A plausible neural circuit for classical conditioning without synaptic plasticity

The cellular bases of learning are currently under active investigation by both experimental and theoretical means. In this paper, a simple neuronal wiring diagram is proposed that can reproduce both simple and higher-order behavioral paradigms seen in invertebrate classical conditioning experiments. Learning in this model does not take place by modification of synaptic strength values. Instead, the model uses a layer of interneurons with modifiable thresholds for spike initiation, as suggested by the plasticity mechanisms thought to operate in Hermissenda [Alkon, D. L. (1983) Sci. Am. 249, 70-84]. The model therefore has an advantage in plausibility compared with more standard models using Hebb synapses or their functional equivalents, which have not yet been demonstrated in any invertebrate organism.     

A plausible two-state model for cytochrome c oxidase.

The catalytic properties of pulsed and resting cytochrome c oxidase (ferrocytochrome c: oxygen oxidoreductase, EC 1.9.3.1), expressed in terms of a minimal kinetic scheme and simulated by numerical computations, were successfully described. A two-state model, in which the relative amounts of the enzyme present in each conformation are regulated by the rates of electron flux and O2 binding on one side and the interconversion rates on the other, accounts for the activation of cytochrome c oxidase during turnover.     

Overlapping neural networks for multiple motor engrams.

The hypothesis was tested that learned movement trajectories of different shapes can be stored in, and generated by, largely overlapping neural networks. Indeed, it was possible to train a massively interconnected neural network to generate different shapes of internally stored, dynamically evolving movement trajectories using a general-purpose core part, common to all networks, and a special-purpose part, specific for a particular trajectory. The weights of connections between the core units do not carry any information about trajectories. The core network alone could generate externally instructed trajectories but not internally stored ones, for which both the core and the trajectory-specific part were needed. All information about the movements is stored in the weights of connections between the core part and the specialized units and between the specialized units themselves. Due to these connections the core part reveals specific dynamical behavior for a particular trajectory and, as the result, discriminates different tasks. The percentage of trajectory-specific units needed to generate a certain trajectory was small (2-5%), and the total output of the network is almost entirely provided by the core part, whereas the role of the small specialized parts is to drive the dynamical behavior. These results suggest an efficient and effective mechanism for storing learned motor patterns in, and reproducing them by, overlapping neural networks and are in accord with neurophysiological findings of trajectory-specific cells and with neurological observations of loss of specific motor skills in the presence of otherwise intact motor control.     

Learning in deep neural networks and brains with similarity-weighted interleaved learning

Understanding how the brain learns throughout a lifetime remains a long-standing challenge. In artificial neural networks (ANNs), incorporating novel information too rapidly results in catastrophic interference, i.e., abrupt loss of previously acquired knowledge. Complementary Learning Systems Theory (CLST) suggests that new memories can be gradually integrated into the neocortex by interleaving new memories with existing knowledge. This approach, however, has been assumed to require interleaving all existing knowledge every time something new is learned, which is implausible because it is time-consuming and requires a large amount of data. We show that deep, nonlinear ANNs can learn new information by interleaving only a subset of old items that share substantial representational similarity with the new information. By using such similarity-weighted interleaved learning (SWIL), ANNs can learn new information rapidly with a similar accuracy level and minimal interference, while using a much smaller number of old items presented per epoch (fast and data-efficient). SWIL is shown to work with various standard classification datasets (Fashion-MNIST, CIFAR10, and CIFAR100), deep neural network architectures, and in sequential learning frameworks. We show that data efficiency and speedup in learning new items are increased roughly proportionally to the number of nonoverlapping classes stored in the network, which implies an enormous possible speedup in human brains, which encode a high number of separate categories. Finally, we propose a theoretical model of how SWIL might be implemented in the brain.

