Update #74: Detecting Postpartum Depression and Kolmogorov-Arnold Networks
We look at Dionysus Digital Health's new ML system for detecting postpartum depression in expectant or new mothers; Kolmogorov-Arnold Networks are getting a lot of hype.
Welcome to the 74th update from the Gradient! If you’re new and like what you see, subscribe and follow us on Twitter. Our newsletters run long, so you’ll need to view this post on Substack to see everything!
Editor Notes
Happy Tuesday. In possibly interesting news, Sam Altman recently gave a talk at Stanford, and students were lining up in droves. I didn’t drive up to Stanford to see the line or the talk, but did some “investigative journalism” by questioning a representative sample of the Stanford population to find out more about what was going on (I think there are now recordings of this talk). Here’s what I learned:
Of course, there was more.
Also notably, both our news and research highlights this week are about things to be skeptical of: diagnosing postpartum depression is pretty hard, as it turns out, and the jury is still very much out on KAN (see the first author’s own comments, which I also linked below).
I also really enjoyed doing the last two podcast episodes—Ryan Tibshirani is incredibly thoughtful, and I love that he sometimes pursues problems for the sake of their beauty; I don’t know how he strikes the balance he does in his research, but it’s admirable. David Thorstad’s bounded rationality program is important, I think that his arguments against longtermism come from a principled place, and I also respect that the longtermism community funds and responds to his critical work. Whatever you think about longtermism, it’s hard to deny that the community pushes for the changes they think best.
Finally,
, a brilliant ML educator and someone the Substack AI community is lucky to have, recently published his new book, Machine Learning Q and AI. If you’re looking for a straightforward read as a refresher on concepts, it’s nicely written.As usual, if you want to write with us, send a pitch using this form.
Can't afford to test in production? Come to The world’s first AI Quality conference
On June 25th in San Francisco, join the first ever AI Qualify conference. Join 1000+ other highly informed attendees to connect and learn about how you can keep grom letting your AI efforts go awry.
Hear speakers from Uber, Groq, Cruise, Torc Robotics, Notion, Anthropic, Open AI, Google and 20+ more organizations.
Use the special discount code 'hallucinate' for 30% off the ticket price.
News Highlight: Dionysus claims its blood test can detect postpartum depression. How much of a difference can this make?
Summary
San Diego-based startup Dionysus Digital Health is pitching a blood test to pitch for postpartum depression, even before symptoms appear. Such a system—if it worked in the way Dionysus imagines—could help healthcare systems funnel vulnerable mothers toward treatment, and even preventative care.
Overview
Postpartum depression, or PPD, is unfortunately common for women who have recently given birth to children—the CDC says about one in 8 new mothers experience symptoms (the NCBI says one in 7). Dionysus claims it has pinpointed a gene linking moods to hormonal changes—their ML system can compare gene expression in blood samples to determine if a mother might develop depression symptoms.
This is the picture Dionysus imagines: providers administer a blood test between the second and third trimesters of pregnancy; that blood test flags women at risk of postpartum depression or other disorders; this, and other methods, funnel vulnerable women towards appropriate treatment and care.
The reality of current medical care is that women are not always screened for postpartum depression during and after pregnancy, despite recommendations. Tools like Dionysus’s could make it easier to identify at-risk mothers.
Of course, such a tool doesn’t exist in a vacuum, and the usual cast of concerns has been raised: bias and cost are two considerations; identifying at-risk mothers alone also makes little difference if those mothers can’t access care.
Our Take
I think this is an interesting story because it highlights the ever-present lesson that AI systems don’t exist in a vacuum, and there is a large gap between solving a problem with an ML system—an important problem, admittedly—and leveraging that system to make a real-world impact.
What I find particularly interesting is that Dionysus says they want to sell their postpartum depression test directly to consumers. And, there’s more than postpartum depression screening: they want to “create bite sized feedback unique to you” and provide actionable tools based on their epigenetic insights. The idea of personalized wellness support isn’t an entirely new one, and interacts with some of the bias concerns that have been raised.
