Zebeth Media Solutions

machine learning

Protein programmers get a helping hand from Cradle’s generative AI • ZebethMedia

Proteins are the molecules that get work done in nature, and there’s a whole industry emerging around successfully modifying and manufacturing them for various uses. But doing so is time consuming and haphazard; Cradle aims to change that with an AI-powered tool that tells scientists what new structures and sequences will make a protein do what they want it to. The company emerged from stealth today with a substantial seed round. AI and proteins have been in the news lately, but largely because of the efforts of research outfits like DeepMind and Baker Lab. Their machine learning models take in easily collected RNA sequence data and predict the structure a protein will take — a step that used to take weeks and expensive special equipment. But as incredible as that capability is in some domains, it’s just the starting point for others. Modifying a protein to be more stable or bind to a certain other molecule involves much more than just understanding its general shape and size. “If you’re a protein engineer, and you want to design a certain property or function into a protein, just knowing what it looks like doesn’t help you. It’s like, if you have a picture of a bridge, that doesn’t tell you whether it’ll fall down or not,” explained Cradle CEO and co-founder Stef van Grieken. “Alphafold takes a sequence and predicts what the protein will look like,” he continued. “We’re the generative brother of that: you pick the properties you want to engineer, and the model will generate sequences you can test in your laboratory.” Predicting what proteins — especially ones new to science — will do in situ is a difficult task for lots of reasons, but in the context of machine learning the biggest issue is that there isn’t enough data available. So Cradle originated much of its own data set in a wet lab, testing protein after protein and seeing what changes in their sequences seemed to lead to which effects. Interestingly the model itself is not biotech-specific exactly but a derivative of the same “large language models” that have produced text production engines like GPT-3. Van Grieken noted that these models are not limited strictly to language in how they understand and predict data, an interesting “generalization” characteristic that researchers are still exploring. Examples of the Cradle UI in action. The protein sequences Cradle ingests and predicts are not in any language we know, of course, but they are relatively straightforward linear sequences of text that have associated meanings. “It’s like an alien programming language,” van Grieken said. Protein engineers aren’t helpless, of course, but their work necessarily involves a lot of guessing. One may know for sure that among the 100 sequences they are modifying is the combination that will produce The model works in three basic layers, he explained. First it assesses whether a given sequence is “natural,” i.e. whether it is a meaningful sequence of amino acids or just random ones. This is akin to a language model just being able to say with 99 percent confidence that a sentence is in English (or Swedish, in van Grieken’s example), and the words are in the correct order. This it knows from “reading” millions of such sequences determined by lab analysis. Next it looks at the actual or potential meaning in the protein’s alien language. “Imagine we give you a sequence, and this is the temperature at which this sequence will fall apart,” he said. “If you do that for a lot of sequences, you can say not just, ‘this looks natural,’ but ‘this looks like 26 degrees Celsius.’ that helps the model figure out what regions of the protein to focus on.” The model can then suggest sequences to slot in — educated guesses, essentially, but a stronger starting point than scratch. And the engineer or lab can then try them and bring that data back to the Cradle platform, where it can be re-ingested and used to fine tune the model for the situation. The Cradle team on a nice day at their HQ (van Grieken is center). Modifying proteins for various purposes is useful across biotech, from drug design to biomanufacturing, and the path from vanilla molecule to customized, effective and efficient molecule can be long and expensive. Any way to shorten it will likely be welcomed by, at the very least, the lab techs who have to run hundreds of experiments just to get one good result. Cradle has been operating in stealth, and now is emerging having raised $5.5 million in a seed round co-led by Index Ventures and Kindred Capital, with participation from angels John Zimmer, Feike Sijbesma, and Emily Leproust. Van Grieken said the funding would allow the team to scale up data collection — the more the better when it comes to machine learning — and work on the product to make it “more self-service.” “Our goal is to reduce the cost and time of getting a bio-based product to market by an order of magnitude,” said van Grieken in the press release, “so that anyone – even ‘two kids in their garage’ – can bring a bio-based product to market.”

