Keynote Talks
Professor Peter Dayan
Learning from scratch: Non-parametric models of task acquisition over the long run
Director, Department of Computational Neuroscience, Max Planck Institute for Biological Cybernetics, Tübingen
Professor Andrew Adamatzky
Fungal brain
Director, Unconventional Computing Laboratory, University of the West of England
Invited Talks
Professor Rahul Bhui (MIT)
Ambiguity and confirmation bias in reward learning
We tend to interpret feedback in ways that confirm our pre-existing beliefs. This confirmation bias is often treated as irrational, but may have adaptive foundations. In this project, we propose a new Bayesian computational model of confirmation bias and a novel experimental paradigm to study its impact on learning. When faced with an ambiguous outcome, we must form the most accurate interpretation we can by making use of all available information, which includes our pre-existing beliefs. Confirmation bias may thus constitute an inductive bias that speeds up learning, analogous to missing data imputation. We test this theory using a reward learning task in which participants are only provided partial information about outcomes, allowing more leeway for subjective interpretation. We find that our Bayesian model better explains the dynamics of behavior and stated beliefs compared to more traditional learning models, supporting an adaptive basis for confirmation biased learning from repeated feedback.
Dr Aenne Brielmann (Max Planck, Tübingen)
Boredom in aesthetic experiences
When was the last time you skipped a song on your playlist or scrolled over an image on your Instagram feed? Why? We wager that aesthetic boredom drove your choice. According to our theory, aesthetic boredom arises when sensory value decreases. We assume that sensory experiences are valuable to the extent that they help increase sensory processing efficiency now and in the future. We distinguish between absolute boredom that arises in response to one experience without need for comparison and relative boredom. Absolute boredom arises when: 1) its continuation leads to a decrease in expected long-term processing efficiency, which 2) outweighs short-term increasing processing efficiency for itself. Relative boredom for a sensory experience arises when other available experiences have higher values. Boredom then serves as a signal for observers that dwelling longer in their current environment is suboptimal in the long-run and motivates them to seek out new sensory stimulation.
Professor Kevin Burrage (QUT)
Coping with tissue heterogeneity: modelling the electrophysiology of the human heart
Professor Carina Curto (Penn State)
Threshold-linear networks (TLNs) display a wide variety of nonlinear dynamics including multistability, limit cycles, quasiperiodic attractors, and chaos. Over the past few years, we have developed a detailed mathematical theory relating stable and unstable fixed points of TLNs to graph-theoretic properties of the underlying network. These results enable us to design networks that count stimulus pulses, track position, and encode multiple locomotive gaits in a single central pattern generator circuit.
Sequences and modularity of dynamic attractors in inhibition-dominated neural networks
Professor Marc Howard (Boston University)
Constructing a continuous estimate of the future
Professor Kobi Kremnitzer (University of Oxford)
Scientific theories of consciousness, the closure of the (current) physical, and collapse
In this talk I will argue that in order to develop scientific theories of consciousness we need to ask two questions: What is consciousness, and how does consciousness interact with the physical world. I will look at possible answers to the second question and relate it to considerations about the closure of the (current) physical. If we assume that current theories of physics are complete, there is very little room for consciousness to interact with the physical world. If we do not assume that current physics gives a complete description of the physical world, I will explain how under mild assumptions quantum collapse theories would be involved in modelling consciousness. I will then give some examples of how such theories look like.
This talk is based on joint work with Johannes Kleiner.
Professor Michael Levin (TUFTS)
Neuroscience outside the nervous system: bioelectric basis of basal cognition in morphogenesis
The remarkable problem-solving capacities of brains have their origin in a far older system for processing information in electrical networks: developmental bioelectricity. What did cell groups think about before there were brains and muscles used to navigate the 3-dimensional world? In this talk, I will describe how evolution pivoted from using bioelectrical networks to control the configuration of the body, using a kind of basal intelligence to solve problems in anatomical morphospace. I will also show our attempts to reprogram this system toward regenerative medicine, extending neuroscience concepts beyond the nervous system to gain control of growth and form in biomedical and synthetic bioengineering contexts.
Andrew Ligeralde (UC Berkeley)
Geometry reveals a role of retinal waves as biologically plausible pre-training signals
Before eye-opening, neurons in the developing mammalian retina spontaneously fire in correlated activity patterns known as retinal waves. Experimental evidence suggests retinal waves strongly influence sensory representations prior to the onset of the visual experience. We aim to elucidate the computational role of retinal waves by using them as pre-training signals for neural networks. We consider simulated activity patterns generated by a model retina as well as real activity patterns observed experimentally in a developing mouse retina.
We show that pre-training a classifier with a biologically plausible Hebbian learning rule on both simulated and real wave patterns improves the separability of the network’s internal representations. In particular, the pre-trained networks achieve higher classification accuracy and exhibit internal representations with higher manifold capacity when compared with randomly initialized networks with no pre-training.
Dr Jonathan Mason (University of Oxford)
Setting the Benchmark Test for Archetypal Models in Mathematical Consciousness Science
The scientific study of consciousness is a very active field these days, and yet a scientific understanding of consciousness is still in need of a breakthrough that can be widely accepted. The problem is not a lack of ideas; indeed one can probably find 20 or more hypotheses in the scientific literature, with varying degrees and types of evidence and justification. Mathematical Consciousness Science (MCS) is turning hypotheses in to predictive mathematical models that are more open to objective scientific testing which can separate out the good hypotheses and models from the bad. In this talk I propose that MCS will benefit from the introduction of Benchmark Testing which has, all be it with some controversy, revolutionised developments in the field of Machine Learning and AI. I will provide an initial outline of how such Benchmark Testing would work through involving what one might describe as the “bulk properties” of consciousness.
Professor Dan V. Nicolau Jr (King’s College London)
Information Complexity and Unified Theories of the Brain
Professor Pedro Resende (Instituto Superior Técnico)
The emergence of geometric worldviews in qualia space
In the model of qualia based on measurement spaces the points are the qualia and the open sets are the pure concepts, which are the communicable finite chunks of classical information. Moreover, algebraic structure conveys subjective time and the subjective experience of logical abstraction. In this talk I discuss the subspace of qualia S associated to any organism which, such as a human being, (1) possesses cognitive capabilities that allow it to compute on real numbers by finite approximations, and (2) has subjective experience associated to its pure concepts. I argue that if (3) such an organism is "macroscopic," thus not having direct awareness of quantum mechanical interference effects, then S is isomorphic to the topology of another space X, of "points," or "states," which is second-countable and locally compact Hausdorff. This suggests an explanation of the apparent fact that humans possess a geometric view of the world, and also suggests that the same is likely to happen to any organism that shares the general features (1), (2) and (3).