But, all the signs (and Dionysus’s own messaging) point to just the kind of consumer product that makes people worry about inequity: if consumers need to pay out of pocket for Dionysus’s technology and it isn’t broadly affordable, there’s a clear exacerbating factor.
Even then, suppose Dionysus’s system works and most people are able to use it—there are a number of non-genetic factors that can also predispose expectant or new mothers to postpartum depression. These are factors that a blood test alone can’t take into account, and would require more extensive familiarity with a patient’s situation (or something like a survey, in which case the system’s ease of detection when compared against older screening methods might not be such a big improvement).
Furthermore, even women who are diagnosed with postpartum depression rarely receive the care they need—this study says just one-third of pregnant mothers with signs of mental disorders received treatment (this treatment included “reassurance” from providers). All this is to say: I think Dionysus’s vision is a nice one and one worth building towards. But we shouldn’t forget that building ML systems alone isn’t going to bring that world about. We need to build better healthcare systems and figure out ways to offer people higher-quality, more consistent care. Again, and again, and again, let’s not make the mistake of techno-solutionism.
—Daniel
I am having war flashbacks to Theranos and am wondering aloud have we learned nothing? While helping those who suffer from postpartum depression is an extremely noble aim, do we have any reliable indication besides their word that Dionysus is capable of reliably detecting postpartum depression? Some clinical research has found that expert doctors can only diagnose depression with a recall of 50%; meaning half of all patients feeling depressed would be misdiagnosed by experts. This suggests that even if Dionysus has a magical AI that can predict their training labels well, the underlying process for label generation (doctors diagnosing people with depression) is incredibly flawed and would on average be mislabeling approximately half of the depressed population as not depressed.
-Justin
Research Highlight: KAN: Kolmogorov-Arnold Networks
Summary
An exciting new paper by researchers at MIT, Caltech, and Northeastern presents Kolmogorov-Arnold Networks (KANs) as an alternative to the traditional Multi-Layer Perceptron (MLP) neural network architecture. The key differentiator between KANs and MLPs come from KANs having learnable activation functions on the edges between nodes, while MLPs have predetermined activation functions on the nodes themselves. The authors achieve this by replacing all linear weights with a univariate function parameterized as a B-spline (a gorgeous explainer on B-splines can be read here). The researchers go on to empirically test KANs on numerous data fitting and PDE (Partial Differential Equations) problems finding that KANs have significant gains in accuracy and interpretability over MLPs. The researchers go on to empirically show KANs can learn and represent numerous laws of nature and physical constraints while also exploring various drawbacks like slow training times.
Overview
While MLPs take their theoretical inspiration from the Universal Approximation Theorem, KANs rest on the theoretical grounds of Kolmogorv-Arnold (KA) representation theorem. This theorem posits that any multivariate continuous function can be expressed as the finite composition of a single variate continuous function. It follows from this theorem that learning a high-dimensional function (an extremely common and crucial task in machine learning) can boil down to learning a polynomial number of one-dimensional functions.
The authors begin by using the KA representation theorem to create a single KAN layer abstraction, and then extend it by introducing a depth-wise expansion. They do so via extending the single composition of one-dimensional functions into a matrix of learnable 1D functions. By introducing the depth wise expansion, they are both strengthening the conceptual relationship to MLPs and improving the accuracy of the functional representation learned by KANs..
The authors propose training a very large and deep representation with introduced sparsity regularization. The authors show that one can prune away all of the unnecessary nodes resulting in better performance and easier interpretability. The final innovation which assists in interpretability is what the authors call Symbolification. The authors suggest that some activations are symbolic functions (like log or sin). They show that these symbolic functions can be determined through learnable affine transformations from the pre-activations to the post-activations. Through node pruning and Symboliification, they empirically demonstrated learning symbolic representations for numerous spline activations and configurations.
These concepts can be best demonstrated in this beautiful animation from the authors, which can be seen below. As the number of learning iterations increases, we can see both the pruning and the symbolic expression in action. By Step 50, we can see the topmost node has learned to represent the exponential function, while the sin function is learned by both nodes in the middle layer. Finally, in the bottom layer, we see that for all four input variables, the splines learn a squared representation while the extraneous inputs are pruned away..
