“Interestingly, I view much of my scientific work as art, and I consider myself an artist.”
I had the pleasure of talking with Dr. Hava Siegelmann. When Hava presented her doctoral research at the University of California, Berkeley, a distinguished computer scientist objected to the conclusion. Her mathematical proofs suggested that certain continuous neural networks could exceed the limits of classical Turing computation, a finding that challenged one of computer science’s most established boundaries. Siegelmann invited him to identify an error. He conceded that he could not.
“I am a soldier in Turing’s army,” he told her.
The remark stayed with Siegelmann because she had never considered her work an attack on Alan Turing. She saw it as an extension of the questions he had opened. Years later, while rereading Turing’s original papers during the 2012 centennial of his birth, she discovered that Turing himself had resisted the rigid interpretation defended in his name. He had explored continuous variables, physical randomness, analog systems, and mechanisms that might better resemble the human brain.
The episode captures a tension that has followed Siegelmann throughout her career. She works in a field built on formal definitions and mathematical limits, yet she is drawn to systems that refuse to remain fixed. Her research examines how brains, machines, and communities learn from new information without erasing what came before. The same interest connects her work in neural computation, lifelong-learning artificial intelligence, neuroscience, music, and even religious practice: How can a system preserve its structure while remaining open to change?
Siegelmann was born in Israel, the youngest of three daughters of German-born parents who had arrived there as children before World War II. Her older sisters were well behaved, she recalls, so her parents felt they still needed “an energetic little boy.” Siegelmann assumed the part. She performed in shows, played piano, and grew up in a religious household where music and art were treated as serious pursuits. Her father held a conventional job but devoted much of his life to helping people and ceramic work.
Her family expected her to study orchestral conducting. She postponed that path to explore computers and found that the subject offered another way to approach the questions that had fascinated her since childhood. She watched the family’s cats learn to navigate obstacle courses she built for them. She wondered how children could progress from one school year to the next without exhausting their capacity for memory. Did new knowledge overwrite the old? Did it pass through some cognitive filter before joining what was already there?
She considered medicine and was accepted to medical school, but computational science gave her a framework for studying memory and learning across machines, biological systems, and human beings. The choice did not feel like a rejection of art. “I view much of my scientific work as art,” she says. Modeling nature, designing algorithms, and searching for new computational structures require the same instinct to recognize form where others may see disorder.
Her academic career took her from Israel’s Technion to the United States, including work at MIT and a faculty position at the University of Massachusetts. Even her hiring story carries the mixture of seriousness and dry humor that runs through her account of her career. She interviewed while pregnant but, following advice about American professional customs, did not mention it. After her son was born and spent his first months in neonatal intensive care, she failed to respond promptly to UMass’s emailed offer. The university apparently concluded that the offer was insufficient and returned with a tenured position.
She jokes that it is “unique negotiation advice,” though not a method she recommends.
DARPA entered her career with similar unpredictability. Siegelmann had never applied for one of the agency’s grants and knew little about the program manager role when a departing manager suggested her as his successor. The Microsystems Technology Office wanted someone who could connect algorithms, biology, neuroscience, hardware, and software. That combination closely matched the territory she had already been exploring.
At DARPA, she translated the theoretical concerns of her early research into practical AI programs. Her Lifelong Learning Machines initiative addressed catastrophic forgetting, the tendency of a neural network to lose earlier capabilities while learning new ones. Rather than treating intelligence as a model trained once and then frozen, the program explored machines that could keep adapting in real time.
The distinction matters beyond engineering. A system that cannot incorporate new information without destroying its past is powerful but brittle. Siegelmann’s work asks whether intelligence is defined less by what a system already knows than by how it changes when the world stops behaving as expected.
That question also appears in her neuroscience research. In one large-scale analysis of brain-imaging studies, Siegelmann and her research lab examined how sensory information develops into increasingly abstract thought. Their results placed naming at the highest level of abstraction. To name something, she explains, is to move beyond its sensory features and identify what appears essential about it.
For Siegelmann, the finding unexpectedly intersected with Jewish thought, where HaShem, “The Name,” is used to refer to God. The Bible portrays naming as something far deeper than labeling. When God brings the animals to Adam “to see what he would call them,” Adam participates in creation by discerning the unique nature of each creature. In that sense, naming becomes a partnership with God: God creates the world, and the human being responds by recognizing, distinguishing, and giving language to creation. This mirrors the way our mind works. We first perceive a forest as a blur. Only when our attention settles on a particular tree can we name it — and only then does it become a distinct object in our awareness. This is one manifestation of the principle behind Super-Turing computation, a computational theory introduced by Hava Siegelmann. Beyond explaining biological intelligence, it provides a blueprint for AI systems that allocate computation selectively where it is most needed, learn continuously, remain resilient, and operate with low energy consumption.