References:
- P. Resende, Qualia as physical measurements: a mathematical model of qualia and pure concepts (2022); arXiv:2203.10602.
- An abstract theory of physical measurements, Found. Phys. 51, 108 (2021); DOI 10.1007/s10701-021-00513-1.
Contributed Talks
Golnaz Baghdadi (Amirkabir University of Technology)
Phase Synchronization in an Oscillatory Network and Response Time Variability during a Sustained Attention Task
Continuous performance tests (CPTs) are sustained attention tasks with the repetitive presentation of similar external stimuli. Intra-individual variability in response to these similar external stimuli is considered a significant cognitive performance index. Although intra-individual response time variability has been studied for a long time, the neurocognitive origin and reason for such changes in response to the same stimuli are not fully understood. So far, this variability has been attributed to various factors, such as brain connectivity, maturation, or some dissociable latent brain state features. This study aims to show how the phase synchronization phenomenon, frequently reported during different neurocognitive tasks, can cause brain response variability.
We have proposed a network of oscillatory units that can couple through forced or mutual synchronization. These units correspond to the bottom-up and top-down processors of the brain involved in the attention control system. The input of the model was a sequence of stimuli similar to those in CPTs. The model's output was the response to these stimuli, and we specifically recorded the model response time to these stimuli. To evaluate the model's performance, we used response times recorded from two groups of typically developing children and children with attention deficit disorder (ADD) during an integrated visual and auditory CPT.
Analytical investigations of the proposed model outputs' pattern can suggest a possible role of phase synchronization phenomenon in brain response variability. Simulations and recorded human data also provided some possible reasons for higher intra-individual response variability in children with ADD compared to typically developing children.
Taylor Beck (The Pierrepont School)
The Motivated Mind: Bipolar Disorder and the Roots of Drive
Psychiatric disease offers a window on an old mystery: How does motivation arise from cells? The question puzzled Nietzsche, William James, and Freud, not to mention Heraclitus, and persists today. Our goals are dynamic, ever mutating: belief-states in constant flux. What metabolic process gives value to these drives, and switches them? We all experience shifts in interest – about a job, a romantic partner, a home; the “flow state” of focused effort that often eludes us. In mental illness, such changes are intense: depressive anhedonia, indifference to food, sex, work; hyperactive mania, obsessed to distraction. Bipolar disorder, characterized by abrupt, episodic shifts in goal seeking and reward, offers a model for these shifts in values and how they happen in us all. Neurons can now be grown from living psychiatric patients in vitro, revealing how these differ from healthy cells: “psychiatry in a dish.” Results from this work suggest new clues to the old riddle of metabolism, motivation and mood. This talk will argue that a basal perspective, rooted in cells, may help us understand this most human of experiences.
Taylor Beck is a writer, a teacher, and a former neuroscience researcher. He holds a BA from Princeton in Neuroscience, an MS from MIT, and an MA from NYU. His journalism and essays have appeared in The Atlantic, Scientific American, and the Los Angeles Review of Books, among other publications.
Dr Spyridon Chavlis (FORTH)
Empowering deep learning architectures by adding biological features
Deep learning algorithms have made remarkable achievements in numerous fields over the past decade. Inspired by biological neuronal networks in our brains, artificial neural networks, a key component of almost all deep learning architectures, are typically constructed with various nodes simulated as a simple linear weighted sum of inputs followed by feeding the result through a nonlinear activation function. A typical biological neuron, as revealed by both experimental and modeling studies, is a far more complex structure than those mentioned above simple linear integrator. More specifically, synaptic inputs can actively modulate neighboring synaptic activity and generate local events, the so-called dendritic spikes, along the dendritic tree via an abundance of ion channels; therefore, dendritic structures are highly nonlinear. Such properties significantly elevate the contribution of dendrites in neuronal outputs and empower neuronal networks with much more significant information processing power. In addition, the most widely-used learning algorithm in artificial neural networks is the backpropagation-of-the-error, which retrogradely broadcasts a global error to fine-tune the model parameters. Nevertheless, in biological systems, neurons communicate with each other via different rules based on features other than a global error. Here, inspired by the rules governing animals’ brains, we modeled dendrites accompanied by their nonlinearities. We also added biologically plausible learning rules. We showed that our algorithm could achieve excellent performance and that our model surpasses traditional deep neural networks given the same architecture complexity. The proposed structure can be adopted in any deep learning architecture substituting the fully connected layer.
Dr Aslan Satary Dizaji (AutocurriculaLab)
A Multi-agent Reinforcement Learning Study of Social Welfare and Fairness under Libertarian and Utilitarian Systems
A society is composed of many individual entities including humans, organizations, a government, and their ecological environment. Their interactions with each other pave a way for progress to build a more prosperous society. The question which has always perplexed me is that what are the causes and roots of what we observe currently in our immediate or distant surroundings and how we can go one step further toward a better future. More specifically, I am fascinated by the following question: What is the origin of prosperity or poverty, in-equality, and morality in a society? While experimental investigations of these questions are valuable, I argue that they are expensive and time-consuming. On the other hand, until very recently, computer simulations have not shown a realistic picture of micro- and macro-foundations of societal processes. I believe this has changed recently due to introduction of more realistic multi-agent reinforcement learning (MARL) environments to simulate actual social, cultural, and economical processes. The most important feature of these new environments is that they can independently produce curriculum which is called Autocurricula. Autocurricula refers to challenges that arise naturally due to interactions among agents and between agents and their changing environment which agents must adapt. Here, I intend to answer three economically and socially relevant questions by modifying one of the most recently introduced MARL environments - the AI-Economist. There is an interesting debate between libertarianism or procedural theories of justice and utilitarianism or consequentialist theories of justice in welfare economics. While the latter insists on the important role of a central planner to determine what is just or unjust in the society, the former emphasizes on the people’s freedom to choose what is just or unjust for them. More specifically, inspired by the welfare economics and using an in-silico MARL experiments, I intend to answer the following three specific questions:
1. What is the best possible role can be defined for a central social planner in a society? Should it only passively determine the tax rates but let the agents decide individually how much to invest their taxes on different economic sectors - more or less similar to libertarian tradition - or should it actively determine the tax rates, collect and invest them on various economic sectors trying to optimize some pre-specified social welfare functions - more or less similar to utilitarian tradition. To answer this question, I introduce a modified version of the original framework of the Gather-Trade-Build environment of the AI-Economist by increasing the complexity of the environment.