Weka announces $135M investment on $750M valuation to change how companies move data • ZebethMedia

If there’s one thing that gets the attention of investors, even in uncertain times like these, it’s data efficiency. Data is the fuel for machine learning models and getting it from point A to point B can be expensive and time consuming. That’s why a startup that can help make that process move faster with less friction is probably going to be valuable. Such is the case with Weka, a company that has come up with a way to virtualize data to make it easier to move between sources without having to make a copy first. Today, the company announced a $135 million Series D investment on a $750 million valuation, big numbers in today’s conservative funding environment. CEO and co-founder Liran Zvibel says the company originally focused on increasing data throughput for high performance computing scenarios. “The initial focus was high performance computing environments. And that’s really where the first phase of the company started. But increasingly, over the last two or three years, people have been taking concepts from high performance computing, scientific computing and applying them in a commercial context, trying to build large enterprise workloads,” Zvibel told ZebethMedia. He says that although network and compute have sped up, especially with the increased use of GPUs to help power data-intensive workloads, storage has remained a legacy bottleneck, and that’s the weak link his company is trying to attack. “Most people have made the leap to GPUs, and they made the leap to fast networking, but they’re trying to underpin it with storage technology and storage architectures from the 1990s,” he said. He claims to have invented an entirely new way of moving data. “We’ve invented a whole set of new algorithms, data structures, control structures, even network protocols. We’re not leveraging TCP/IP. We have a network protocol that allows you to leverage RDMA zero copy-like performance even on a public cloud,” he said. “And we sat down, we implemented all of that new new kind of computer science theory. And now we can actually show customers the huge advantages they can get with this new approach.” He admits the technology in some ways sounds like it’s science fiction, so companies often start small to prove it works, and once they do they sign much bigger deals. “When we come in and we tell the story, it sounds like a fairy tale or a science fiction. So customers tend to start small. When they realize we’re actually doing what we say we do then then they really go into hyperdrive,” he said. The company offers their solution as a service, but sometimes it’s delivered on prem and increasingly in the cloud with 43% of transactions in Q3 coming from public cloud business. They currently have 300 employees and expect to double in the next 12-18 months. He says that hiring a diverse workforce is top of mind for the company. “At the end of the day, we’re hiring the best talent we can find because building that is the most important thing, but we are putting a lot of thought and effort on turning over every rock to make sure we are more and more diverse.” Today’s round of funding came from a large group that includes traditional VC firms, as well as many strategic investors. The list includes 10D, Atreides Management, Celesta Capital, Gemini Israel Ventures, Hewlett Packard Enterprise, Hitachi Ventures, Key1 Capital, Lumir Ventures, Micron Ventures, Mirae Asset Capital, MoreTech Ventures, Norwest Venture Partners, NVIDIA, Qualcomm Ventures and Samsung Catalyst. Today’s $750 million valuation doubles the previous valuation, according to the company. Weka has now raised over $293 million, per Crunchbase.

AI that sees with sound, learns to walk, and predicts seismic physics • ZebethMedia

