How sensory representations in the sensory cortex are dynamically and rapidly modulated to support flexible, goal-directed behavior remains a fundamental open question. Using rare simultaneous single-neuron recordings across multiple human brain regions, including the ventral temporal cortex (VTC), a high-level visual area not traditionally associated with context-dependent coding, we show that visual representations in VTC are rapidly reconfigured by the presence or absence of preceding verbal task instructions. Identical visual stimuli evoked distinct responses in VTC depending on whether task instructions were provided or not, with categorical representations sharpened when goals were specified in advance. In contrast, dorsal anterior cingulate cortex (dACC) and the hippocampus carried strong signals related to instruction timing, consistent with known regional specialization in top-down control. Coupling between dACC and VTC increased during periods of instructional uncertainty and high cognitive demand on a rapid timescale during stimulus presentation, and its strength predicted correct performance. Together, these findings uncover a rapid, online feedback mechanism through which the medial frontal cortex dynamically reorganizes sensory population codes in VTC, linking flexible cortical coordination to successful goal-directed computation.

Studies of human decision-making demonstrate that environmental regularities, such as natural image statistics or intentionally nonuniform stimulus probabilities, can be exploited to improve efficiency (termed `efficient-coding'). Conversely, from a machine learning perspective, such nonuniform stimulus properties can lead to biased neural networks with poor generalization performance. Understanding how the brain flexibly leverages stimulus bias while maintaining robust generalization could lead to novel architectures that adaptively exploit environmental structure without sacrificing performance on out-of-distribution data. To address this disconnect, we investigated the impact of stimulus regularities in a 3-layer hierarchical continuous-time recurrent neural network (ctRNN) to better understand how artificial networks might exploit statistical regularities to improve efficiency while avoiding undesirable biases. We trained the model to reproduce one of six possible inputs under biased conditions (stimulus 1 more probable than stimuli 2-6) or unbiased conditions (all stimuli equally likely). Across all hidden layers, more information was encoded about high-probability stimuli, consistent with the efficient-coding framework. Importantly, reducing feedback from the final hidden layer of trained models selectively magnified representations of high-probability stimuli, at the expense of low-probability stimuli, across all layers. Together, these results suggest that models exploit nonuniform input statistics to improve efficiency, and that feedback pathways evolve to protect the processing of low-probability stimuli by regulating the impact of biased input statistics.

Working memory (WM) enables temporary retention of information essential for flexible cognition. Although persistent population activity has long been regarded as a principal mechanism of memory maintenance, continuous single-neuron firing is energetically demanding and difficult to reconcile with the heterogeneous firing properties of cortical neurons. Applying single-trial analyses to 902 neurons recorded from 21 neurosurgical patients performing a WM task, we found that maintenance was supported by transient, burst-like episodes of coordinated activity rather than sustained firing. Cross-temporal decoding exhibited localized generalization, and decoding accuracy increased with wider temporal windows, indicating that apparent persistence can emerge from temporally interleaved activity across neurons. We further developed a feature-based, putative cell-type classifier that revealed distinct circuit contributions: pyramidal neurons expressed content in burst-aligned events during maintenance, whereas interneurons were strongly modulated by memory load and behavior. Together, these findings reconcile dynamic and persistent accounts, indicating that human WM can emerge from temporally interleaved, cell-type-specific dynamics that provide a flexible and potentially metabolically efficient substrate for maintaining information over time. ### Competing Interest Statement The authors have declared no competing interest. Air Force Office of Scientific Research American Academy of Neurology (AAN) Resident Research Scholarship

Recurrent neural networks (RNNs) based on model neurons that communicate via continuous signals have been widely used to study how cortical neural circuits perform cognitive tasks. Training such networks to perform tasks that require information maintenance over a brief period (i.e., working memory tasks) remains a challenge. Inspired by the robust information maintenance observed in higher cortical areas such as the prefrontal cortex, despite substantial inherent noise, we investigated the effects of random noise on RNNs across different cognitive functions, including working memory. Our findings reveal that random noise not only speeds up training but also enhances the stability and performance of RNNs on working memory tasks. Importantly, this robust working memory performance induced by random noise during training is attributed to an increase in synaptic decay time constants of inhibitory units, resulting in slower decay of stimulus-specific activity critical for memory maintenance. Our study reveals the critical role of noise in shaping neural dynamics and cognitive functions, suggesting that inherent variability may be a fundamental feature driving the specialization of inhibitory neurons to support stable information processing in higher cortical regions.