2. Using Shapley value - a concept from coalitional game theory - I suggest to quantify the fairness of the system towards individual agents, and also the root of inequality in the society. More specifically, I would like to investigate if the way that the system counts the votes, collects the taxes, and invests them can be considered fair considering the contributions of the agents in the society.
3. What kind of society would be the product of a libertarian passive social planner compared to a utilitarian active central planner? Do the agents will be more moral and fair in the libertarian society compared to the utilitarian one? To answer this question, I use a metric - inequity aversion as a signature of fairness - to quantify the fairness of the agents trained under libertarian system versus the agents trained under utilitarian system. Moreover, I will mix the populations of the agents trained under both systems to compare directly their behaviors in the same environment.
Generally, I believe that MARL Autocurricula has a great potential to perform in-silico experimental economics with much less financial and time burden.
Professor Constantine Dovrolis (Georgia Institute of Technology)
The hourglass architecture of multi-sensory integration and lifelong learning
What is an appropriate neural architecture for lifelong learning, where both the sensory stream (inputs) and the cognitive tasks (outputs) are often subject to novelty? How can the brain exploit the modularity that is common in different perceptual and cognitive tasks to re-use intermediate-level representations and sub-tasks? We investigate these questions in the context of multi-sensory integration in the mouse connectome using the mathematical framework of network diffusion. In particular, we model the flow of evoked activity, initiated by stimuli at primary sensory regions, using the Asynchronous Linear Threshold (ALT) diffusion model. The ALT model captures how evoked activity that originates at a given region of the cortex “ripples through” other brain regions (referred to as an activation cascade). We have validated the accuracy of this model using Voltage-Sensitive Dye (VSD) imaging data from earlier experiments. We find that a small number of brain regions – the claustrum and the parietal temporal cortex being at the top of the list – are involved in almost all cortical sensory streams. A major result of this study is that relatively few cortical regions are responsible for integrating almost all sensory information. This finding supports the idea that multisensory integration is performed through an “hourglass architecture”. The benefit of an hourglass architecture is that it first reduces the input dimensionality of the sensory stimulus at few intermediate-level modules that reside at the “hourglass waist”. Second, it re-uses those compressed intermediate-level representations across higher-level tasks, reducing redundancy between the latter.
Dr Andrew Duggins (University of Sydney)
The flat plane between experiential horizons of motion
An event that a self-rotating subject approaches is perceived to precede a simultaneous event in the receding environment. One explanation for such simultaneity distortion invokes covert orienting of attention to the upcoming region of space, shortening the latency to representation of stimuli presented in this region. The explanation is problematic, particularly where the subject infers accumulation of angular velocity from vestibular input, in that rotatory vestibular stimulation displaces the subjective ‘straight ahead’ away from the purported direction of covert attention.
In an alternative reference frameshift theory, I adapt the Lorentz transformations of special relativity to the rotating reference frames of subjects and their environments: Just as objective motion is bounded by the speed of light, self-rotation is perceived within finite limits of angular velocity. I equate these limits with asymptotic vestibular nystagmus slow-phase velocity of 200º/s (during whole-body constant angular acceleration in the dark) to derive the experimental magnitude of simultaneity distortion. Anisometry of an angularly accelerating reference frame explains the ‘straight ahead’ shift.
A hyperbolic tangent transformation from the infinite domain of objective angular velocity, derived from special relativity, fits the gain function of the vestibulo-ocular reflex. It should also be the ideal objective-to-subjective transfer function. The efficient coding principle, describing in information theoretic terms the optimal transformation into neural code of an environmental stimulus variable, is proposed also to constrain perceptual experience. Hypothetically, the statistics of body rotation during prolonged environmental immersion are such that the tanh transformation maximizes the entropy of self-rotation perception.
Dr Hamza Giaffar (Halıcıoğlu Data Science Institute; UCSD)
Life in the brain: statistical learning and discovery processes via Darwinian Neurodynamics
The process of discovery involves efficient search in vast combinatorial and compositional spaces, while the process of life-long learning requires statistical inference over multiple timescales. These two seemingly disparate aspects of brain function can be unified under a single mathematical framework, learning theory, and can be implemented by a single computational system based on information replication.
We start from the isomorphism between Bayesian hierarchical learning and a model of multilevel selection, which links statistical learning over multiple timescales to replicator (evolutionary) dynamics. We show how a finite neural replicator system evolving under selection can perform approximate statistical inference over shorter time scales and multi-level discovery processes akin to biological evolution over longer timescales.
With the essential mathematical framework in place, we will discuss neural systems that can realize the proposed Darwinian process, which operates over sequential cycles of imperfect copying and selection of neural informational patterns. We elucidate the key concepts in a proof-of-principle model based on dynamical output states of reservoir computers as units of evolution. We show that a population of reservoir computing units is capable of maintaining and continually improving upon existing solutions over rugged combinatorial reward landscapes.
We discuss the high degree of representational redundancy implied by Darwinian neurodynamics by turning to supporting fMRI data and new theoretical ideas. Finally, we will return to the underlying mathematics and demonstrate how learning theory can offer an account of the evolution of a full-blown evolutionary process within the brain.
Dr Chris Hillar (Awecom, Inc)
Retina-inspired Representations of Natural Signals
Theoretical constructs from biophysics have profoundly impacted signal processing and engineering from McCulloch-Pitt's theory of mind seeding modern computer design (von Neumann) to the hierarchical neural network approach underlying recent breakthroughs in deep learning. This history will be discussed as the backdrop for a new biologically-inspired paradigm for digital sensory encoding, originally developed at the Redwood Center for Theoretical Neuroscience (founded by Jeff Hawkins) and now fueling initial applications in signal processing. Consequences for source/channel coding, learning in AI, and experimental neuroscience will also be explored.