Research in the field of machine learning and AI, now a key technology in practically every industry and company, is far too voluminous for anyone to read it all. This column, Perceptron, aims to collect some of the most relevant recent discoveries and papers — particularly in, but not limited to, artificial intelligence — and explain why they matter. This month, engineers at Meta detailed two recent innovations from the depths of the company’s research labs: an AI system that compresses audio files and an algorithm that can accelerate protein-folding AI performance by 60x. Elsewhere, scientists at MIT revealed that they’re using spatial acoustic information to help machines better envision their environments, simulating how a listener would hear a sound from any point in a room. Meta’s compression work doesn’t exactly reach unexplored territory. Last year, Google announced Lyra, a neural audio codec trained to compress low-bitrate speech. But Meta claims that its system is the first to work for CD-quality, stereo audio, making it useful for commercial applications like voice calls. Image Credits: An architectural drawing of Meta’s AI audio compression model. Using AI, Meta’s compression system, called Encodec, can compress and decompress audio in real time on a single CPU core at rates of around 1.5 kbps to 12 kbps. Compared to MP3, Encodec can achieve a roughly 10x compression rate at 64 kbps without a perceptible loss in quality. The researchers behind Encodec say that human evaluators preferred the quality of audio processed by Encodec versus Lyra-processed audio, suggesting that Encodec could eventually be used to deliver better-quality audio in situations where bandwidth is constrained or at a premium. As for Meta’s protein folding work, it has less immediate commercial potential. But it could lay the groundwork for important scientific research in the field of biology. Protein structures predicted by Meta’s system. Meta says its AI system, ESMFold, predicted the structures of around 600 million proteins from bacteria, viruses and other microbes that haven’t yet been characterized. That’s more than triple the 220 million structures that Alphabet-backed DeepMind managed to predict earlier this year, which covered nearly every protein from known organisms in DNA databases. Meta’s system isn’t as accurate as DeepMind’s. Of the ~600 million proteins it generated, only a third were “high quality.” But it’s 60 times faster at predicting structures, enabling it to scale structure prediction to much larger databases of proteins. Not to give Meta outsize attention, the company’s AI division also this month detailed a system designed to mathematically reason. Researchers at the company say that their “neural problem solver” learned from a data set of successful mathematical proofs to generalize to new, different kinds of problems. Meta isn’t the first to build such a system. OpenAI developed its own, called Lean, that it announced in February. Separately, DeepMind has experimented with systems that can solve challenging mathematical problems in the studies of symmetries and knots. But Meta claims that its neural problem solver was able to solve five times more International Math Olympiad than any previous AI system and bested other systems on widely-used math benchmarks. Meta notes that math-solving AI could benefit the the fields of software verification, cryptography and even aerospace. Turning our attention to MIT’s work, research scientists there developed a machine learning model that can capture how sounds in a room will propagate through the space. By modeling the acoustics, the system can learn a room’s geometry from sound recordings, which can then be used to build visual renderings of a room. The researchers say the tech could be applied to virtual and augmented reality software or robots that have to navigate complex environments. In the future, they plan to enhance the system so that it can generalize to new and larger scenes, such as entire buildings or even whole towns and cities. At Berkeley’s robotics department, two separate teams are accelerating the rate at which a quadrupedal robot can learn to walk and do other tricks. One team looked to combine the best-of-breed work out of numerous other advances in reinforcement learning to allow a robot to go from blank slate to robust walking on uncertain terrain in just 20 minutes real-time. “Perhaps surprisingly, we find that with several careful design decisions in terms of the task setup and algorithm implementation, it is possible for a quadrupedal robot to learn to walk from scratch with deep RL in under 20 minutes, across a range of different environments and surface types. Crucially, this does not require novel algorithmic components or any other unexpected innovation,” write the researchers. Instead, they select and combine some state-of-the-art approaches and get amazing results. You can read the paper here. Robot dog demo from EECS professor Pieter AbbeelÕs lab in Berkeley, Calif. in 2022. (Photo courtesy Philipp Wu/Berkeley Engineering) Another locomotion learning project, from (ZebethMedia’s pal) Pieter Abbeel’s lab, was described as “training an imagination.” They set up the robot with the ability to attempt predictions of how its actions will work out, and though it starts out pretty helpless, it quickly gains more knowledge about the world and how it works. This leads to a better prediction process, which leads to better knowledge, and so on in feedback until it’s walking in under an hour. It learns just as quickly to recover from being pushed or otherwise “purturbed,” as the lingo has it. Their work is documented here. Work with a potentially more immediate application came earlier this month out of Los Alamos National Laboratory, where researchers developed a machine learning technique to predict the friction that occurs during earthquakes — providing a way to forecast earthquakes. Using a language model, the team says that they were able to analyze the statistical features of seismic signals emitted from a fault in a laboratory earthquake machine to project the timing of a next quake. “The model is not constrained with physics, but it predicts the physics, the actual behavior of the system,” said Chris Johnson. one of the research leads on the

Dataloop secures cash infusion to expand its data annotation tool set • ZebethMedia