Supplementary Figure 1 from JAGGED1 Expression Is Associated with Prostate Cancer Metastasis and Recurrence

Recent studies have proposed employing biologically plausible recurrent neural networks (RNNs) to investigate flexible decision-making in the brain. However, the mechanisms underlying the integration of bottom-up sensory inputs and temporally varying top-down factors (such as task instructions and selective attention) remain poorly understood, both within the context of these models and the brain. To address this knowledge gap, we trained biologically inspired RNNs on complex cognitive tasks that require adaptive integration of these factors. Through comprehensive analyses of RNNs and neural activity from mouse primary visual cortex, we show that sensory neurons in low-level areas possess the remarkable ability to multiplex and dynamically combine both bottom-up and top-down information via local inhibitory-to-inhibitory connections. Our results shed light on the role of disinhibitory circuits in the intricate interplay between bottom-up and top-down factors to enable flexible decision processes.### Competing Interest StatementThe authors have declared no competing interest.

Recent studies have proposed employing biologically plausible recurrent neural networks (RNNs) to investigate flexible decision-making in the brain. However, the mechanisms underlying the integration of bottom-up sensory inputs and temporally varying top-down factors (such as task instructions and selective attention) remain poorly understood, both within the context of these models and the brain. To address this knowledge gap, we trained biologically inspired RNNs on complex cognitive tasks that require adaptive integration of these factors. Through comprehensive analyses of RNNs and neural activity from mouse primary visual cortex, we show that sensory neurons in low-level areas possess the remarkable ability to multiplex and dynamically combine both bottom-up and top-down information via local inhibitory-to-inhibitory connections. Our results shed light on the role of disinhibitory circuits in the intricate interplay between bottom-up and top-down factors to enable flexible decision processes.

Visual inputs are often highly structured, and statistical regularities of these signals can be used to guide future visuomotor associations and thus optimize behavior. Through a recurrent neural network (RNN) model, human psychophysics, and electroencephalography (EEG), we probed the neural mechanisms for processing probabilistic structures of visual signals to guide behavior. We first constructed and trained a biophysically constrained RNN model to perform a series of probabilistic visual discrimination tasks similar to paradigms designed for humans. Specifically, the training environment was probabilistic such that one stimulus was more probable than the others. We showed that both humans and RNNs successfully learned the stimulus probability and integrated this knowledge into their decisions and task strategy in a new environment. Performance of both humans and RNNs varied with the degree to which the stimulus probability of the new environment matched the formed expectation. In both cases, this expectation effect was more prominent when the strength of sensory evidence was low, suggesting that like humans, the RNNs placed more emphasis on prior expectation (top-down signals) when the available sensory information (bottom-up signals) was limited, thereby optimizing task performance. By dissecting the trained RNNs, we demonstrated how competitive inhibition and recurrent excitation form the basis for neural circuitry optimized to perform probabilistic visual processing.

Recurrent neural networks (RNNs) based on model neurons that communicate via continuous signals have been widely used to study how cortical neurons perform cognitive tasks. Training such networks to perform tasks that require information maintenance over a brief period (i.e., working memory tasks) remains a challenge. Critically, the training process becomes difficult when the synaptic decay time constant is not fixed to a large constant number for all the model neurons. We hypothesize that the brain utilizes intrinsic cortical noise to generate a reservoir of heterogeneous synaptic decay time constants optimal for maintaining information. Here, we show that introducing random, internal noise to the RNNs not only speeds up the training but also produces stable models that can maintain information longer than the RNNs trained without internal noise. Importantly, this robust working memory performance induced by incorporation of internal noise during training is attributed to an increase in synaptic decay time constants of a sub-population of inhibitory units.

Recurrent neural network (RNN) models trained to perform cognitive tasks are a useful computational tool for understanding how cortical circuits execute complex computations. However, these models are often composed of units that interact with one another using continuous signals and overlook parameters intrinsic to spiking neurons. Here, we developed a method to directly train not only synaptic-related variables but also membrane-related parameters of a spiking RNN model. Training our model on a wide range of cognitive tasks resulted in diverse yet task-specific synaptic and membrane parameters. We also show that fast membrane time constants and slow synaptic decay dynamics naturally emerge from our model when it is trained on tasks associated with working memory (WM). Further dissecting the optimized parameters revealed that fast membrane properties are important for encoding stimuli, and slow synaptic dynamics are needed for WM maintenance. This approach offers a unique window into how connectivity patterns and intrinsic neuronal properties contribute to complex dynamics in neural populations.