Our Take
One of the most exciting things to me about this work is the speed in which the machine learning community seems to be adopting , contributing , and expanding the domain of knowledge around KANs. Some notable community highlights include implementations trying to extend the methodology to different deep learning frameworks (Pytorch) and other domains (reinforcement learning). Some of those can be seen here and here. Additionally the physicist in me loves to see both the symbolic representation regarding interpretability and that the learned representations seem to be respecting the laws of nature. I am eagerly awaiting to see what comes next regarding innovations and improvements as well as how scientists can adopt KANs for new and exciting use cases as highlighted by some of the authors from the website formerly known as Twitter here and here.
-Justin
I like this paper and am excited at what seems to be a new idea, but the jury’s far from out yet on KANs. See this CompactNet repo: the KAN paper’s comparisons to the noted DeepMind paper had some issues, and the authors found they could match KAN’s accuracy on the mathematics dataset using networks with as few as 122 parameters. This makes me a tad skeptical of the other KAN/MLP comparisons; did the authors train the MLPs as well as they could have? Once again: baselines are important! I also want to mention this thread and notebook, which shows you can rewrite a KAN into an ordinary MLP. It’s worth noting the argument starts by simplifying its focus to piecewise linear functions; it’s not entirely obvious to me if the argument extends in general, so I won’t comment on the veracity of the claim, but it’s worth looking at.
—Daniel
New from the Gradient
David Thorstad: Bounded Rationality and the Case Against Longtermism
Ryan Tibshirani: Statistics, Nonparametric Regression, Conformal Prediction
Other Things That Caught Our Eyes
News
Generative A.I. Arrives in the Gene Editing World of CRISPR
A new AI technology developed by Profluent, a startup based in Berkeley, California, is generating blueprints for microscopic biological mechanisms that can edit DNA—Profluent's AI technology analyzes large amounts of biological data to create new gene editors. This advancement in AI-driven gene editing could lead to more precise and faster methods for battling illnesses and diseases. The research paper describing this technology will be presented at the annual meeting of the American Society of Gene and Cell Therapy.
Meta Says It Plans to Spend Billions More on A.I.
Meta, the parent company of Facebook, Instagram, WhatsApp, and Messenger, reported strong revenue and profits for the first quarter of the year. Revenue was $36.5 billion, up 27% from the previous year, and profit was $12.4 billion, more than double the previous year. However, Meta plans to increase its spending on AI efforts, with a projected spending forecast of $35 billion to $40 billion for the year. This increase is driven by investments in AI infrastructure, including data centers, chip designs, and research and development. Despite the positive financial results, Meta's revenue forecast for the current quarter is lower than analysts' expectations.
Ex-athletic director accused of framing principal with AI arrested at airport with gun
A former athletic director at Pikesville High School in Baltimore County has been arrested and charged with crimes related to the use of AI to impersonate the school's principal and spread racist and antisemitic comments. Dazhon Darien, 31, was apprehended at the airport with a gun as he attempted to board a flight. Investigators determined that Darien faked the principal's voice using AI and circulated the audio on social media, causing significant disruptions at the school. The audio clip led to the temporary removal of the principal and triggered a wave of hate-filled messages. Darien is also charged with theft and retaliating against a witness.
Lethal AI Weapons Are on the Rise. What’s Next?
The development of lethal autonomous weapons (LAWs), including AI-equipped drones, is on the rise. These weapons have the ability to find and kill targets without human intervention. The United Nations has taken a step towards addressing this issue by adding LAWs to the agenda of the UN General Assembly meeting in September. However, progress has been slow due to a lack of consensus on what constitutes an autonomous weapon. AI weapons offer advantages such as increased accuracy and the ability to operate in environments with electronic jamming. However, there are concerns about the potential for catastrophic mistakes and the ethical implications of delegating life-and-death decisions to machines.
AI Can Tell Your Political Affiliation Just by Looking at Your Face, Researchers Find
A recent study conducted by researchers at Stanford University suggests that facial recognition technology combined with artificial intelligence can accurately determine a person's political affiliation by analyzing their facial features. The study involved 591 participants who completed a questionnaire about their political beliefs and were then scanned by an AI algorithm. The algorithm was able to predict political orientation with a high degree of accuracy, even when participants' identities were anonymized. The researchers found that liberals and conservatives have distinct facial morphologies, with liberals having smaller faces. The study highlights the potential implications of biometric surveillance technologies and the need for caution in targeting political messaging online.