Data annotation, or the process of adding labels to images, text, audio and other forms of sample data, is typically a key step in developing AI systems. The vast majority of systems learn to make predictions by associating labels with specific data samples, like the caption “bear” with a photo of a black bear. A system trained on many labeled examples of different kinds of contracts, for example, would eventually learn to distinguish between those contracts and even extrapolate to contracts that it hasn’t seen before. The trouble is, annotation is a manual and labor-intensive process that’s historically been assigned to gig workers on platforms like Amazon Mechanical Turk. But with the soaring interest in AI — and in the data used to train that AI — an entire industry has sprung up around tools for annotation and labeling. Dataloop, one of the many startups vying for a foothold in the nascent market, today announced that it raised $33 million in a Series B round led by Nokia Growth Partners (NGP) Capital and Alpha Wave Global. Dataloop develops software and services for automating aspects of data prep, aiming to shave time off of the AI system development process. “I worked at Intel for over 13 years, and that’s where I met Dataloop’s second co-founder and CPO, Avi Yashar,” Dataloop CEO Eran Shlomo told ZebethMedia in an email interview. “Together with Avi, I left Intel and founded Dataloop. Nir [Buschi], our CBO, joined us as third co-founder, after he held executive positions [at] technology companies and [lead] business and go-to-market at venture-backed startups.” Dataloop initially focused on data annotation for computer vision and video analytics. But in recent years, the company has added new tools for text, audio, form and document data and allowed customers to integrate custom data applications developed in-house. One of the more recent additions to the Dataloop platform is data management dashboards for unstructured data. (As opposed to structured data, or data that’s arranged in a standardized format, unstructured data isn’t organized according to a common model or schema.) Each provides tools for data versioning and searching metadata, as well as a query language for querying datasets and visualizing data samples. Image Credits: Dataloop “All AI models are learned from humans through the data labeling process. The labeling process is essentially a knowledge encoding process in which a human teaches the machine the rules using positive and negative data examples,” Shlomo said. “Every AI application’s primary goal is to create the ‘data flywheel effect’ using its customer’s data: a better product leads to more users leads to more data and subsequently a better product.” Dataloop competes against heavyweights in the data annotation and labeling space, including Scale AI, which has raised over $600 million in venture capital. Labelbox is another major rival, having recently nabbed more than $110 million in a financing round led by SoftBank. Beyond the startup realm, tech giants, including Google, Amazon, Snowflake and Microsoft, offer their own data annotation services. Dataloop must be doing something right. Shlomo claims the company currently has “hundreds” of customers across retail, agriculture, robotics, autonomous vehicles and construction, although he declined to reveal revenue figures. An open question is whether Dataloop’s platform solves some of the major challenges that exist in data labeling today. Last year, a paper published out of MIT found that data labeling tends to be highly inconsistent, potentially harming the accuracy of AI systems. A growing body of academic research suggests that annotators introduce their own biases when labeling data — for example, labeling phrases in African American English (a modern dialect spoken primarily by Black Americans) as more toxic than the general American English equivalents. These biases often manifest in unfortunate ways; think moderation algorithms that are more likely to ban Black users than white users. Data labelers are also notoriously underpaid. The annotators who contributed captions to ImageNet, one of the better-known open source computer vision libraries, reportedly made a median of $2 per hour in wages. Shlomo says it’s incumbent on the companies using Dataloop’s tools to affect change — not necessarily Dataloop itself. “We see the underpayment of annotators as a market failure. Data annotation shares many qualities with software development, one of them being the impact of talent on productivity,” Shlomo said. “[As for bias,] bias in AI starts with the question that the AI developer chooses to ask and the instructions they supply to the labeling companies. We call it the ‘primary bias.’ For example, you could never identify color bias unless you ask for skin color in your labeling recipe. The primary bias issue is something the industry and regulators should address. Technology alone will not solve the issue.” To date, Dataloop, which has 60 employees, has raised $50 million in venture capital. The company plans to grow its workforce to 80 employees by the end of the year.