A data-driven investigation had returned her to a concept embedded in her religious life. Science and faith were not being collapsed into one another. They had arrived, by different routes, at a shared question about language, essence, and human understanding.
Her observance of Shabbat provides another boundary in a career devoted to continuous learning. She disconnects while the technological world continues operating, and colleagues across the international research community have learned to plan around her Friday departures. The practice is not a retreat from science. It is a reminder that an adaptive system still needs limits.
Siegelmann’s argument for diversity follows the same logic. Neural networks become more capable when their components operate across different functions and time scales. Human communities, she believes, also require genuine differences in experience and thought. Uniformity may make a system easier to control, but it can also make it dangerously narrow.
Her work has repeatedly challenged the idea that intelligence, whether artificial or human, can be reduced to one model, one authority, or one permanent set of rules. The conversation that follows moves through the questions that have occupied her since childhood: how memory survives change, how difference strengthens a system, and how naming the world may be one of the most sophisticated things the human brain can do.
Yitzi: Hava, it is so nice to meet you. Before we dive in and talk about your amazing work, our readers would love to learn about your personal origin story. Can you share with us the story of your childhood, how you grew up, and particularly the seeds of the genesis of all the amazing creativity that has come since then?
Hava Siegelmann: To start from childhood — I have never tried telling it quite like this, so let us try. I was born in Israel as the third and youngest child to German parents. Both of my parents came to Israel as children just before the war and met here, so my grandparents spoke German.
As a child, I had two older sisters who were very well-behaved, so my parents felt they needed an energetic little boy — and that role fell to me. I was very active, played music, played the piano, and was on stage in shows all the time. My family is religious, so I attended a religious elementary and high school, spending four years at an all-girls high school.
I was about to finish high school early, so my parents advised me to transition to another school. I spent two years at the Reali School in Haifa, a prestigious private science school in the city where I grew up. Afterwards came my army service, the Technion, and computer science.
All throughout this time, I was deeply involved in music and played piano. Originally, the expectation within my family was that I would study orchestral conducting. I postponed that path slightly just to explore computers, and I fell in love with the connection between the human brain and computer systems. From a young age, I was fascinated by how memory and learning operate. I was curious about people as well as animals; my family always had cats, and I used to set up little “Olympics” obstacle courses for each new kitten. I was captivated by how they learned, trained, and remembered.
I considered medical school, took the medical entrance exams, and was accepted, but I postponed entry. I realized I preferred using computer science to explore these questions within a computational framework, ultimately dedicating my career to memory and learning across computers, biological systems, and humans.
The core question I pondered early on was memory capacity: How much memory can we actually retain? How can we complete second grade, move into third grade, and still have enough capacity left? Does new information clash with old memories? When we gain a new perspective to understand the world, how does that affect everything we previously knew? Does it integrate directly into memory, or must memories pass through cognitive filters before coming together? These questions were deeply fascinating to me.
Music remains an essential part of my life. Interestingly, I view much of my scientific work as art, and I consider myself an artist. While society labels composing and playing piano as art, I view my work in modeling nature, building computational frameworks, studying learning, and designing new algorithms in a similar artistic light. Behind me, you can see ceramic artwork; my father was an artist who held a standard job while dedicating his life to compassionate mitzvot and creation.
Yitzi: Please tell us the next chapter: the story of how you became a program manager and director at DARPA.
Hava Siegelmann: It is an amusing story. I started my academic career in Israel at the Technion in Haifa. I then went on sabbatical to the United States, stayed, and raised my family here. Following a period at MIT, I secured a faculty position at the University of Massachusetts.
I have a humorous story about that hiring process. When I interviewed, I was pregnant and was advised that in American professional settings one does not discuss pregnancy during interviews, so I did not say a word. I wondered if they simply assumed I was gaining weight. After giving birth, my son experienced a difficult medical start — Baruch Hashem for how thriving he is today, but he spent his first few months in the NICU.
During that period, UMass sent me a job offer via email, but I was not checking my inbox. Because I did not respond immediately, the university assumed their initial offer was too low and revised it to a tenured offer! I often joke that I have unique negotiation advice for people, though it does not require spending months in the NICU.