Artificial neural networks (ANNs) tend to lose previously acquired knowledge abruptly when new information is incorporated too quickly (“catastrophic interference”) (1, 2). Successful lifelong learners (e.g., humans) do not suffer from this problem, potentially by using mechanisms suggested in the Complementary Learning Systems Theory (CLST) (3) (see also ref. 4). CLST states that the brain relies on complementary learning systems: the hippocampus (HC) for rapid acquisition of new memories and the neocortex (NC) for the gradual incorporation of the new data into context-independent structured knowledge. During “offline periods,” such as sleep and quiet awake rest, the HC triggers replay of recent experiences in the NC, while the NC spontaneously retrieves and interleaves representations of existing classes (5–7). The interleaved replay allows gradual adjustment of NC synaptic weights, in a gradient-descent manner, to create context-independent category representations, thereby gracefully integrating new memories and overcoming catastrophic interference. Numerous studies have since successfully used interleaved replay to achieve lifelong learning in neural networks (8, 9).In practice, however, the CLST raises two significant issues. First, how can the brain possibly perform a comprehensive interleaving when it does not have access to all the old data? One potential solution is “Pseudorehearsal” (10), where random inputs can elicit generative replay of internal representations without requiring explicit access to previously learned examples. Attractor-like dynamics may allow the brain to accomplish pseudorehearsal, but it is unclear what to pseudorehearse. Thus, the second problem is that there is not enough time to interleave all of the previously learned information after each new learning event. “Similarity Weighted Interleaved Learning” (SWIL) was proposed as a solution to this second problem, suggesting that it may be sufficient to interleave only old items with substantial representational similarity to new items (11). Empirical behavioral studies showed that highly consistent new items could be rapidly integrated into NC structured knowledge with little or no interference (12, 13). This indicates that the speed of integrating new information depends on its consistency with the prior knowledge (14). Inspired by this behavioral result, and by a reexamination of the distribution of catastrophic interference among previously acquired classes, which is described below, McClelland et al. (11) demonstrated that SWIL allowed learning new information using 2.5x fewer item presentations per epoch in a simple dataset with two superordinate categories and achieved the same performance as training the network on the entire data. However, the authors did not find a similar effect when using more complex datasets, raising concerns about the algorithm’s scalability.The current study has overcome these limitations by modifying the SWIL algorithm to work with Convolutional Neural Networks (CNNs) on traditional classification datasets (Fashion-MNIST, CIFAR10, and CIFAR100). We exploit the hierarchical structure of existing knowledge to selectively interleave only the old items that have higher representational similarity to new items. With this strategy, we can reach performance levels comparable to that achieved by using the entire training dataset, thereby substantially reducing the amount of data required (data-efficient) and learning time (speedup). We then show that SWIL can also be used in a sequential learning framework. Additionally, we show that learning a new class can be extremely data-efficient—i.e., a much smaller number of old items being presented—if it shares similarities with far fewer previously learned classes, which is likely the case in human learning. Finally, we present a theoretical model of how SWIL might be implemented in the brain using previously stored attractors with an excitability bias proportional to their overlap with new items.     

Interactions of neural networks: models for distraction and concentration.

We present a model of neural group interactions, which are projections from one neural network (network B) of McCulloch-Pitts neurons connected via a Hebbian rule, to another network (network A) of the same structure. We first consider the case in which the projecting network B is in a pattern different from the initial attracting state of network A. A critical projecting strength lambda c is found such that for lambda below this value there exists a noise threshold sigma lambda corresponding to each lambda. For the case where lambda less than lambda c and the noise level sigma less than sigma lambda, there are two possible retrievals, with different probabilities: the initial attracting state of network A and the projecting pattern. If lambda less than lambda c and sigma greater than sigma lambda, stable states of network A disappear. In the case lambda greater than lambda c, network A is pulled out of its initial basin of attraction and into that of the projecting pattern. This analysis provides a model for distraction. Second-order interactions reduce the distraction. When the projecting network B is in the same pattern as the initial attracting state of network A, the projection acts as an external reinforcement, which enables network A to retrieve in highly noisy conditions. Sharp noise thresholds for nonzero retrievals are shown to be eliminated by the projection. Higher-order connectivity improves the retrieval ability of the network. The second case serves as a model of concentration. We discuss the model of distraction and concentration (i) in connection with common experience of expectation of recognition and (ii) in connection with recent T-maze experiments on infant rats; finally, we suggest a refined version of the Bruner-Potter experiment to test our prediction of the disappearance of hysteresis.     