8 major newspapers join legal backlash against OpenAI, Microsoft
Eight major daily newspapers, including the Chicago Tribune and the New York Daily News, have filed a lawsuit against OpenAI and Microsoft. The lawsuit claims that the companies used copyrighted work from the newspapers to train their artificial intelligence algorithms without compensating the content owners. The newspapers involved in the lawsuit are owned by Alden Global Capital, an investment fund based in New York City. The lawsuit specifically mentions OpenAI's ChatGPT as one of the AI tools that allegedly used the news articles for training.
MTA banned from using facial recognition to enforce fare evasion
The new state budget in New York includes a ban on the use of facial recognition technology by the Metropolitan Transportation Authority (MTA) to enforce fare evasion rules. The law prohibits the MTA from using biometric identifying technology, including facial recognition, to enforce fare payment. The measure was added to protect New Yorkers' privacy and prevent the potential invasion of people's lives through expanded surveillance. Privacy advocates and good government groups have praised the ban, particularly as the state Legislature increased the maximum penalty for fare evasion. However, some civil rights groups argue that the ban does not go far enough and are calling for legislation to fully outlaw the use of facial recognition by government agencies. Facial recognition technology has been criticized for its imperfections and potential for biased results. The ban is seen as a signal of growing skepticism among New York legislators regarding law enforcement's use of facial recognition technology.
Google urges US to update immigration rules to attract more AI talent
Google is urging the US government to update its immigration rules to attract more AI and tech talent. The company believes that current policies, such as Schedule A, which lists occupations with a shortage of American workers, need to be more flexible and regularly updated to meet the demand in AI and cybersecurity. Google argues that the US risks losing out on highly sought-after talent if immigration policies are not modernized. The company suggests including AI and cybersecurity on Schedule A and considering multiple data sources to reflect workforce gaps. The US's strict immigration policies have made it difficult for companies to attract AI specialists, resulting in a shortage of talent in the country.
China unveils Sora challenger able to produce videos from text
China has unveiled its own text-to-video AI tool called Vidu, which is similar to OpenAI's Sora. Developed by start-up Shengshu Technology in collaboration with Tsinghua University, Vidu can produce 1080p resolution videos based on simple text prompts, although the videos are limited to 16 seconds compared to Sora's 60 seconds. Vidu is described as "imaginative" and able to simulate the physical world, producing videos with consistent characters, scenes, and timelines. The lack of sufficient computing power has been a hindrance to Chinese firms in developing similar AI models, as Sora requires eight Nvidia A100 GPUs to run for over three hours to produce a one-minute clip. Vidu's debut has raised hopes in China as the country aims to catch up with global generative AI players.
NIST launches a new platform to assess generative AI
The National Institute of Standards and Technology (NIST) has launched a new program called NIST GenAI to assess generative AI technologies, including text- and image-generating AI. The program aims to release benchmarks, develop content authenticity detection systems, and encourage the creation of software to identify the source of fake or misleading AI-generated information. NIST GenAI's first project is a pilot study to differentiate between human-created and AI-generated media, starting with text. Teams from academia, industry, and research labs are invited to submit AI systems to generate content or identify AI-generated content. The launch of NIST GenAI is in response to President Joe Biden's executive order on AI and aims to address the growing concern of AI-generated misinformation and deepfakes. The program will inform the work of NIST's AI Safety Institute.
In 2019, Microsoft's CTO, Kevin Scott, expressed concern about Google's advancements in artificial intelligence (AI) in an email to CEO Satya Nadella and Bill Gates. Scott specifically mentioned Google's AI-powered "auto-complete in Gmail" as being "scarily good." He also acknowledged that Microsoft was behind in terms of machine learning (ML) scale. These emails were made public as part of the Department of Justice's antitrust case against Google. In response, Nadella highlighted the need for Microsoft to invest in AI, which led to their partnership with OpenAI. Microsoft's timely investment allowed them to incorporate OpenAI's technology into products like Bing and Microsoft 365, potentially surpassing Google in the AI race.