Purpose: Geographic atrophy (GA), a late stage of age-related macular degeneration (AMD), is a major cause of blindness. Even while central visual acuity remains relatively well preserved, GA often causes considerable compromise of visual function and quality of life. No treatment currently exists. We evaluated the safety and efficacy of pegcetacoplan, a complement C3 inhibitor, for treatment of GA. Design: Prospective, multicenter, randomized, sham-controlled phase 2 study. Participants: Two hundred forty-six patients with GA. Methods: Patients with GA were assigned randomly in a 2:2:1:1 ratio to receive intravitreal injections of 15 mg pegcetacoplan monthly or every other month (EOM) or sham intravitreal injections monthly or EOM for 12 months with follow-up at months 15 and 18. Area and growth of GA were measured using fundus autofluorescence imaging. Main Outcome Measures: The primary efficacy end point was mean change in square root GA lesion area from baseline to month 12. Secondary outcome measures included mean change from baseline in GA lesion area without the square root transformation, distance of GA lesion from the fovea, best-corrected visual acuity (BCVA), low-luminance BCVA, and low-luminance visual acuity deficit. The primary safety end point was the number and severity of treatment-emergent adverse events. Results: In patients receiving pegcetacoplan monthly or EOM, the GA growth rate was reduced by 29% (95% confidence interval [CI], 9-49; P = 0.008) and 20% (95% CI, 0-40; P = 0.067) compared with the sham treatment group. Post hoc analysis showed that the effect was greater in the second 6 months of treatment, with observed reductions of 45% (P = 0.0004) and 33% (P = 0.009) for pegcetacoplan monthly and EOM, respectively. Two cases of culture-positive endophthalmitis and 1 case of culture-negative endophthalmitis occurred in the pegcetacoplan monthly group. New-onset investigator-determined exudative AMD was reported more frequently in pegcetacoplan-treated eyes (18/86 eyes [20.9%] and 7/79 eyes [8.9%] in monthly and EOM groups, respectively) than in sham-treated eyes (1/81 eyes [1.2%]). Conclusions: Local C3 inhibition with pegcetacoplan resulted in statistically significant reductions in the growth of GA compared with sham treatment. Phase 3 studies will define the efficacy and safety profile further. (C) 2019 by the American Academy of Ophthalmology.

Cortical neurons process information on multiple timescales, and areas important for working memory (WM) contain neurons capable of integrating information over a long timescale. However, the underlying mechanisms for the emergence of neuronal timescales stable enough to support WM are unclear. By analyzing a spiking recurrent neural network model trained on a WM task and activity of single neurons in the primate prefrontal cortex, we show that the temporal properties of our model and the neural data are remarkably similar. Dissecting our recurrent neural network model revealed strong inhibitory-to-inhibitory connections underlying a disinhibitory microcircuit as a critical component for long neuronal timescales and WM maintenance. We also found that enhancing inhibitory-to-inhibitory connections led to more stable temporal dynamics and improved task performance. Finally, we show that a network with such microcircuitry can perform other tasks without disrupting its pre-existing timescale architecture, suggesting that strong inhibitory signaling underlies a flexible WM network.

Geographic atrophy (GA), a late stage of age-related macular degeneration (AMD), is a major cause of blindness. Even while central visual acuity remains relatively well preserved, GA often causes considerable compromise of visual function and quality of life. No treatment currently exists. We evaluated the safety and efficacy of pegcetacoplan, a complement C3 inhibitor, for treatment of GA.Prospective, multicenter, randomized, sham-controlled phase 2 study.Two hundred forty-six patients with GA.Patients with GA were assigned randomly in a 2:2:1:1 ratio to receive intravitreal injections of 15 mg pegcetacoplan monthly or every other month (EOM) or sham intravitreal injections monthly or EOM for 12 months with follow-up at months 15 and 18. Area and growth of GA were measured using fundus autofluorescence imaging.The primary efficacy end point was mean change in square root GA lesion area from baseline to month 12. Secondary outcome measures included mean change from baseline in GA lesion area without the square root transformation, distance of GA lesion from the fovea, best-corrected visual acuity (BCVA), low-luminance BCVA, and low-luminance visual acuity deficit. The primary safety end point was the number and severity of treatment-emergent adverse events.In patients receiving pegcetacoplan monthly or EOM, the GA growth rate was reduced by 29% (95% confidence interval [CI], 9-49; P = 0.008) and 20% (95% CI, 0-40; P = 0.067) compared with the sham treatment group. Post hoc analysis showed that the effect was greater in the second 6 months of treatment, with observed reductions of 45% (P = 0.0004) and 33% (P = 0.009) for pegcetacoplan monthly and EOM, respectively. Two cases of culture-positive endophthalmitis and 1 case of culture-negative endophthalmitis occurred in the pegcetacoplan monthly group. New-onset investigator-determined exudative AMD was reported more frequently in pegcetacoplan-treated eyes (18/86 eyes [20.9%] and 7/79 eyes [8.9%] in monthly and EOM groups, respectively) than in sham-treated eyes (1/81 eyes [1.2%]).Local C3 inhibition with pegcetacoplan resulted in statistically significant reductions in the growth of GA compared with sham treatment. Phase 3 studies will define the efficacy and safety profile further.