Sequential state generation by model neural networks.

Sequential patterns of neural output activity form the basis of many biological processes, such as the cyclic pattern of outputs that control locomotion. I show how such sequences can be generated by a class of model neural networks that make defined sets of transitions between selected memory states. Sequence-generating networks depend upon the interplay between two sets of synaptic connections. One set acts to stabilize the network in its current memory state, while the second set, whose action is delayed in time, causes the network to make specified transitions between the memories. The dynamic properties of these networks are described in terms of motion along an energy surface. The performance of the networks, both with intact connections and with noisy or missing connections, is illustrated by numerical examples. In addition, I present a scheme for the recognition of externally generated sequences by these networks.     

Chemical implementation of neural networks and Turing machines.

We propose a reversible reaction mechanism with a single stationary state in which certain concentrations assume either high or low values dependent on the concentration of a catalyst. The properties of this mechanism are those of a McCulloch-Pitts neuron. We suggest a mechanism of interneuronal connections in which the stationary state of a chemical neuron is determined by the state of other neurons in a homogeneous chemical system and is thus a "hardware" chemical implementation of neural networks. Specific connections are determined for the construction of logic gates: AND, NOR, etc. Neural networks may be constructed in which the flow of time is continuous and computations are achieved by the attainment of a stationary state of the entire chemical reaction system, or in which the flow of time is discretized by an oscillatory reaction. In another article, we will give a chemical implementation of finite state machines and stack memories, with which in principle the construction of a universal Turing machine is possible.     

Chemical implementation and thermodynamics of collective neural networks.

The chemical implementation of a neuron and connections among neurons described in prior work is used to construct collective neural networks. With stated approximations, these chemical networks are reduced to networks of the Hopfield type. Chemical networks approaching a stationary or equilibrium state provide a Liapunov function with the same extremal properties as Hopfield's energy function. Numerical comparisons of chemical and Hopfield networks with small numbers (2-16) of neurons show agreement on the results of given computations.     

Self-organized phase transitions in neural networks as a neural mechanism of information processing.

Transitions between dynamically stable activity patterns imposed on an associative neural network are shown to be induced by self-organized infinitesimal changes in synaptic connection strength and to be a kind of phase transition. A key event for the neural process of information processing in a population coding scheme is transition between the activity patterns encoding usual entities. We propose that the infinitesimal and short-term synaptic changes based on the Hebbian learning rule are the driving force for the transition. The phase transition between the following two dynamical stable states is studied in detail, the state where the firing pattern is changed temporally so as to itinerate among several patterns and the state where the firing pattern is fixed to one of several patterns. The phase transition from the pattern itinerant state to a pattern fixed state may be induced by the Hebbian learning process under a weak input relevant to the fixed pattern. The reverse transition may be induced by the Hebbian unlearning process without input. The former transition is considered as recognition of the input stimulus, while the latter is considered as clearing of the used input data to get ready for new input. To ensure that information processing based on the phase transition can be made by the infinitesimal and short-term synaptic changes, it is absolutely necessary that the network always stays near the critical state corresponding to the phase transition point.     

Synchronous neural networks of nonlinear threshold elements with hysteresis.

We use Hoffmann's suggestion [Hoffmann, G. W. (1986) J. Theor. Biol. 122, 33-67] of hysteresis in a single neuron level and determine its consequences in a synchronous network made of such neurons. We show that the overall retrieval ability in the presence of noise and the memory capacity of the network in the present model are better than in conventional models without such hysteresis. Second-order interaction further improves the retrieval ability of the network and causes hysteresis in the retrieval-noise curve for any arbitrary width of the bistable region. The convergence rate is increased by the hysteresis at high noise levels but is reduced by the hysteresis at low noise levels. Explicit formulae are given for calculations of average final convergence and noise threshold as functions of the width of the bistable region. There is neurophysiological evidence for hysteresis in single neurons, and we propose optical implementations of the present model by using ZnSe interference filters to test the predictions of the theory.     

