Unifying the Principles of Learning with and without Brains

The field of collective intelligence studies how teams can achieve better results than any of the team members alone. Inspired by the history of "freestyle" chess tournaments, in which human-machine teams known as "centaurs" outperformed the best humans and best machines alone, we set out to study the relative advantages of humans and machines in chess. I will show that, in principle, there is high potential for synergy between humans and machines in a complex sequential decision environment such as chess. However, a chess expert can identify only a small part of these relative advantages. By contrast, we could train a network using reinforcement learning to recognize some of these relative advantages without chess expertise. Our findings contribute to the study of collective intelligence and human-centric AI.

Designing machine learning agents that can accumulate knowledge and refine skills over an extended learning lifetime remains a central challenge in artificial intelligence. This talk explores recent theoretical advances in continual learning—the study of agents that learn from a continuous stream of tasks while balancing memory retention, adaptability, and model expressiveness. We focus on the fundamental problem of forgetting, where acquiring new knowledge can impair performance on previously learned tasks, and present data-dependent and oracle PAC-Bayesian bounds that quantify and, in some cases, predict the extent of forgetting. Emphasizing exemplar-free methods, we examine how forgetting unfolds in environments characterized by various forms of task shift—gradual, abrupt, or highly similar—and how this behavior reflects deeper tensions between plasticity and stability. We illuminate the complex relationship between generalization and forgetting, showing that trade-offs are often necessary and may depend sensitively on task structure and computational constraints.

In 1972 Phil Andersen articulated the motto of condensed matter physics as “More is different.” However, for most many-body systems the behavior of a trillion bodies is nearly the same as that of a thousand. Here I argue for a class of condensed matter, “tunable matter," in which many more is different. The ultimate example of tunable matter is the brain, whose cognitive capabilities increase as size increases from 302 neurons (C. Elegans) to a million neurons (honeybees) to 100 billion neurons (humans). I propose that tunable matter provides a unifying conceptual framework for understanding not only a wide range of systems that perform biological functions, but also physical systems capable of being trained to develop special collective behaviors without using a processor.

Models of machine learning offer the opportunity of studying higher cognitive functions and their relation within simple models. This follows from Marr's intuition that such functions can be studied at the conceptual level irrespective of whether they are implemented in-silico or in a biological brain. This requires an approach to learning as "making sense of data that make sense" which is independent of what is learned, how and why. I'd be happy to share my thoughts on issues such as the difference between learning and understanding, the determinants of the emergence of abstraction and, possibly, a minimal approach to awareness.

While mechanobiological regulation of many epithelial carcinomas has been well-studied for decades, its impact on oral squamous cell carcinoma (OSCC) is less understood, despite overall survival rates for OSCC significantly dropping once the disease has metastasized. Here we examined how niche stiffness drives epithelial-to-mesenchymal transition (EMT), how it enables OSCC cells to acquire “mechanical memory”, and what mechanisms are used to recall such memories. We found that invasive cells overexpress myosin II (vs. noninvasive cells) consistent with OSCC. However, prolonged exposure of noninvasive cells to a stiff niche or contractile agonists up-regulated myosin and EMT markers and enabled them to migrate as fast as invasive cells, which persisted even when the niche softened and indicated “memory” of their prior niche. Stiffness-mediated mesenchymal phenotype acquisition required AKT signaling and was also observed in patient samples, whereas phenotype recall on soft substrates required focal adhesion kinase (FAK) activity. Phenotype durability was further observed in transcriptomic differences between preconditioned cells cultured without or with FAK or AKT antagonists, and such transcriptional differences corresponded to discrepant patient outcomes. While intrinsic mechanisms are important, cell extrinsic signaling within tumor has not been studied well. Co-cultlure of heterotypic OSCC cell lines enhanced the invasiveness of naïve cells via a specific cytokine signature. Supplementing normal growth media with those cytokines comparatively increased naïve cells’ motility via MAPK and AKT signaling pathways, and their withdrawal did not alter the phenotype. These data suggest that mechanical memory, mediated by contractility and paracrine signals via distinct kinase signaling, may be necessary for cancer dissemination.

Take a thin plastic sheet and crumple it into a ball. It might not be immediately evident, but this seemingly mundane object exhibits many of the hallmark behaviors shared by non-equilibrium disordered systems. These include logarithmic aging spanning many timescales, intermittent mechanical responses, crackling noise, avalanche dynamics, and a range of memory effects. In particular, the sheet can adapt its response to repeated external driving, and thereby consolidate a memory of the drive. Multiple such memories can be encoded in and retrieved from a single sheet. Using experiments that combine global mechanical measurements, local probing, acoustic measurements, and 3D imaging of crumpled sheets, we build a mesoscopic description of their mechanics. These reveal that the adaptation and memory behaviors emerge from the collective dynamics of mesoscopic, bistable elements within the sheet: localized geometric instabilities that act as coupled, hysteretic, two-state degrees of freedom. Based on this picture, we develop a numerical model of a disordered network of bistable elastic elements that corroborates all our observations. The model highlights the role of interactions and frustration between instabilities in driving these behaviors. Finally, inspired by our findings, we create mechanical metamaterials composed of highly frustrated by-stable elements, and show that these can be designed to perform additional information processing tasks such as sequence dependence. D. Shohat, D. Hexner and Y. Lahini, PNAS 119 (28) 2200028119 (2022) D. Shohat and Y. Lahini, PRL 130, 048202 (2023) Sirote-Katz, Shohat, et al., Nat. Comm. (2024)