Dr Kamila Jozwik (University of Cambridge)
Face dissimilarity judgments are predicted by representational distance in morphable and image-computable models
Human vision is attuned to the subtle differences between individual faces. Yet we lack a quantitative way of predicting how similar two face images look and whether they appear to show the same person. Principal component–based three-dimensional (3D) morphable models are widely used to generate stimuli in face perception research. These models capture the distribution of real human faces in terms of dimensions of physical shape and texture. How well does a “face space” based on these dimensions capture the similarity relationships humans perceive among faces? To answer this, we designed a behavioral task to collect dissimilarity and same/different identity judgments for 232 pairs of realistic faces. Stimuli sampled geometric relationships in a face space derived from principal components of 3D shape and texture (Basel face model [BFM]). We then compared a wide range of models in their ability to predict the data, including the BFM from which faces were generated, an active appearance model derived from face photographs, and image-computable models of visual perception. Euclidean distance in the BFM explained both dissimilarity and identity judgments surprisingly well. In a comparison against 16 diverse models, BFM distance was competitive with representational distances in state-of-the-art deep neural networks (DNNs), including novel DNNs trained on BFM synthetic identities or BFM latents. Models capturing the distribution of face shape and texture across individuals are not only useful tools for stimulus generation. They also capture important information about how faces are perceived, suggesting that human face representations are tuned to the statistical distribution of faces.
Dr Nir Lahav (Bar Ilan University)
Can physics solve the hard problem of consciousness? The new solution of the relativistic theory of consciousness
There is an explanatory gap between our scientific knowledge of consciousness and its subjective, phenomenal aspects, referred to as the “hard problem” of consciousness. How come we can only measure in the brain neural activity (third-person point of view) and not the actual conscious experience itself (first-person point of view)? And how can the brain create such a phenomenon that cannot be measured from third-person point of view? No theory of consciousness can be complete without answering these fundamental questions. Naturalistic dualists argue that consciousness is composed of a private element. Illusionists, on the other hand, argue that it’s a cognitive illusion. We contend that both the positions are flawed because they assume consciousness to be an absolute property that doesn’t depend on the observer. We developed a new physical approach of a relativistic theory of consciousness in which a system either has or doesn’t have phenomenal consciousness with respect to some observer. Phenomenal consciousness is neither private nor delusional, just relativistic. In the frame of reference of the cognitive system, it will be observable (first-person perspective) and in the other cognitive frame of reference it will not (third-person perspective). These two cognitive frames of reference are both correct, just as in the case of an observer that claims to be at rest while another will claim that the observer has constant velocity. Neither observers frame can be privileged, as they both describe the same underlying reality. Based on relativistic phenomena in physics we developed a mathematical formalization for consciousness which dissolves the hard problem.
Dr Steeve Laquitaine (Leuven University)
Pruning for efficiency in Hopfield networks
The mammalian brain forms intricate connectivity patterns, yet its connectivity is ubiquitously sparse, enabling efficient information processing. Hopfield networks have been proposed as a schematic model of auto-associative memory retrieval in the brain; they learn to store memories by modifying their connections’ weights, proxies for biological synaptic strengths. But Hopfield networks are fully connected, which is at odds with the brain’s sparse connectivity. In this work, we ask how a Hopfield network’s memory retrieval accuracy changes when some of its connections are pruned. We hypothesize that the network’s most unstable weights are uninformative for the retrieval task. To test this, we measured the retrieval accuracy of a 32-neuron Hopfield network trained to classify white noise images, before and after we pruned its most variable weights. We found that the network maintains maximal accuracy up to a critical sparsity level, above which accuracy suddenly drops. We also found that the relationship between sparsity level and accuracy depends on the network's storage capacity known as “loading ratio”- the number of images to the number of neurons in the network. When the ratio is low, the accuracy is maximal up to a critical sparsity, then drops. Surprisingly, in the challenging regimes, the relationships are concave, rising to an intermediate sparsity level then dropping again, rather than linearly decreasing when pruning increases. As a control we found that uninformed stochastic pruning produced much poorer accuracy/sparsity tradeoffs than variance-guided pruning, demonstrating that the network can exploit its weights’ variance to achieve better energy-efficiency/accuracy tradeoff.
Dr Thomas Langlois (Princeton)
3D Perspective Memory Priors Reflect Efficient Semantic Categories
An essential function of the human visual system is to encode noisy sensory percepts into memory such as perspective views of 3D visual objects. Due to limited perceptual resources, the visual system forms internal representations by combining noisy sensory information with strong perceptual priors about the structure of these percepts in order to optimize a trade-off between accuracy and efficiency during inference. We reveal detailed priors in 3D perspective memory by iterating a visual memory task (serial reproduction) where the response of one participant becomes the stimulus for the next. We reveal 3D perspective memory priors in unprecedented detail using data from hundreds of respondents on Amazon Mechanical Turk (AMT). We propose a category adjustment model of human 3D memory based on the Information Bottleneck (IB) principle that simulates memory biases as a pull towards efficient semantic categories derived from a lexicon of view words. Our results show that memory results do not correspond to canonical views and that traditional explanations such as the frequency hypothesis or the maximal information hypothesis cannot account for biases in memory.
Li Xin Lim (Purdue University)
A neurobiological inspired computational model of the interaction of multiple systems in category learning
For years, attention has been drawn to explaining how perceptual category learning is mediated by multiple psychological and biological learning systems. However, the interaction between learning systems is assumed but rarely the theoretical focus. Generally, multiple-system theories include a model-based decision system supported by the prefrontal cortex and a model-free decision system supported by the striatum. We propose a neurobiological model to describe the interactions and switching between model-based and model-free decision systems in category learning. The model focuses on the switching mechanism with spiking neurons to simulate neuronal activity of the hyperdirect pathway of the basal ganglia. The hyperdirect pathway gates the transmission of responses from the model-free learning system (in the striatum) for action selection. We propose that the model-free system’s response is inhibited when the model-based system (in the prefrontal cortex) is in control of the response. However, if model-based system fails, inhibition to the model- free system’s response is reduced. The reduction in inhibition results in the acceptance of responses from both learning systems to compete for action control. The model was used to simulate information-integration category learning data from undergraduate students published in Lim & Hélie (2019) (for uncued switching) as well as cue-based trial-by-trial switching category learning data from different age groups from Hélie & Fansher (2018). The simulation results suggest that poor system-switching capability may be related to lower tonic dopamine level, higher susceptibility to proactive interference, and poor strategy searching ability.