A plausible model for reversal of neoplastic transformations in plants based on multiple steady states.

We offer a plausible interpretation of some experiments on the reversal of neoplastic transformations in plants. We suggest that normal cells and tumorous cells represent multiple stable-steady states corresponding to a reaction feedback mechanism. The (autocatalytic) feedback loop is constructed from observations on the role played by myo-inositol: it increases the permeability of ions through the membrane and the biosynthetic pathway to myo-inositol is activated by ions. Provided that the permeabilities of nutrients (sugars and salts) are a product-enhanced function of myo-inositol, then we have a (oversimplified) model that can exhibit multiple stationary stable states, one or two depending on the exogenous nutrients and myo-inositol concentrations, and reversible and irreversible transitions from one of these states to the other are possible. From this model, straightforward simple experiments are suggested. We also propose that recent models dealing with the intracellular calcium regulation by hormones, where one key step requires the hydrolysis of inositol phospholipids, take into account free myo-inositol and endogenous hormone concentrations (e.g., auxins).     

A clinical prediction rule for pulmonary tuberculosis in emergency departments.

SETTING: University-affiliated hospital located in an area with a high incidence of pulmonary tuberculosis (PTB). OBJECTIVE: To develop a clinical prediction rule (CPR) based on information obtainable on admission, to permit rapid identification of patients with PTB. DESIGN: Information from patients with respiratory symptoms who attended the emergency department of Cayetano Heredia Hospital, Lima, Peru, was collected prospectively. Clinical symptoms, past medical history, demographic data and results of chest X-rays (CXRs), sputum smear and culture in L?wenstein-Jensen media were obtained. Based on logistic regression, we constructed a scoring system to predict PTB. RESULTS: A total of 345 patients were enrolled in the study, including 109 (31%) culture-proven PTB cases. In logistic regression analysis, we found age, previous history of PTB, weight loss, presence of cavities, upper lobe infiltrate and miliary pattern on CXR as independent predictors of PTB. We designed a scoring system with these variables, taking into account their statistical weight. The score attained 93% sensitivity and 42% specificity. CONCLUSION: The CPR that was developed performed well in our population. It merits further validation in other settings. It should not, however, replace, but should complement sputum microscopy when deciding on isolation, and it does not preclude microbiology in making a definitive diagnosis.     

A prediction model based on artificial neural networks for the diagnosis of obstructive sleep apnea

Background

Recently, artificial neural networks (ANNs) have been widely applied in science, engineering, and medicine. In the present study, we evaluated the ability of artificial neural networks to be used as a computer program and assistant tool in the diagnosis of obstructive sleep apnea (OSA). Our hypothesis was that ANNs could use clinical information to precisely predict cases of OSA.

Method

The study population in this clinical trial consisted of 201 patients with suspected OSA (140 with a positive diagnosis of OSA and 61 with a negative diagnosis of OSA). The artificial neural network was trained by assessing five clinical variables from 201 patients; efficiency was then estimated in this group of 201 patients. The patients were classified using a five-element input vector. ANN classifiers were assessed with the multilayer perceptron (MLP) networks.

Results

Use of the MLP classifiers resulted in a diagnostic accuracy of 86.6 %, which in clinical practice is high enough to reduce the number of patients evaluated by polysomnography (PSG), an expensive and limited diagnostic resource.

Conclusions

By establishing a pattern that allows the recognition of OSA, ANNs can be used to identify patients requiring PSG.

     

A prognostic rule for elderly patients admitted with community-acquired pneumonia.