MLOps platform Galileo lands $18M to launch a free service • ZebethMedia

Galileo, a startup launching a platform for AI model development, today announced that it raised $18 million in a Series A round led by Battery Ventures with participation from The Factory, Walden Catalyst, FPV Ventures, Kaggle co-founder Anthony Goldbloom and other angel investors. The new cash brings the company’s total raised to $23.1 million and will be put toward growing Galileo’s engineering and go-to-market teams and expanding the core platform to support new data modalities, CEO Vikram Chatterji told ZebethMedia via email. As the use of AI becomes more common throughout the enterprise, the demand for products that make it easier to inspect, discover and fix critical AI errors is increasing. According to one recent survey (from MLOps Community), 84.3% of data scientists and machine learning engineers say that the time required to detect and diagnose problems with a model is a problem for their teams, while over one in four (26.2%) admit that it takes them a week or more to detect and fix issues. Some of those issues include mislabeled data, where the labels used to train an AI system contain errors, like a picture of a tree mistakenly labeled “houseplant.” Others pertain to data drift or data imbalance, which happens when data evolves to make an AI system less accurate (think a stock market model trained on pre-pandemic data) or the data isn’t sufficiently representative of a domain (e.g., a data set of headshots has more light-skinned people than dark-skinned). Galileo’s platform aims to systematize AI development pipelines across teams using “auto-loggers” and algorithms that spotlight system-breaking issues. Built to be deployable in an on-premises environment, Galileo scales across the AI workflow — from predevelopment to postproduction — as well as unstructured data modalities like text, speech and vision. In data science, “unstructured” data usually refers to data that’s not arranged according to a preset data model or schema, like invoices or sensor data. Atindriyo Sanyal — Galileo’s second co-founder — makes the case that the Excel- and Python script–based processes to ensure quality data is being fed into models are manual, error-prone and costly. A screenshot of the Galileo Community Edition. Image Credits: Galileo “When inspecting their data with Galileo, users instantly uncover the long tail of data errors such as mislabeled data, underrepresented languages [and] garbage data that they can immediately take action upon within Galileo by removing, re-labeling or by adding additional similar data from production,” Sanyal told ZebethMedia in an email interview. “It has been critical for teams that Galileo supports machine learning data workflows end to end — even when a model is in production, Galileo automatically lets teams know of data drifts, and surfaces the highest-value data to train with next.” The co-founding team at Galileo spent more than a decade building machine learning products, where they say they faced the challenges of developing AI systems firsthand. Chatterji led product management at Google AI, while Sanyal spearheaded engineering at Uber’s AI division and was an early member of the Siri team at Apple. Third Galileo co-founder Yash Sheth is another Google veteran, having previously led the company’s speech recognition platform team. Galileo’s platform falls into the burgeoning category of software known as MLOps, a set of tools to deploy and maintain machine learning models in production. It’s in serious demand. By one estimation, the market for MLOps could reach $4 billion by 2025. There’s no shortage of startups going after the space, like Comet, which raised $50 million last November. Other vendors with VC backing include Arize, Tecton, Diveplane, Iterative and Taiwan-based InfuseAI. But despite having launched just a few months ago, Galileo has paying customers from “high-growth” startups to Fortune 500 companies, Sanyal claims. “Our customers are using Galileo while building machine learning applications such as hate speech detection, caller intent detection at contact centers and customer experience augmentation with conversational AI,” he added. Sanyal expects the launch of Galileo’s free offering — Galileo Community Edition — will boost sign-ups further. The Community Edition enables data scientists working on natural language processing to build machine learning models using some of the tools included in the paid version, Sanyal said. “With Galileo Community Edition, anyone can sign up for free, add a few lines of code while training their model with labeled data or during an inference run with unlabeled data to instantly inspect, find and fix data errors, or select the right data to label next using the powerful Galileo UI,” he added. Sanyal declined to share revenue figures when asked. But he noted that San Francisco–based Galileo’s headcount has grown in size from 14 people in May to “more than” 20 people as of today.

Pinecone vector database can now handle hybrid keyword-semantic search • ZebethMedia