Dr Christina Merrick (UCSF)
Selective reduction in loss-related behavior during subthalamic deep brain stimulation in Parkinson’s disease
Behavioral symptoms such as apathy and impulsivity represent a prevalent and disabling feature of Parkinson’s disease (PD), yet there exist limited treatments. Currently available therapies for PD to treat motor symptoms such as deep brain stimulation (DBS) and dopaminergic medication may even exacerbate these behavioral symptoms. To investigate how impulsivity might be affected by DBS, twenty patients with PD implanted with the Percept™ PC, the first FDA approved DBS device that is sense enabled during stimulation, were recruited (STN = 10; GPi = 10). Patients played multiple rounds of a value-based decision-making task while DBS was ON or OFF. For each trial, the participant chose between a gamble option or a guaranteed option to win points across three different trial types: Loss trials (potential to loss points), Gain trials (potential to gain points) and Mixed trials (potential to gain or loss points). Patients with STN DBS gambled more on Loss trials when DBS was ON compared to OFF. When DBS was OFF, Theta (4-8 Hz) power, time-locked to stimulus onset, correlated with expected value of the gamble option. This relationship was disrupted when DBS was ON. In comparison, the same relationship was found for Theta power in Gain and Mixed trials when DBS was ON or OFF. Currently, adaptive DBS (aDBS) is being investigated to treat motor symptoms; one goal of investigating the neural correlates of behavioral symptoms is the hope that they can be incorporated into aDBS algorithms in the future, improving the quality of life for PD patients.
Anja Meunier (University of Vienna)
A mathematical framework for bridging Marr's levels
The question of how algorithms in general and cognitive skills in particular are implemented by our brains is at the core of cognitive computational neuroscience. Bridging the gap between Marr's levels, the computational, the algorithmic and the implementational (Marr 1982), has become one of the fundamental challenges of the field. Nevertheless, the notions of what it means for physical systems such as our brains to implement an algorithm remain surprisingly vague. Even without a precise definition, modern methods of comparison have yielded fascinating insights into similarities between artificial and biological neural networks (Seeliger 2018, Güçlü 2016). However, these results can only hint at a possible structural resemblance as long as there is no established theoretical basis which can justify a more definitive conclusion about whether these obviously different physical systems implement the same algorithms. As such, we argue that a rigorous mathematical theory of algorithms and their physical implementations is indispensable to truly leverage the power of artificial intelligence in cognitive computational neuroscience. In this work we suggest mathematical definitions of the terms “algorithm” and “implementation” and propose an encoding-decoding method to test empirically whether a physical system implements a specific algorithm, which we demonstrate on simulated toy data. Our conceptual framework thus contributes to the efforts in cognitive computational neuroscience to develop rigorous theories that can link the computational, algorithmic, and implementational levels.
Kerry Nix (University of Pennsylvania)
Detection of Language Lateralization using Spectral Analysis of High-Density EEG
We assessed if lateralized brain responses to a language task can be detected with spectral analysis of high-density EEG.
Twenty right-handed, neurotypical adults (28+/-10 years; 5 males) performed a verb-generation [VG] task along with two control tasks (passive listening and word repetition) while EEG was recorded. The event related spectral perturbation (ERSP; a ratio of task:baseline EEG power) of theta, alpha, beta, and gamma frequency bands were calculated for 3 regions of interest bilaterally, including 2 language (superior temporal and inferior frontal [ST, IF]) and 1 control (Occipital) region. We tested whether tasks elicited: (1) changes in spectral power from baseline; and (2) asymmetries in ERSP between matched left and right regions.
Compared to baseline, VG elicited a significant decrease in beta power in the left IF (-0.37 +/- 0.53, t=-3.12, p=0.006) and in all regions bilaterally in the gamma band. VG also elicited asymmetric ERSP responses between left vs. right hemispheres in: (1) IF in the beta band (right ERSP greater than left by 0.23 +/- 0.37, t(19)=-2.80, p=0.0114); and (2) ST in the alpha band (right ERSP greater by 0.48 +/- 0.63, t(19)=-3.36, p=0.003). No changes from baseline or hemispheric asymmetries were noted in language regions during control tasks. On the individual level, 80% of participants showed decrease from baseline in left IF beta activity and 80% showed greater right than left ST alpha ERSP; 90% of participants showed at least one of these two findings.
Spectral EEG analysis detects lateralized responses during language tasks in frontal and temporal regions. With further development, quantitative EEG analysis could provide an inexpensive and readily available language lateralization modality.
Sotirios Panagiotou (Erasmus MC)
EDEN: a NeuroML-native neural simulator
Modern neuroscience employs in silico experimentation on ever-increasing and more detailed neural networks. The high modeling detail goes hand in hand with the need for high model reproducibility, reusability and transparency. Besides, the size of the models and the long timescales under study mandate the use of a simulation system with high computational performance, so as to provide an acceptable time to result. In this work, we present EDEN (Extensible Dynamics Engine for Networks), a new general-purpose, NeuroML-based neural simulator that achieves both high model flexibility and high computational performance, through an innovative model-analysis and code-generation technique. The simulator runs NeuroML-v2 models directly, eliminating the need for users to learn yet another simulator-specific, model-specification language. EDEN's functional correctness and computational performance were assessed through NeuroML models available on the NeuroML-DB and Open Source Brain model repositories. In qualitative experiments, the results produced by EDEN were verified against the established NEURON simulator, for a wide range of models. At the same time, computational-performance benchmarks reveal that EDEN runs from one to nearly two orders-of-magnitude faster than NEURON on a typical desktop computer, and does so without additional effort from the user. Finally, and without added user effort, EDEN has been built from scratch to scale seamlessly over multiple CPUs and across computer clusters, when available.