My path to DARPA was equally unexpected. I had no prior grants with DARPA. A departing Program Manager felt I would be the ideal candidate to take over his portfolio. When DARPA initially reached out, I knew very little about the agency or what a PM role entailed.
The division that contacted me was MTO (Microsystems Technology Office), which focuses on hardware. They wanted to combine algorithm design, biology, and neuroscience to bring brain-inspired concepts into integrated hardware and software architectures — a perfect match for my research. When I did not immediately accept their invitation to visit, DARPA representatives traveled to UMass to meet with me, my department chairman, and the dean.
I am very glad I accepted. It was a wonderful fit and a privilege to serve the research community — extending across the United States and globally into the fields of AI and neuro-AI.
Yitzi: Can you tell us about a few products or initiatives that grew out of your work at DARPA?
Hava Siegelmann: Certainly. My first program aimed to translate theoretical research — specifically Super-Turing computational theories — into practical AI applications.
To understand Super-Turing models, it helps to review standard computing history. In 1936, Alan Turing formulated his foundational computing model by observing how human “calculators” — clerks sitting at desks performing arithmetic manually — processed information. Turing modeled this process using a central processing unit and an unbounded memory source (like a notebook where pages can be written, erased, and replaced). A program loaded into the processing unit dictated how incoming data were processed. For example, if the program specified how to multiply two vectors, the processor would apply those instructions whenever two vectors were provided as input. This concept birthed the universal computer, establishing the modern architecture of separate processing and memory, with computation directed by a stored program.
During my PhD in the early 1990s, I began examining neural networks. Unlike traditional computers, neural networks do not separate memory from processing; computation occurs directly through interconnected artificial neurons where synaptic weights store memory. Computation and memory are unified.
At the time, the prevailing view in computer science was that the computational power of neural networks was limited to that of finite automata — models far weaker than Turing machines. I set out to prove this accepted claim, but quickly discovered that neural networks could already simulate Turing machines. Since the Turing machine was considered the strongest realistic model of computation, I then sought to prove that the two models were computationally equivalent. I quickly proved that a neural network could simulate a Turing machine, but every attempt to prove the reverse — that a Turing machine could simulate any neural network — broke down. To uncover the missing insight, I tried proving the opposite: that neural networks were more powerful than Turing machines, expecting the proof to fail in a way that would reveal how to complete the equivalence proof. Instead, it went through cleanly. The result was striking: neural networks are Turing-equivalent only when they are deterministic, non-learning, and have discrete state spaces. Once they exhibit even a single natural feature — such as continuous state, probabilistic behavior, asynchronous operation, or ongoing plasticity: recurrent neural network architectures strictly transcend classical Turing computability, establishing a framework we named Super-Turing computation.
This revealed a broader hierarchy of computational classes based on how a system dynamically incorporates external real-time information to adapt its processing.
This framework suggested a new paradigm for artificial intelligence. Rather than training a neural network on a static dataset and permanently freezing its parameters, a system can continuously draw targeted information from its environment as needed.
Consider driving to work on a routine route: you expend minimal cognitive effort and attention. However, if your car skids on ice, you instantly recruit maximum focus and energy. Traditional computers consume uniform energy and clock cycles regardless of task complexity. Super-Turing principles allow a system to operate at high efficiency with low baseline energy while dynamically scaling resources for complex, novel events. But Super-Turing is far more than an energy-saving principle. In the Turing paradigm, changing the input changes only the output while the program itself remains fixed. In the Super-Turing paradigm, an input can also modify the program itself. Future outputs are then generated by the updated program, allowing the system to adapt rather than merely react. Humans work this way: our experiences do not simply produce immediate responses — they also change how we think and behave in the future. In this sense, part of the input acts as advice or context that reshapes the program itself. Years later, while studying memory reconsolidation in neuroscience — how stored memories become malleable upon recall before being re-stored — I realized human memory operates directly on these Super-Turing dynamics.
When I arrived at DARPA, I established the Lifelong Learning Machines (L2M) program to bring these principles into mainstream AI. Traditional deep learning models suffer from “catastrophic forgetting”: if you attempt to train a pre-trained model on new data, the new weight adjustments overwrite existing knowledge across the network. L2M developed new algorithmic paradigms allowing AI systems to learn continuously in real-time without erasing past capabilities. This initiative sparked widespread research across the field and laid groundwork for modern adaptive AI architectures.
I also created a biomedical application of lifelong learning focused on closed-loop physiological monitoring. In severe trauma, blood sugar volatility can be fatal. Standard medical pumps were open-loop systems: they could administer insulin for high blood sugar, but could not automatically administer glucose if levels crashed. We funded closed-loop systems capable of co-administering insulin and dextrose simultaneously to keep trauma patients stable.