PURPOSE: We sought to identify admission characteristics predicting mortality in elderly patients hospitalized with community-acquired pneumonia and to develop a prognostic staging system and discriminant rule. PATIENTS AND METHODS: We retrospectively analyzed data from 2,356 patients aged > or = 65 years admitted with community-acquired pneumonia. Multivariable analyses of a derivation cohort (n = 1,000) identified characteristics associated with hospital mortality. A staging system and discriminant rule based on these characteristics were tested in a validation cohort (n = 1,356). Our discriminant rule was compared with a rule formulated from a heterogeneous adult population with community-acquired pneumonia. RESULTS: Hospital mortality rates were 9% (derivation cohort) and 12% (validation cohort). We identified five independent predictors of mortality: age > or = 85 years [odds ratio 1.8 (95% confidence interval 1.1-3.1)], comorbid disease [odds ratio 4.1 (2.1-8.1)], impaired motor response [odds ratio 2.3 (1.4-3.7)], vital sign abnormality [odds ratio 3.4 (2.1-5.4)], and creatinine level > or = 1.5 mg/dL [odds ratio 2.5 (1.5-4.2)]. These variables stratified patients into four distinct stages with increasing mortality in the derivation cohort (Stage 1, 2%; Stage 2, 7%; Stage 3, 22%; Stage 4, 45%; P = 0.001) as well as in the validation cohort (Stage 1, 4%; Stage 2, 11%; Stage 3, 23%; Stage 4, 41%; P = 0.001). The discriminant rule developed from the derivation cohort had greater overall accuracy (77.1%) in the validation cohort than a rule formulated from a heterogeneous adult population (68.0%, P = 0.001). CONCLUSION: Elderly patients with community-acquired pneumonia have characteristics at admission that can predict mortality. Our staging system and discriminant rule improve prognostic stratification of these patients.     

From the Cover: A geometric constraint, the head-to-tail exclusion rule, may be the basis for the isolated-pentagon rule in fullerenes with more than 60 vertices

Carbon atoms self-assemble into the famous soccer-ball shaped Buckminsterfullerene (C₆₀), the smallest fullerene cage that obeys the isolated-pentagon rule (IPR). Carbon atoms self-assemble into larger (n > 60 vertices) empty cages as well—but only the few that obey the IPR—and at least 1 small fullerene (n ≤ 60) with adjacent pentagons. Clathrin protein also self-assembles into small fullerene cages with adjacent pentagons, but just a few of those. We asked why carbon atoms and clathrin proteins self-assembled into just those IPR and small cage isomers. In answer, we described a geometric constraint—the head-to-tail exclusion rule—that permits self-assembly of just the following fullerene cages: among the 5,769 possible small cages (n ≤ 60 vertices) with adjacent pentagons, only 15; the soccer ball (n = 60); and among the 216,739 large cages with 60 < n ≤ 84 vertices, only the 50 IPR ones. The last finding was a complete surprise. Here, by showing that the largest permitted fullerene with adjacent pentagons is one with 60 vertices and a ring of interleaved hexagons and pentagon pairs, we prove that for all n > 60, the head-to-tail exclusion rule permits only (and all) fullerene cages and nanotubes that obey the IPR. We therefore suggest that self-assembly that obeys the IPR may be explained by the head-to-tail exclusion rule, a geometric constraint.     

A plausible model for the digital response of p53 to DNA damage

Recent observations show that the single-cell response of p53 to ionizing radiation (IR) is "digital" in that it is the number of oscillations rather than the amplitude of p53 that shows dependence on the radiation dose. We present a model of this phenomenon. In our model, double-strand break (DSB) sites induced by IR interact with a limiting pool of DNA repair proteins, forming DSB-protein complexes at DNA damage foci. The persisting complexes are sensed by ataxia telangiectasia mutated (ATM), a protein kinase that activates p53 once it is phosphorylated by DNA damage. The ATM-sensing module switches on or off the downstream p53 oscillator, consisting of a feedback loop formed by p53 and its negative regulator, Mdm2. In agreement with experiments, our simulations show that by assuming stochasticity in the initial number of DSBs and the DNA repair process, p53 and Mdm2 exhibit a coordinated oscillatory dynamics upon IR stimulation in single cells, with a stochastic number of oscillations whose mean increases with IR dose. The damped oscillations previously observed in cell populations can be explained as the aggregate behavior of single cells.