Paper: https://www.frontiersin.org/articles/10.3389/fninf.2022.724336/full
Orr Paradise (UC Berkeley)
A Theory of Unsupervised Machine Translation Motivated by Understanding Whale Communication
Unsupervised translation generally refers to the challenging task of translating between two languages without parallel translations, i.e., from two separate monolingual corpora. In this work, we propose an information-theoretic framework of unsupervised translation that can be well suited even for the case where the source language is that of highly intelligent animals, such as whales, and the target language is a human language, such as English.
We identify three conditions that combined allow for unsupervised translation: (1) a ground-truth translation function exists and could be learned if enough source-target data pairs were available, that is, in a supervised setting; (2) there is access to an prior distribution over the target language that estimates the likelihood that a sentence was translated from the source language (a `plausible' translation); and (3) most alterations (ambiguations) of translations are deemed implausible (i.e., unlikely) by the prior.
We then give an (inefficient) algorithm which, given oracle access to the prior and enough unlabeled source examples as input, outputs a provably accurate translation function. We prove upper bounds on the number of samples needed by our algorithm to succeed when conditions (1-3) are met. Surprisingly, our analysis suggests that the amount of source data required (information theoretically) for unsupervised translation is not significantly greater than that of supervised translation, i.e., the standard case where one has parallel translated data for training.
To support the viability of our theory, we propose a simplified probabilistic language model: the random sub-tree language model, in which sentences correspond to paths in a randomly-labeled tree. We prove that random sub-tree languages satisfy conditions (1-3) with high probability, and are therefore translatable by our algorithm.
Our theory is motivated by a recent initiative to translate whale communication using modern machine translation techniques. The recordings of whale communications that are being collected have no parallel human-language data. Our work seeks to inform this ambitious effort by modeling unsupervised translation. We are further motivated by recent empirical work, reported in the machine learning literature, demonstrating that unsupervised translation is possible in certain settings.
Joint work with Shafi Goldwasser, David Gruber, and Adam Tauman Kalai.
Dr Michael A. Popov (Kant Mission ESA; UK Synaesthesia Association; OMCAN)
Radical thought experiments with human consciousness in Number theory
Number theory may be not necessary be " Divine" as in legendary Kronecker's claim ( " God made the integers, all else is work of man"), but it itself has own history of mathematical experiments with human consciousness: from Zeno's Gedanken thought experiments with evaporation of the mental line of numbers to Srinivasa Ramanujan's thought experiments with directionality of the whole numbers in studies of zeta - function. When we are asking to point out in what place an arrow is when it moves, Zeno (and later William James) showed that it is impossible to stop any thought for introspection before it reaches a conclusion. In other words, mental line of numbers may evaporate before it can be examinated. A phenomenon closely connected with Ramanujan's directionality of the whole numbers was described by St.Dehaene in 1993. When requested to judge whether a number is even or odd, humans are faster when small numbers are on the Left and large numbers on the Right ( human neonates tended to chouse the left side also in Di Giorgio et al experiments). In contrast, in his famous letter to Hardy, Ramanujan suggested that his calculations may prove that 1+2+3+4+...= -1/12 ( i.e.large numbers on the left must produce highly unexpected small numbers on the right! ). Similarly, some neuropsychological experiments with nonhuman creatures, which have not developed sociocultural bias, demonstrated merely weak tendency to prefer the left side Indeed ( G.Vallortigara, 2012, 18). Hence, Ramanujan's paradoxical result could be connected with further neuropsychological experiments suggesting some new " Deep Platonic Model" , where preexisting innate mathematical knowledge is considered as an ordering factor into consciousness.
Dr Ilias Rentzeperis (Universite Paris-Saclay)
Sparser models for visual coding with homeostasis constraints
Experimental evidence suggests that activity in the visual cortex is sparse: only few neurons, out of the large pool that could respond to sensed stimuli, are active at a time. A theoretical support of such evidence is that filters emerging from generative learning models which try to replicate such sparse coding behavior are shaped similarly to the receptive field of V1 neurons, i.e. they are band-pass, oriented, and localized. Such filters are estimated by solving an optimization problem where the square error between the actual image and the reconstructed one is minimized together with a sparsity-promoting term (typically an L1 norm or more complex continuous relaxations of the L0 pseudo-norm) keeping the neural activities as sparse as possible. However, it can be observed that only a subset of the estimated filters converge to useful representations, with the rest being the losers in the competition for activity and never catching up. We probe a possible solution to this issue by adding a homeostatic constraint to the optimization problem and examining the development of the unit receptive fields with less relevant activity. Homeostatic mechanisms offer an alternative way for balancing out activity in populations of neurons by favoring the subsequent activation of the least active neurons. Further, it may also offer a way to adapt the trade-off between image reconstruction and sparsity and subsequently provide systems that can adjust to stimuli with varying degrees of noise.
Chris Rourk (Citizen Scientist)
Modelling action selection in large substantia nigra pars compacta dopamine neurons
Neural networks and “big data” are used to model human action selection. Recent research on the axons of large dopamine neurons of the substantia nigra pars compacta demonstrates that they release dopamine in a manner that is spatially and temporally specific, and that results in action selection. [1, 2] Cortical neurons are directly connected to many synapses of those dopamine neuron axons on dendritic boutons of striatal neurons [3], and afferent signals from those cortical neurons lead activation of the dopamine neurons [4]. Based on this work, it appears that the specific cortical inputs to these striatal neurons are associated with action selection, as predicted by a hypothesized communications mechanism associated with these neurons [5-9]. Neural networks could be used to model the action selection behavior of a specific individual, under certain circumstances and given the right inputs for training.
The types of action selections made in mere exposure studies and other similar studies are instructive. For example, [10] presents a meta-analysis of 268 mere exposure studies, which include tests that range from reading comprehension [11] to ordering samples [12]. In that regard, the reading comprehension studies would be less amenable to modelling than the ordering of samples, which involves action selection. The presentation will discuss ways in which such tests could be designed to generate training inputs to a neural network for predicting the action selections of individuals, such as by developing a range of action selection responses to different scenarios that could provide a boundary or workspace for predicting other action selections within that workspace.
[1] Liu, Changliang, Pragya Goel, and Pascal S. Kaeser. "Spatial and temporal scales of dopamine transmission." Nature Reviews Neuroscience 22.6 (2021): 345-358.