Ads on Facebook, Instagram for explicit 'AI girlfriends' prompt Meta crackdown
An investigation by Wired revealed that there were over 29,000 explicit advertisements for "AI girlfriend" apps on Meta's platforms, including Facebook, Instagram, and Messenger. These ads featured sexually explicit messaging and AI-generated images of scantily-clad women. More than half of the ads included the acronym "NSFW," indicating that they were not safe for work. Meta, the parent company of these platforms, has policies in place that prohibit adult content, and a spokesperson stated that they are working to remove the violating ads. Meta acknowledges that they are constantly evaluating and updating their approach to address new tactics used by individuals or groups to evade detection.
Friends From the Old Neighborhood Turn Rivals in Big Tech’s A.I. Race
Mustafa Suleyman and Demis Hassabis, childhood friends from London, have become powerful executives in the tech industry's race to build AI. Suleyman is the chief executive of Microsoft AI, while Hassabis is the chief executive of Google DeepMind. In 2010, they co-founded DeepMind, an AI research lab aimed at preventing the profit-driven race to build and deploy AI that they are now both involved in.
Mysterious “gpt2-chatbot” AI model appears suddenly, confuses experts
A mysterious chatbot named "gpt2-chatbot" has appeared in the LMSYS Chatbot Arena, sparking speculation that it may be a secret test version of OpenAI's upcoming GPT-4.5 or GPT-5 language model. The new model is currently only available through the Chatbot Arena website, with a limited rate of eight queries per day. Despite rumors and hype surrounding the model's capabilities, some users have found that it does not represent a significant leap beyond GPT-4 Turbo. While the origins of the model remain unknown, AI researcher Simon Willison believes it may be an OpenAI stealth preview.
Papers
Daniel: Our own
—who is now at Scale AI—led an incredible paper trying to understand whether data contamination might be going on in notable LLMs. Thy evaluated LLMs on a new test set of the grade school arithmetic benchmark GSM8K, called GSM1K, and observed interesting accuracy discrepancies that pointed to systematic overfitting in models like Phi and Mistral. On a similar theme, I like this paper that presents a new metric for assessing memorization in LLMS: the Adversarial Compression Ratio (ACR). A string from the training data is considered “memorized” if it can be elicited with a prompt shorter than the original string.On another topics, this paper proposes an evaluation framework for subword tokenization. This is pretty important, given that tokenizer evals and comparison are still open problems. The authors’ UniMorph Labeller classifies a subword tokenization as either “morphological” (subword tokenizations that “respect morpheme boundaries,” or: are not present in the vocabulary and aren’t alien) or “alien” (linguistically implausible subword compositions, e.g. those that do not align with semantic compositions that humans understand). They also propose the Out-of-Vocabulary Generalization challenge for evaluating sub-word tokenization in downstream NLP tasks—they find alien compositions lead to poor tokenization compared to morphological ones.
You should also check out this Bayesian learning model for LLMs—I think it’s a nice framework to think about LLMs, and, if you’ve been following work on in-context learning, the paper notes implications for ICL as well. Maybe it’s notable the a Bayesian picture, at least for ICL, still seems to explain the evidence we have now.
Finally, I think Transformers Can Represent n-gram Language Models is a really neat paper. The authors argue that language acceptance is an ill-suited problem for studying LMs, and that transformer LMs using hard or sparse attention can exactly represent any n-gram LM (with stronger results for hard attention). The concrete lower bound on LMs’ representational capacity is useful, and helps contextualize and justify explanations and intuitions we might have about what’s causing LMs to behave as they do.
Closing Thoughts
Have something to say about this edition’s topics? Shoot us an email at editor@thegradient.pub and we will consider sharing the most interesting thoughts from readers to share in the next newsletter! For feedback, you can also reach Daniel directly at dbashir@hmc.edu or on Twitter. If you enjoyed this newsletter, consider donating to The Gradient via a Substack subscription, which helps keep this grad-student / volunteer-run project afloat. Thanks for reading the latest Update from the Gradient!