When Pinecone announced a vector database at the beginning of last year, it was building something that was specifically designed for machine learning and aimed at data scientists. The idea was that you could query this data in a format that machines understand, making it much faster. Originally this involved semantic searches where users could search based on meaning instead of specific words. It turns out, however, that as people put Pinecone to work, there were use cases where specific keywords mattered, and today the company announced that it’s now possible to conduct searches combining both semantic and keyword searches, what company founder and CEO Edo Liberty calls hybrid search. “We’ve conducted a lot of research on this topic and we found that, in fact, hybrid search ends up being better [in many cases]. It’s better in the sense that if you can combine both semantic search, this is the deep NLP encoding of sentences that gets the context and the meaning and so on, but you can also infuse that with specific keywords…the combination of those two ends up being significantly better,” Liberty told ZebethMedia. In fact he says the two complement each other well, especially in cases where industry-specific terms matter. This could be something like a doctor searching for keywords related to a specific disease. In those cases, the medical context may return better results by combining a question and some specific keywords around a given disease. He says that the keywords never take precedence over the semantic question the user is asking, but they provide some extra information to help return more meaningful results. “You might know exactly what you’re looking for, and you might be able to provide extra oomph when you make your semantic search keyword-aware – and that actually helps a lot. So I don’t want to throw away the good parts of keyword search [by relying completely on semantic search]. I don’t want the keywords to be in the driver’s seat, but I don’t to ignore them completely either,” he said. As Liberty told us at the time of the company’s $28 million Series A last year, search has become a big use case for the company: “The predominant use of the vector databases is for search, and search in the broad sense of the word. It’s searching through documents, but you can think about search as information retrieval in general, discovery, recommendation, anomaly detection and so on,” he said at the time. Pinecone launched in 2019 and has raised $38 million, per Crunchbase.

LatticeFlow raises $12M to eliminate computer vision blind spots • ZebethMedia

LatticeFlow, a startup that was spun out of Zurich’s ETH in 2020, helps machine learning teams improve their AI vision models by automatically diagnosing issues and improving both the data and the models themselves. The company today announced that it has raised a $12 million Series A funding round led by Atlantic Bridge and OpenOcean, with participation from FPV Ventures. Existing investors btov Partners and Global Founders Capital, which led the company’s $2.8 million seed round last year, also participated in this round. As LatticeFlow co-founder and CEO Petar Tsankov told me, the company currently has more than 10 customers in both Europe and the U.S., including a number of large enterprises like Siemens and organizations like the Swiss Federal Railways, and is currently running pilots with quite a few more. It’s this customer demand that led LatticeFlow to raise at this point. “I was in the States and I met with some investors in Palo Alto, Tsankov explained. “They saw the bottleneck that we have with onboarding customers. We literally had machine learning engineers supporting customers and that’s not how you should run the company. And they said: ‘OK, take $12 million, bring these people in and expand.’ That was great timing for sure because when we talked to other investors, we did see that the market has changed.” As Tsankov and his co-founder CTO Pavol Bielik noted, most enterprises today have a hard time bringing their models into production and then, when they do, they often realize that they don’t perform as well as they expected. The promise of LatticeFlow is that it can auto-diagnose the data and models to find potential blind spots. In its work with a major medical company, its tools to analyze their datasets and models quickly found more than half a dozen critical blind spots in their state-of-the-art production models, for example. The team noted that it’s not enough to only look at the training data and ensure that there is a diverse set of images — in the case of the vision models that LatticeFlow specializes in — but also examine the models. LatticeFlow founding team (from left to right): Prof. Andreas Krause (scientific advisor), Dr. Petar Tsankov (CEO), Dr. Pavol Bielik (CTO) and Prof. Martin Vechev (scientific advisor). Image Credits: LatticeFlow “If you only look at the data — and this is a fundamental differentiator for LatticeFlow because we not only find the standard data issues like labeling issues or poor-quality samples, but also model blind spots, which are the scenarios where the models are failing,” Tsankov explained. “Once the model is ready, we can take it, find various data model issues and help companies fix it.” He noted, for example, that models will often find hidden correlations that may confuse the model and skew the results. In working with an insurance customer, for example, who used an ML model to automatically detect dents, scratches and other damage in images of cars, the model would often label an image with a finger in it as a scratch. Why? Because in the training set, customers would often take a close-up picture with a scratch and point at it with their finger. Unsurprisingly, the model would then correlate “finger” with “scratch,” even when there was no scratch on the car. Those are issues, the LatticeFlow teams argues, that go beyond creating better labels and need a service that can look at both the model and the training data. LatticeFlow uncovers a bias in data for training car damage inspection AI models. Because people often point at scratches, this causes models to learn that fingers indicate damage (a spurious feature). This issue is fixed with a custom augmentation that removes fingers from all images. Image Credits: LatticeFlow LatticeFlow itself, it is worth noting, isn’t in the training business. The service works with pre-trained models. For now, it also focuses on offering its service as an on-prem tool, though it may offer a fully managed service in the future, too, as it uses the new funding to hire aggressively, both to better service its existing customers and to build out its product portfolio. “The painful truth is that today, most large-scale AI model deployments simply are not functioning reliably in the real world,” said Sunir Kapoor, operating partner at Atlantic Bridge. “This is largely due to the absence of tools that help engineers efficiently resolve critical AI data and model errors. But, this is also why the Atlantic Bridge team so unambiguously reached the decision to invest in LatticeFlow. We believe that the company is poised for tremendous growth, since it is currently the only company that auto-diagnoses and fixes AI data and model defects at scale.”