Non-genetic inheritance plays a fundamental role in determining cellular properties in future generations and in restraining the proliferation of non-genetic cell-to-cell variation over time. This in turn can influence the cell’s ability to respond and adapt to new environmental conditions. It is, therefore, important to be able to measure it and quantitatively characterize it reliably. In this talk I will present our newly developed method that allows measuring and characterizing non-genetic inheritance (or cellular memory) in the simple bacterial model organism E. coli. The method utilizes a novel microfluidic device, coined “sisters machine”, that enables us to track and measure how two sister cells become different from each other over time. Our measurements reveal how non-genetic inheritance contributes to regulating the various cellular properties (e.g. size, growth rate, etc.) in future generations. We find that non-genetic cellular memory is property specific, and can last up to ∼10 generations, but decreases under stress. The results obtained from this study can help uncover mechanisms of non-genetic inheritance and adaptation to stress.

I will discuss structure and dynamics that can lead to the automatic generation of behaviorally relevant neural circuits, without the necessity for extrinsic learning mechanisms.

Many complex systems in nature exhibit self-organization and emergence without a central processing unit akin to the animal brain. In what sense can such phenomena be considered learning? As a particular context, we present a model of anticipatory systems that uses past information to predict the future of the environment and take action accordingly. These are thus dynamical systems that can be considered to learn (and forget) from their environment, on time scales given by the external inputs. We show that anticipation provides the system with certain performance advantages, which may be favorable in the evolutionary landscape, but it also comes with some drawbacks.

Neural networks in the brain and artificial neural networks (ANNs) in silico are both able to learn complex functionalities. While each artificial neuron is updated based on global information, using a central processor (CPU) and memory, each real neuron in the brain updates itself without external CPU. In this talk I will describe laboratory realizations of such self-learning without use of CPU or memory. Our systems consist of a network of identical variable-resistive elements that self-adjust using a local rule based on the voltage drops they experience under contrastive boundary conditions. As such, they have many brain-like advantages over ANNs and enable study of learning as a bottom-up emergent process.

Traditional view of adaptation focuses on selection of adaptive variations without regard to how these variations come about in the first place (no consideration of emergence). And yet, every single animal is constantly undergoing newly forming variations in its epigenome, microbiome and even in its somatic genome. Many of these variations appear in novel combinations that are unique to the individual. Since any new variation is potentially harmful, it is not clear how every individual can tolerate large numbers of novel variations that are forming during their lifetime. We previously hypothesized that every individual acquires new adaptations by undergoing stochastic variations under (existing) mechanistic constraints that suppress the likelihood of undergoing non-viable changes (i.e. reaching non-viable states). Our group is testing this hypothesis using experimental models of coping with severe conditions of stress mimicking unforeseen challenges. Experimental work-in-progress provide direct evidence supporting emergent adaptation by constrained exploration (“adaptive improvisation”) within the lifetime of individual flies, as well as during the generation time of individual cancer cells in culture. The feasibility of emergent adaptation is further supported by theoretical models of coping with “unforeseen challenges” presented to a specific class of complex systems. I will describe the conceptual problem of emergent adaptation and its hypothesized solution, present the experimental support, and discuss implications to our view of evolution. If time permits, I will also present and discuss theoretical work-in-progress of emergent adaptation.

Plants solve complex navigational problems, continuously negotiating their unstructured and changing environment. They strategically redirect their growth to optimize photosynthesis in response to fluctuating light sources, while minimizing mechanical strains. While they have no brain or neural system, they can sense their environment, process sensory information, and plan strategic growth movements. Since plants are distributed systems, with no central control, underlying computational processes must be emergent properties of the tissue. I will discuss how plants accomplish decentralized computation at the tissue level, underpinning integration of sensory information over space and time.

Biological and artificial systems encode information through several complex nonlinear operations, making their exact study a formidable challenge. These internal mechanisms often occur across multiple timescales and process external signals to enable functional output responses. In this work, we focus on two widely implemented paradigms: nonlinear summation, where signals are first processed independently and then combined; and nonlinear integration, where they are combined first and then processed. We study a general model where the input signal is propagated to an output unit through a processing layer via nonlinear activation functions. Further, we distinguish between the two cases of fast and slow processing timescales. We demonstrate that integration and fast-processing capabilities systematically enhance input-output mutual information over a wide range of parameters for large-scale systems, while simultaneously enabling tunable input discrimination. Moreover, we reveal that high-dimensional embeddings and low-dimensional projections emerge naturally as optimal competing strategies. Our results uncover the foundational features of nonlinear information processing with profound implications for both biological and artificial systems.

We examine overparameterized fully-connected and convolutional neural network training, with $N$ ($ \ll$ \#parameters) samples labeled by a small teacher network. We examine training with noisy gradient flow, i.e. gradient descent with infinitesimal step size, perturbed with $\Omega(1/N)$-variance Gaussian noise (i.e., Langevin flow). We prove that, if the training achieves a small error, the model generalizes well (i.e., with roughly the sample complexity of the teacher). We assume uniform initialization, that the flow is constrained to some finite volume, and is quantized post-training, similarly to common practices. The proof builds on a generalized second law of thermodynamics, by which the noisy gradient flow cannot ``squeeze'' the distribution of iterates into a tiny ``bad'' volume in parameter space (where the geometry of optimization specifies the ``volume'' here).