[2] Liu, Changliang, et al. "An action potential initiation mechanism in distal axons for the control of dopamine release." Science 375.6587 (2022): 1378-1385.
[3] Schultz, Wolfram. "Predictive reward signal of dopamine neurons." Journal of neurophysiology 80.1 (1998): 1-27.
[4] Park, Junchol, Luke T. Coddington, and Joshua T. Dudman. "Basal ganglia circuits for action specification." Annual review of neuroscience 43 (2020): 485-507.
[5] Rourk, Christopher John. "Ferritin and neuromelanin “quantum dot” array structures in dopamine neurons of the substantia nigra pars compacta and norepinephrine neurons of the locus coeruleus." Biosystems 171 (2018): 48-58.
[6] Rourk, Christopher J. "Indication of quantum mechanical electron transport in human substantia nigra tissue from conductive atomic force microscopy analysis." Biosystems 179 (2019): 30-38.
[7] Rourk, Christopher J. "Functional neural electron transport." Advances in Quantum Chemistry. Vol. 82. Academic Press, 2020. 25-111.
[8] Rourk, Christopher, et al. "Indication of Strongly Correlated Electron Transport and Mott Insulator in Disordered Multilayer Ferritin Structures (DMFS)." Materials 14.16 (2021): 4527.
[9] Rourk, Chris. "Application of the Catecholaminergic Neuron Electron Transport (CNET) Physical Substrate for Consciousness and Action Selection to Integrated Information Theory." Entropy 24.1 (2022): 91.
[10] Montoya RM, Horton RS, Vevea JL, Citkowicz M, Lauber EA. A re-examination of the mere exposure effect: The influence of repeated exposure on recognition, familiarity, and liking. Psychol Bull. 2017 May;143(5):459-498. doi: 10.1037/bul0000085. Epub 2017 Mar 6. PMID: 28263645.
[11] Lu, Xi, Xiaofei Xie, and Lu Liu. "Inverted U-shaped model: How frequent repetition affects perceived risk." Judgment and Decision making 10.3 (2015): 219.
[12] Hekkert, Paul, Clementine Thurgood, and TW Allan Whitfield. "The mere exposure effect for consumer products as a consequence of existing familiarity and controlled exposure." Acta psychologica 144.2 (2013): 411-417.
Dr Sophia Sanborn (UC Santa Barbara)
Tutorial: Manifolds, Lie Groups, and Geometric Machine Learning for Neuroscience
An emerging set of findings in sensory and motor neuroscience is beginning to illuminate a new perspective on neural coding. Across sensory and motor regions of the brain, neural circuits appear to mirror the geometric and topological structure of the systems they represent--either in their synaptic structure, or in the implicit manifold generated by their activity. This suggests a general computational strategy that is employed throughout the brain to preserve the geometric structure of data throughout stages of information processing. In parallel, there has been a growing recognition of the importance of respecting the geometry of data in deep neural network architectures for learning useful representations for downstream tasks. This has given rise to the nascent sub-field of Geometric Deep Learning.
There is high potential for synergy between these two emerging fields. Geometric deep learning models may allow us to identify and model geometric structure in neural data, while a richer understanding of how the brain represents geometric structure may inform the development of more elegant, expressive, and efficient geometric deep learning models. However, across both fields, there is a need to more deeply engage with the mathematics of differential geometry and group theory, to strengthen the foundation of these lines of work and illuminate new tools for application.
In this tutorial, I provide a concise introduction to differential geometry, group theory, and geometric machine learning as it pertains to understanding representations in the brain, illuminating new possibilities for both theoretical development and application.
Christian Shewmake (UC Berkeley)
Learning Group Structure with Bispectral Neural Networks
We present a neural network architecture for learning representations that are invariant to the actions of groups. The model incorporates the ansatz of the bispectrum, an analytically defined group invariant that is complete—that is, it preserves all signal structure while removing only the variation due to group actions. A bispectral neural network (BNN) is composed of two layers, a linear transform whose weights are learned, followed by a product layer in which each unit computes a fixed multiplicative combination of three units from the output of the first layer. The objective function for training a BNN consists of two terms: one that aims to collapse all variations of a signal due to group action to a single point in the output layer (invariance), and another term that forces the first-layer linear transform to be invertible (information preservation). We show that a BNN trained with this objective learns the irreducible representations of a given group purely from symmetries implicit in the data. Additionally, our parameterization of the model enables one to extract the group’s Cayley table from the learned weights and thus understand its mathematical structure. Because the model has learned the fundamental structure of the group, it can generalize to arbitrary unseen data with the same group structure. Furthermore, the completeness property endows these networks with strong adversarial robustness.
Dr Lisanne Stock (Great Ormond Street Hospital Trust)
Don’t sweat the micro, sweat the macro—a clinical approach to neuroscience
Dr Ilia Sucholutsky (Princeton University)
Learning from almost no data
Many modern machine learning methods (e.g., deep neural networks) require large training sets and suffer from high computational costs and long training times. The ability to generalize from much smaller training sets while maintaining nearly the same accuracy - much like humans - would be very beneficial. To probe the limits of data-efficient learning, we explore a recently proposed regime called 'less-than-one-shot' learning, where agents (human or artificial) must learn N new classes from fewer than N examples. We investigate several paths, both empirical and theoretical, towards understanding and achieving this level of sample efficiency in learning.
Dr Son Tran (RMIT)
Exploring a Formal Structure of the Common Model of the Intelligent Agent
Designing a model of the decision maker that is substantive and widely held across scientific disciplines is an old quest that promises great benefits but has yielded little results due to difficulties in finding common ground in the design process. Recently, new ideas in reinforcement learning have made substantial progress toward constructing a common model of the intelligent agent by devising a neutral terminology that can be used across disciplines. The neutrality relies on a set of ground rules that seek to describe the process of intelligent decision making in a framework independent of specific application contexts. The framework, however, hasn’t been structured in any formal language amenable to mathematical or computational analysis. This paper discusses and explores the use of probabilistic bigraphs as a possible option that can make three contributions to the process of establishing a science of intelligent decision making independent of biology and engineering application. First, the neutrality of the common model’s terminology and its content are mapped to clearly defined mathematical structures that enable interactive adoption of interdisciplinary ideas. Second, the mapping process enables extension to the common model to be implemented via a standard set of interfaces, thus ensuring communication across disciplines remain robust when the model needs to be adapted for specific application environments. Finally, computational simulation of the model’s dynamics can be carried out to analyze its comparative behaviors in different implementation contexts.
