolve, we’re going to get to a point where there’s no discernible difference between an AI-generated video and an actual video,” Piehota stated.

Chris Piehota, executive assistant director, FBI science and technology division

The word ‘deepfake’ is a portmanteau of “deep learning” and “fake.” It refers to a branch of synthetic media in which artificial neural networks are used to generate fake images or videos based on a person’s likeness.

The FBI has created its own deepfakes in a test lab, that have been able to create artificial personas that can pass some measures of biometric authentication, Piehota stated. The technology can also be used to create realistic images of people who do not exist. And 3-D printers powered with AI models can be used to copy someone’s fingerprints—so far, FBI examiners have been able to tell the difference between real and artificial fingerprints.

Threat to US Elections Seen

Some are quite concerned about the impact of deepfakes on US democratic elections and on the attitude of voters. The AI-enhanced deepfakes can undermine the public’s confidence in democratic institutions, even if proven false, warned Suzanne Spaulding, a senior adviser at the Center for Strategic and International Studies, a Washington-based nonprofit.

“It really hastens our move towards a post-truth world, in which the American public becomes like the Russian population, which has really given up on the idea of truth, and kind of shrugs its shoulders. People will tune out, and that is deadly for democracy,” she stated in the WSJ Pro account.

Suzanne Spaulding, senior adviser, Center for Strategic and International Studies

Deepfake tools rely on a technology called generative adversarial networks (GANs), a technique invented in 2014 by Ian Goodfellow, a Ph.D. student who now works at Apple, according to an account in Live Science.

A GAN algorithm generates two AI streams, one that generates content such as photo images, and an adversary that tries to guess whether the images are real or fake. The generating AI starts off with the advantage, meaning its partner can easily distinguish real from fake photos. But over time, the AI gets better and begins producing content that looks lifelike.

For an example, see NVIDIA’s project www.thispersondoesnotexist.com which uses a GAN to create completely fake—and completely lifelike—photos of people.

Example material is starting to mount. In 2017, researchers from the University of Washington in Seattle trained a GAN can change a video of former President Barack Obama, so his lips moved consistent with the words, but from a different speech. That work was published in the journal ACM Transactions on Graphics (TOG). In 2019, a deepfake could generate realistic movies of the Mona Lisa talking, moving and smiling in different positions. The technique can also be applied to audio files, to splice new words into a video of a person talking, to make it appear they said something they never said.

All this will cause attentive viewers to be more wary of content on the internet.

High tech is trying to field a defense against deepfakes.

Google in October 2019 released several thousand deepfake videos to help researchers train their models to recognize them, according to an account in Wired. The hope is to build filters that can catch deepfake videos the way spam filters identify email spam.

The clips Google released were created in collaboration with Alphabet subsidiary Jigsaw. They focused on technology and politics, featuring paid actors who agreed to have their faces replaced. Researchers can use the videos to benchmark the performance of their filtering tools. The clips show people doing mundane tasks, or laughing or scowling into the camera. The face-swapping is easy to spot in some instances and not in others.

Some researchers are skeptical this approach will be effective. “The dozen or so that I looked at have glaring artifacts that more modern face-swap techniques have eliminated,” stated Hany Farid, a digital forensics expert at UC Berkeley who is working on deepfakes, to Wired. “Videos like this with visual artifacts are not what we should be training and testing our forensic techniques on. We need significantly higher quality content.”

Going further, the Deepfake  Detection Challenge competition was launched in December 2019 by Facebook — along with Amazon Web Services (AWS), Microsoft, the Partnership on AI, Microsoft, and academics from Cornell Tech, MIT, University of Oxford, UC Berkeley; University of Maryland, College Park; and State University of New York at Albany, according to an account in VentureBeat.

Facebook has budged more than $10 million to encourage participation in the competition; AWS is contributing up to $1 million in service credits and offering to host entrants’ models if they choose; and Google’s Kaggle data science and machine learning platform is hosting both the challenge and the leaderboard.

“‘Deepfake’ techniques, which present realistic AI-generated videos of real people doing and saying fictional things, have significant implications for determining the legitimacy of information presented online,” noted Facebook CTO Mike Schroepfer in a blog post. “Yet the industry doesn’t have a great data set or benchmark for detecting them. The [hope] is to produce technology that everyone can use to better detect when AI has been used to alter a video in order to mislead the viewer.”

The data set contains 100,000-plus videos and was tested through a targeted technical working session in October at the International Conference on Computer Vision, stated Facebook AI Research Manager Christian Ferrer.  The data does not include any personal user identification and features only participants who have agreed to have their images used. Access to the dataset is gated so that only teams with a license can access it.

The Deepfake Detection Challenge is overseen by the Partnership on AI’s Steering Committee on AI and Media Integrity. It is scheduled to run through the end of March 2020.

Read the source articles in  WSJProLive Science, Wired and VentureBeat.

Source: AI Trends