AI saving whales, steadying gaits and banishing traffic • ZebethMedia

Research in the field of machine learning and AI, now a key technology in practically every industry and company, is far too voluminous for anyone to read it all. This column, Perceptron, aims to collect some of the most relevant recent discoveries and papers — particularly in, but not limited to, artificial intelligence — and explain why they matter. Over the past few weeks, researchers at MIT have detailed their work on a system to track the progression of Parkinson’s patients by continuously monitoring their gait speed. Elsewhere, Whale Safe, a project spearheaded by the Benioff Ocean Science Laboratory and partners, launched buoys equipped with AI-powered sensors in an experiment to prevent ships from striking whales. Other aspects of ecology and academics also saw advances powered by machine learning. The MIT Parkinson’s-tracking effort aims to help clinicians overcome challenges in treating the estimated 10 million people afflicted by the disease globally. Typically, Parkinson’s patients’ motor skills and cognitive functions are evaluated during clinical visits, but these can be skewed by outside factors like tiredness. Add to that fact that commuting to an office is too overwhelming a prospect for many patients, and their situation grows starker. As an alternative, the MIT team proposes an at-home device that gathers data using radio signals reflecting off of a patient’s body as they move around their home. About the size of a Wi-Fi router, the device, which runs all day, uses an algorithm to pick out the signals even when there’s other people moving around the room. In study published in the journal Science Translational Medicine, the MIT researchers showed that their device was able to effectively track Parkinson’s progression and severity across dozens of participants during a pilot study. For instance, they showed that gait speed declined almost twice as fast for people with Parkinson’s compared to those without, and that daily fluctuations in a patient’s walking speed corresponded with how well they were responding to their medication. Moving from healthcare to the plight of whales, the Whale Safe project — whose stated mission is to “utilize best-in-class technology with best-practice conservation strategies to create a solution to reduce risk to whales” — in late September deployed buoys equipped with onboard computers that can record whale sounds using an underwater microphone. An AI system detects the sounds of particular species and relays the results to a researcher, so that the location of the animal — or animals — can be calculated by corroborating the data with water conditions and local records of whale sightings. The whales’ locations are then communicated to nearby ships so they can reroute as necessary. Collisions with ships are a major cause of death for whales — many species of which are endangered. According to research carried out by the nonprofit Friend of the Sea, ship strikes kill more than 20,000 whales every year. That’s destructive to local ecosystems, as whales play a significant role in capturing carbon from the atmosphere. A single great whale can sequester around 33 tons of carbon dioxide on average. Image Credits: Benioff Ocean Science Laboratory Whale Safe currently has buoys deployed in the Santa Barbara Channel near the ports of Los Angeles and Long Beach. In the future, the project aims to install buoys in other American coastal areas including Seattle, Vancouver, and San Diego. Conserving forests is another area where technology is being brought into play. Surveys of forest land from above using lidar are helpful in estimating growth and other metrics, but the data they produce aren’t always easy to read. Point clouds from lidar are just undifferentiated height and distance maps — the forest is one big surface, not a bunch of individual trees. Those tend to have to be tracked by humans on the ground. Purdue researchers have built an algorithm (not quite AI but we’ll allow it this time) that turns a big lump of 3D lidar data into individually segmented trees, allowing not just canopy and growth data to be collected but a good estimate of actual trees. It does this by calculating the most efficient path from a given point to the ground, essentially the reverse of what nutrients would do in a tree. The results are quite accurate (after being checked with an in-person inventory) and could contribute to far better tracking of forests and resources in the future. Self-driving cars are appearing on our streets with more frequency these days, even if they’re still basically just beta tests. As their numbers grow, how should policy makers and civic engineers accommodate them? Carnegie Mellon researchers put together a policy brief that makes a few interesting arguments. Diagram showing how collaborative decision making in which a few cars opt for a longer route actually makes it faster for most. The key difference, they argue, is that autonomous vehicles drive “altruistically,” which is to say they deliberately accommodate other drivers — by, say, always allowing other drivers to merge ahead of them. This type of behavior can be taken advantage of, but at a policy level it should be rewarded, they argue, and AVs should be given access to things like toll roads and HOV and bus lanes, since they won’t use them “selfishly.” They also recommend that planning agencies take a real zoomed-out view when making decisions, involving other transportation types like bikes and scooters and looking at how inter-AV and inter-fleet communication should be required or augmented. You can read the full 23-page report here (PDF). Turning from traffic to translation, Meta this past week announced a new system, Universal Speech Translator, that’s designed to interpret unwritten languages like Hokkien. As an Engadget piece on the system notes, thousands of spoken languages don’t have a written component, posing a problem for most machine learning translation systems, which typically need to convert speech to written words before translating the new language and reverting the text back to speech. To get around the lack of labeled examples of language, Universal Speech Translator converts speech into “acoustic units”