Eric Weiss
TBA
Dr James Whittington (University of Oxford; Stanford University)
Why do neurons look the way they do? From cell-types and modules to mixed selectivity and warping
Neurons in the brain are often finely tuned to specific task variables. Such disentangled representations are highly sought after in the machine learning community. Here we mathematically prove that biological constrains on neurons, namely non-negativity and finite energy resources for activity and weights, enforce neurons to care about single factors of variation, i.e. disentangled representations. We demonstrate these constraints lead to disentangling in a variety of tasks and architectures, including variational autoencoders. We use this theory to explain why the brain partitions its cells into distinct cell types such as grid and object-vector cells, and also explain when the brain instead entangles representations due to entangled tasks. This work tells us populations of neurons encode multiple different variables in discrete modules. It does not, however, tell us how each module should represent its variable of interest, or why grid cells look like grid cells. Thus we also mathematically prove that, if a module is trying to encode physical space, then grid cells are the mathematically optimal representations. This work provides a mathematical understanding of why, when, and how neurons represent factors in both brains and machines, and is a first step on an overall understanding of how task demands structure neural representations.
Jamal Williams (Princeton University)
Novel application of hidden Markov models for naturalistic event processing in high-order cortical areas
Recent fMRI studies of event segmentation have found that default mode regions represent high-level event structure during movie watching. In these regions, neural patterns are relatively stable during events and shift at event boundaries. Music, like narratives, contains hierarchical event structure (e.g., sections are composed of phrases). Here, we tested the hypothesis that brain activity patterns in default mode regions reflect the high-level event structure of music. We used fMRI to record brain activity from 25 participants (male and female) as they listened to a continuous playlist of 16 musical excerpts and additionally collected annotations for these excerpts by asking a separate group of participants to mark when meaningful changes occurred in each one. We then identified temporal boundaries between stable patterns of brain activity using a hidden Markov model and compared the location of the model boundaries to the location of the human annotations. We identified multiple brain regions with significant matches to the observer-identified boundaries, including auditory cortex, medial prefrontal cortex, parietal cortex, and angular gyrus. From these results, we conclude that both higher-order and sensory areas contain information relating to the high-level event structure of music. Moreover, the higher-order areas in this study overlap with areas found in previous studies of event perception in movies and audio narratives, including regions in the default mode network.
Dr Nicolas Zadeh (ULB)
Derivation and numerical analysis of a Fokker-Planck equation describing a population of spiking resonate-and-fire neurons
Based on Izhikhevich’s resonate-and-fire model of membrane potential, we derive a two-dimensional Fokker-Planck equation describing the probability density ρ(v1, v2, t) of finding neurons in a population (subject to excitatory as well as inhibitory inputs) having a potential v1, with a time derivative v2, at a given time t. After discussing the main theoretical difficulties raised by the newfound model, we will study numerical simulations of this partial differential equation (PDE), then conjecture some mathematical properties of the solutions to this problem, in particular the presence of privileged trajectories, attractors and mass conservation or loss. In parallel, and to stay as close as possible to the physical reality of the phenomenon, we build a series of animations of a neuronal network obeying the same rules as before, but without going into the approximations needed to obtain the PDE. This not only serves a didactic purpose (i.e. clearly showing the evolution over time of a population, coming from a given initial condition, for different parameter values), but also allows us to compare it to the PDE simulations mentioned above, thus checking for consistency of the reality depicted by the PDE with the physical phenomenon. We finally try to build a bridge with the clinical side by seeing how parameters of our model could be retrieved when considering an EEG, and by studying the way some drugs could influence the behavior of solutions to the equation.
Amey Zhang (University of Oxford)
Thinking with comics: How the medium can reflect the mind
As researchers, much of the way we communicate about the mind is through linear prose. However, in many ways, this writing - entirely verbal, linearly organized - does not reflect the way we actually think. We think in a jumble of language and image, through fragments and associations, and in patterns particular to a moment, to an individual, or to a species.
While comics are often associated with a genre, they are, like prose or film, a medium. In this talk, I will argue that many of the distinct properties of the comics medium are particularly well-suited to exploring the mind. These properties - such as the combination of image and language, the creation of meaning by juxtaposition, and the development of idiosyncratic symbol systems - can provide tools for both science communication and the active exploration of the mind.
Amey Yun Zhang just finished a MSc in cognitive and evolutionary anthropology at the University of Oxford. She holds a BA from Dartmouth College in ecology, evolutionary biology, and studio art. She is the graduate convener of the Oxford TORCH Comics Network and will be pursuing a PhD at the Rachel Carson Center for environment and society studying how children across cultures learn concepts of the nonhuman.
Liu Zhang (Princeton)
Topology-based Comparison of Neural Population Responses via Persistent Homology and p-Wasserstein Distance
Breakthroughs in neuroscience have prompted research on the manifold view of neural population activity, which poses challenges: the data are high-dimensional; it is unclear which coordinates and metrics for the neural manifold are biologically justified. Topology could address these challenges: topological properties are generalized to high-dimensional surfaces: they are invariant under different coordinates and robust to the choice of metrics. Topological methods have been successfully applied to analyze neural population response. However, these works have not considered how different neural population responses can be appropriately compared.
We develop a topology-based approach and apply it to compare neural population responses in the mouse retina to different visual stimuli. We use nonlinear dimensionality reduction to obtain a lower-dimensional neural manifold of retinal ganglion cell population activity. Topological features are then extracted using persistent homology and represented as persistence diagrams. Finally, we compute the pairwise p-Wasserstein distance between these persistence diagrams. Our experiments show that in terms of topological structures, the neural population response to low-frequency gratings is significantly different from other types of flow stimuli, informing further neuroscientific investigations into this selective preference. Moreover, the p-Wasserstein distance induces a metric space of persistence diagrams where standard statistical objects are well-defined, allowing statistical inference on a distribution of persistence diagrams for the respective neural population responses, such as the expected diagram and the variance over the diagrams. The proposed approach can be used to compare neural population responses arising from a variety of artificial and biological neural networks.