Additionally, I launched the GARD program (Guaranteeing AI Robustness against Deception). Just as humans can be manipulated through cognitive biases, AI models can be tricked by exploiting their pattern-recognition architecture — such as placing small, targeted stickers on a military vehicle to trick a vision algorithm into classifying it as an ambulance. GARD focused on identifying these structural vulnerabilities and engineering robust defenses.
Returning to deception: tricking an AI system usually involves exploiting its greatest strength. Criminals trick honest people by exploiting their trust. Similarly, AI excels at pattern recognition, so adversaries exploit that feature. Because humans and AI possess fundamentally different cognitive structures, combining human oversight with AI systems creates the strongest safeguard against deception on both sides.
Yitzi: You have had a remarkable career. Can you share two stories that stand out to you from your professional journey?
Hava Siegelmann: I have never viewed myself through the lens of being “successful” — it is simply how I approach my work. However, I can share an early story and a later story from my career.
When I was a young researcher presenting my PhD work on Super-Turing computation at UC Berkeley, I presented mathematical proofs showing how continuous-domain neural networks extend beyond classical Turing limits. During the Q&A, a highly distinguished computer scientist became visibly upset. I invited him to point out any mathematical flaws in the proof so we could evaluate them together. He acknowledged there were no mathematical errors, but stated, “I am a soldier in Turing’s army.”
That reaction startled me. I respected the immense value of Turing’s work while demonstrating that additional computational paradigms exist in biological systems.
Years later, in 2012, during the centennial celebration of Alan Turing’s birth, I spent time re-reading Turing’s original papers from the 1930s and 1940s. I was astonished to discover that Turing himself held the opposite view of a rigid “soldier.”
In his writings, Turing explicitly noted that what we call the “Turing machine” was designed for mechanical, repetitive arithmetic — not as the ultimate boundary for human-like intelligence. He actively sought advanced computational mechanisms, exploring continuous variables, true physical randomness (even suggesting introducing radioactive decay into hardware), and hybrid analog-digital processing to emulate the human brain.
It was deeply moving to realize that Turing shared those exact foundational questions about brain-like computation decades earlier.
To tell you a story — and I do not know if you will share this with everyone — I observe Shabbat. When I am invited to events, I usually explain that I cannot attend on Friday or that I must leave early. For instance, the International Neural Network Society (INNS), where I am active, always holds organizing meetings on Friday and Saturday, with pre-conference activities starting Sunday. I would explain that I could only arrive later.
After doing this for years, without even having to explicitly bring it up anymore, whenever I receive conference invitations — whether from China, Japan, or anywhere else — people write, “We understand the conference runs Monday through Saturday, but that you will be leaving before Friday.” Somehow word spread across the global community, and they accommodate it automatically. It is quite funny.
Yitzi: That is so special. Part of it may be a growing awareness of how ingenious Shabbat is on physical, emotional, societal, and intellectual levels. Today especially, when technology allows us to work 24 hours a day, 7 days a week nonstop — something that used to be impossible — there is a real appreciation for that boundary.
Hava Siegelmann: A colleague at MIT told me she had been reading about this and adopted a new practice: taking one 24- to 25-hour period every week where she disconnects completely from emails and work. I smiled, knowing I have been practicing something similar for years, but it is certainly a good idea!
I have another research story to share involving a postdoc of mine. Coming from neuroscience, we were exploring a fundamental question: our senses receive very tangible, raw inputs — bits of light, auditory signals — so how does the brain assemble these into abstract concepts like democracy, fairness, or honesty? We are not born with direct sensory inputs for these abstract ideas, yet even young children quickly grasp notions of fairness, bias, and equity without explicit instruction.
To investigate this, we constructed a model of human brain architecture incorporating different sensory inputs — including visual cortex, auditory cortex, gustatory cortex, limbic regions, and body signals. Using resting-state fMRI data, we set neuronal distances based on how long an input takes to propagate through the brain. We combined structural pathways with temporal processing speeds.
Next, we analyzed every fMRI study available across major databases, evaluating which brain regions were primarily recruited for various tasks. By aggregating massive amounts of data, any individual study noise washes out, revealing the broader organizational pattern.