Deep Render believes AI holds the key to more efficient video compression • ZebethMedia

Chri Besenbruch, CEO of Deep Render, sees many problems with the way video compression standards are developed today. He thinks they aren’t advancing quickly enough, bemoans the fact that they’re plagued with legal uncertainty and decries their reliance on specialized hardware for acceleration. “The codec development process is broken,” Besenbruch said in an interview with ZebethMedia ahead of Disrupt, where Deep Render is participating in the Disrupt Battlefield 200. “In the compression industry, there is a significant challenge of finding a new way forward and searching for new innovations.” Seeking a better way, Besenbruch co-founded Deep Render with Arsalan Zafar, whom he met at Imperial College London. At the time, Besenbruch was studying computer science and machine learning. He and Zafar collaborated on a research project involving distributing terabytes of video across a network, during which they say they experienced the shortcomings of compression technology firsthand. The last time ZebethMedia covered Deep Render, the startup had just closed a £1.6 million seed round ($1.81 million) led by Pentech Ventures with participation from Speedinvest. In the roughly two years since then, Deep Render has raised an additional several million dollars from existing investors, bringing its total raised to $5.7 million. “We thought to ourselves, if the internet pipes are difficult to extend, the only thing we can do is make the data that flows through the pipes smaller,” Besenbruch said. “Hence, we decided to fuse machine learning and AI and compression technology to develop a fundamentally new way of compression data getting significantly better image and video compression ratios.” Deep Render isn’t the first to apply AI to video compression. Alphabet’s DeepMind adapted a machine learning algorithm originally developed to play board games to the problem of compressing YouTube videos, leading to a 4% reduction in the amount of data the video-sharing service needs to stream to users. Elsewhere, there’s startup WaveOne, which claims its machine learning-based video codec outperforms all existing standards across popular quality metrics. But Deep Render’s solution is platform-agnostic. To create it, Besenbruch says that the company compiled a dataset of over 10 million video sequences on which they trained algorithms to learn to compress video data efficiently. Deep Render used a combination of on-premise and cloud hardware for the training, with the former comprising over a hundred GPUs. Deep Render claims the resulting compression standard is 5x better than HEVC, a widely used codec and can run in real time on mobile devices with a dedicated AI accelerator chip (e.g., the Apple Neural Engine in modern iPhones). Besenbruch says the company is in talks with three large tech firms — all with market caps over $300 billion — about paid pilots, though he declined to share names. Eddie Anderson, a founding partner at Pentech and board member at Deep Render, shared via email: “Deep Render’s machine learning approach to codecs completely disrupts an established market. Not only is it a software route to market, but their [compression] performance is significantly better than the current state of the art. As bandwidth demands continue to increase, their solution has the potential to drive vastly improved commercial performance for current media owners and distributors.” Deep Render currently employs 20 people. By the end of 2023, Besenbruch expects that number will more than triple to 62.

Subscribe to Zebeth Media Solutions

You may contact us by filling in this form any time you need professional support or have any questions. You can also fill in the form to leave your comments or feedback.

We respect your privacy.
business and solar energy