Author(s): Kim, Robert | Advisor(s): Sejnowski, Terrence J | Abstract: Schizophrenia is a complex neuropsychiatric disorder characterized by a wide range of clinical manifestations. Even though the etiology of schizophrenia is not known, the heterogeneous nature of the disorder strongly suggests that multiple pathways and brain areas are affected by a combination of internal and external factors. These factors include and not limited to: psychological, genetic, social, and environmental determinants. Cognitive impairment is one of the commonly observed clinical manifestations of schizophrenia. Working memory, which is an ability to encode and hold information over a short period, is severely impaired in schizophrenia. (1) Characterizing how such deficits manifest in large-scale dynamics and (2) understanding the pathophysiology and circuit mechanisms behind working memory deficits associated with schizophrenia are the two main questions that I address in my dissertation.To answer the first question, I employed a method based on nonlinear systems theory to quantify large-scale dynamical states of time-series data and to identify dynamically distinct subgroups (Chapter 2). I demonstrate that the method, which utilizes delay differential analysis (DDA), can effectively extract features reflective of significant state changes and detect subgroups with similar features. Applying the method to brain signals obtained from a large cohort of schizophrenia patients further revealed subgroups with distinct dynamical characteristics aligned with neurophysiological and clinical parameters.To answer the second question, I first developed a biologically realistic computational model based on spiking recurrent neural networks (RNNs) capable of learning cognitive tasks that involve working memory (Chapter 3). By taking advantage of a close relationship between continuous and spike RNNs that emerges under certain conditions, the method provides an extremely simple platform that can be utilized to investigate how power-efficient network dynamics lead to complex cognitive computations. By employing the framework, I uncover and characterize important circuit mechanisms critical for working memory maintenance in Chapter 4. The uncovered microcircuitry underscores the importance of disinhibitory gating exerted by specific subtypes of inhibitory interneurons, further confirming recent experimental findings.Overall, my dissertation provides important computational tools for probing both micro- and macro-scale circuit dynamics associated with cognitive deficits in schizophrenia.

Natural systems, including the brain, often seem chaotic, since they are typically driven by complex nonlinear dynamical processes. Disruption in the fluid coordination of multiple brain regions contributes to impairments in information processing and the constellation of symptoms observed in neuropsychiatric disorders. Schizophrenia (SZ), one of the most debilitating mental illnesses, is thought to arise, in part, from such a network dysfunction, leading to impaired auditory information processing as well as cognitive and psychosocial deficits. Current approaches to neurophysiologic biomarker analyses predominantly rely on linear methods and may, therefore, fail to capture the wealth of information contained in whole EEG signals, including nonlinear dynamics. In this study, delay differential analysis (DDA), a nonlinear method based on embedding theory from theoretical physics, was applied to EEG recordings from 877 SZ patients and 753 nonpsychiatric comparison subjects (NCSs) who underwent mismatch negativity (MMN) testing via their participation in the Consortium on the Genetics of Schizophrenia (COGS-2) study. DDA revealed significant nonlinear dynamical architecture related to auditory information processing in both groups. Importantly, significant DDA changes preceded those observed with traditional linear methods. Marked abnormalities in both linear and nonlinear features were detected in SZ patients. These results illustrate the benefits of nonlinear analysis of brain signals and underscore the need for future studies to investigate the relationship between DDA features and pathophysiology of information processing.

Cortical neurons process and integrate information on multiple timescales. In addition, these timescales or temporal receptive fields display functional and hierarchical organization. For instance, areas important for working memory (WM), such as prefrontal cortex, utilize neurons with stable temporal receptive fields and long timescales to support reliable representations of stimuli. Despite of the recent advances in experimental techniques, the underlying mechanisms for the emergence of neuronal timescales long enough to support WM are unclear and challenging to investigate experimentally. Here, we demonstrate that spiking recurrent neural networks (RNNs) designed to perform a WM task reproduce previously observed experimental findings and that these models could be utilized in the future to study how neuronal timescales specific to WM emerge.