The study revealed a phenomenon we termed a “slope hierarchy.” Every cognitive task — even simple ones — engages multiple brain depths, but with different weighting (“slopes”) or engagement. Remarkably, when we ranked tasks by their slope — from the most negative to the most positive — they also fell naturally along a continuum from the most tangible to the most abstract. Motor actions, such as snapping your fingers, had negative slopes and occupied the tangible baseline, whereas delayed-memory tasks had positive slopes and lay toward the abstract end.
The single highest, most abstract human cognitive task across all data was naming.
When you consider it, naming is profound. We name our children, and according to the Bible, human history begins when God asks Adam to name the animals. Perhaps naming is itself a way of communicating with God — not because God lacks names for His creations, but because naming is humanity’s response to creation. It is as though creation does not fully become part of the human world until it is recognized and named. This mirrors the way our minds work. We first perceive a forest as a blur of colors and textures. Only when our attention settles on a particular tree, branch, or leaf can we name it — and only then does it become a distinct entity in our awareness. Naming transforms perception into understanding. This is the essence of Super-Turing computation. Intelligence is not about processing everything equally, but about deciding what deserves attention. We devote only as much computation to a scene as it requires, selectively focusing our resources rather than analyzing every detail. Naming, in this sense, is the cognitive expression of selective attention. Only what enters our focus becomes fully represented. If everything were always in full focus, we would be overwhelmed rather than intelligent. This is how Super-Turing computation works. By allocating computation only where it is needed, it mirrors the way biological intelligence perceives and learns. That is why Super-Turing computation is more than a theory of biological intelligence. It provides the foundation for a new generation of AI systems that are more capable, more resilient, more energy efficient, and ultimately more responsible.
Naming requires understanding the core essence of something perceived and distilling it into a single word. It ranked above memory and other complex cognitive tasks as the highest, most distinctly human capacity.
Yitzi: Isn’t it interesting that the phrase we often use for God is HaShem — literally, “The Name”?
Hava Siegelmann: You are right! It is above everything. I had not made that connection before.
Yitzi: Ultimately, that aligns with our purpose: to create a Kiddush Hashem, to sanctify and spread That Name. Even the Hebrew word for heaven, Shamayim, contains the root for name — shem — doubled as a plural.
Hava Siegelmann: That is fascinating — a double name! I had always thought of it in terms of sham mayim (“there is water”), but a double name is a wonderful perspective. It is remarkable how a purely data-driven, neuroscience investigation involving massive fMRI datasets leads right back to the concept of naming.
Yitzi: What is also interesting is that the root of “etymology” — the origin of words and names — comes from the Greek etymos, meaning truth, which mirrors the Hebrew word emet. The discipline of giving things names was literally viewed as the search for truth.
Hava Siegelmann: Exactly. When we name something, we seek an exact term that captures its essence.
Yitzi: Right, exactly — the essential truth of the matter.
Hava Siegelmann: It sits deep in the cognitive architecture, beyond all sensory input cortices. That is the height of human cognitive reach.
Yitzi: Dr. Siegelman, because of your remarkable work and the platform you have built, you are a person of significant influence. If you could spread an idea or inspire a movement that would bring the greatest good to the most people, what would that be? You never know what an idea might inspire.
Hava Siegelmann: Let me reflect on that for a moment. At the core of my research is the concept of diversity — not superficial categorization, but functional diversity of components. The most effective neural network architectures incorporate diverse elements, such as varying temporal time scales. The same principle applies to human communities: generating great ideas requires listening to diverse perspectives and thoughts.
Progress in science and society relies on natural diversity rather than forced quotas. The most dangerous path would be relying on a single large language model trained within a narrow geographic or political echo chamber, forcing uniform thinking. That is anti-humanity.
My core message would be to respect natural diversity in computational models, AI, and human society — without labels, grading, or rigid dogmatism. The opposite of dogmatic conformity is art, science, and the open-minded embrace of human differences.
Yitzi: That is so beautiful. In English, “peace” denotes the absence of hostility, but in Hebrew, shalom signifies wholeness, completion, and harmony — a positive state where all diverse pieces come together into a complete whole (shalem).
Hava Siegelmann: Exactly — having all the pieces unite. True shalom begins internally: making peace with yourself, your family, your group, and your neighbors. It is an active state of growth rather than a passive cessation of conflict.
Yitzi: Dr. Siegelman, thank you so much for your time and for sharing these extraordinary insights. I would love to stay in touch and continue learning from you. I feel honored to have met you. I used to live in Sharon, Massachusetts.
Hava Siegelmann: I have learned from you as well. Thank you.
Beyond Turing’s Army: How DARPA Trailblazer Dr. Hava Siegelmann Is Rewriting the Limits of AI was originally published in Authority Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.