A model of visual-spatial memory across saccades.

This paper describes a neural network model that directs saccades back to targets after they disappear and other saccades intervene. This is a simple example of knowing where something is after it is no longer visible and the observer has moved. These tasks require a short-term memory that can store continuous values of spatial location. The model was generated by training a neural network with a recurrently connected hidden layer to specify memory-guided saccades. The trained network maintains stored locations accurately for a few seconds. It uses a leaky integrator mechanism in which there is a slow decay of the stored value to a small number of fixed point attractors. Similar mechanisms have been used to model oculomotor integration (Cannon, S., Robinson, D., & Shamma, S. (1983). A proposed neural network for the integrator of the oculomotor system. Biological Cybernetics, 49, 127-136; Seung, H. (1998). Continuous attractors and oculomotor control. Neural Networks, 11, 1253-1258). The mechanism is robust to parameters such as the input and output format and the constraints in training. However, the receptive field properties of the hidden units do depend on these parameters. It was possible to find biologically plausible parameters that produced hidden unit behavior similar to that of real neurons involved in saccade memory. In particular, training the model to simultaneously represent the target location in both eye- and head-based reference frames produces units similar to neurons in parietal saccade areas.

Neural network models of bilateral coordination.

Two mechanisms are described for controlling the movement of a pair of arms. The first is an engineered motion planner that finds solutions to the ill-posed problem of making noncolliding, goal-directed movements. The second uses neural networks that learn to emulate the coordinated behaviors of the motion planner using considerably less computational resources. Analysis of the networks shows in general terms how they work, and allows us to make testable predictions about some of the response properties that might be observed in the brain systems serving bilateral coordination.

A model of reaching dynamics in primary motor cortex.

Features of virtually all voluntary movements are represented in the primary motor cortex. The movements can be ongoing, imminent, delayed, or imagined. Our goal was to investigate the dynamics of movement representation in the motor cortex. To do this we trained a fully recurrent neural network to continually output the direction and magnitude of movements required to reach randomly changing targets. Model neurons developed preferred directions and other properties similar to real motor cortical neurons. The key finding is that when the target for a reaching movement changes location, the ensemble representation of the movement changes nearly monotonically, and the individual neurons comprising the representation exhibit strong, nonmonotonic transients. These transients serve as internal recurrent signals that force the ensemble representation to change more rapidly than if it were limited by the time constants of individual neurons. These transients, if they exist, could be observed in experiments that require only slight modifications of the standard paradigm used to investigate movement representation in the motor cortex.

Analysis of direction selectivity arising from recurrent cortical interactions.

The relative contributions of feedforward and recurrent connectivity to the direction-selective responses of cells in layer IVB of primary visual cortex are currently the subject of debate in the neuroscience community. Recently, biophysically detailed simulations have shown that realistic direction-selective responses can be achieved via recurrent cortical interactions between cells with nondirection-selective feedforward input (Suarez et al., 1995; Maex & Orban, 1996). Unfortunately these models, while desirable for detailed comparison with biology, are complex and thus difficult to analyze mathematically. In this article, a relatively simple cortical dynamical model is used to analyze the emergence of direction-selective responses via recurrent interactions. A comparison between a model based on our analysis and physiological data is presented. The approach also allows analysis of the recurrently propagated signal, revealing the predictive nature of the implementation.

A model that accounts for activity in primate frontal cortex during a delayed matching-to-sample task.

A fully recurrent neural network model was optimized to perform a spatial delayed matching-to-sample task (DMS). In DMS, a stimulus is presented at a sample location, and a match is reported when a subsequent stimulus appears at that location. Stimuli elsewhere are ignored. Computationally, a DMS system could consist of memory and comparison components. The model, although not constrained to do so, worked by using two corresponding classes of neurons in the hidden layer: storage and comparator units. Storage units form a dynamical system with one fixed point attractor for each sample location. Comparator units constitute a system receiving input from these storage units as well as from current input stimuli. Both unit types were tuned directionally. These two sources of information combine to create unique patterns of activity that determine whether a match has occurred. In networks with abundant hidden units, the storage and comparator functions were distributed so that individual units took part in both. We compared the model with single-neuron recordings from premotor (PM) and prefrontal (PF) cortex. As shown previously, many PM and PF neurons behaved like storage units. In addition, both regions contain neurons that behave like the comparator units of the model and appear to have dual functionality similar to that observed in the model units. No neuron in either area had properties identical to those of the match output neuron of the model. However, four PF neurons and one PM neuron resembled the output signal more closely than any of the hidden units of the model.

A spiking network model of short-term active memory.

Studies of cortical neurons in monkeys performing short-term memory tasks have shown that information about a stimulus can be maintained by persistent neuron firing for periods of many seconds after removal of the stimulus. The mechanism by which this sustained activity is initiated and maintained is unknown. In this article we present a spiking neural network model of short-term memory and use it to investigate the hypothesis that recurrent, or "re-entrant," networks with constant connection strengths are sufficient to store graded information temporarily. The synaptic weights that enable the network to mimic the input-output characteristics of an active memory module are computed using an optimization procedure for recurrent networks with non-spiking neurons. This network is then transformed into one with spiking neurons by interpreting the continuous output values of the nonspiking model neurons as spiking probabilities. The behavior of the model neurons in this spiking network is compared with that of 179 single units previously recorded in monkey inferotemporal (IT) cortex during the performance of a short-term memory task. The spiking patterns of almost every model neuron are found to resemble closely those of IT neurons. About 40% of the IT neuron firing patterns are also found to be of the same types as those of model neurons. A property of the spiking model is that the neurons cannot maintain precise graded activity levels indefinitely, but eventually relax to one of a few constant activities called fixed-point attractors. The noise introduced into the model by the randomness of spiking causes the network to jump between these attractors. This switching between attractor states generates spike trains with a characteristic statistical temporal structure. We found evidence for the same kind of structure in the spike trains from about half of the IT neurons in our test set. These results show that the behavior of many real cortical memory neurons is consistent with an active storage mechanism based on recurrent activity in networks with fixed synaptic strengths.

Identification models of the nervous system.

It has been widely observed that when artificial neural networks are trained by supervised learning to do computations that also occur in the nervous system, the behavior of the model neurons often closely resembles that of the real neurons involved in the task. It is not immediately clear why this should be the case or what use can be made of models generated by supervised learning. Here, recent developments are reviewed and analysed in an attempt to clarify these issues. This analysis is facilitated by treating supervised learning models of the brain as a special case of system identification, a general and well-studied modeling paradigm. The neural systems identification paradigm provides a systematic way to generate realistic models starting with a high-level description of a hypothesized computation and some architectural and physiological constraints about the area being modeled. There is no inherent limitation to the realism that can be incorporated into identification models. This approach eliminates the need to find neural implementation algorithms by ad hoc means and provides neuroscientists with a convenient way to build models that account for observed data.

The role of the teacher in learning-based models of parietal area 7a.

The back-propagation learning procedure can be used to train simulated neural networks to compute arbitrary functions. We have recently shown that when such a network is trained to carry out the transformation of stimulus location to head-centered coordinates that occurs in parietal area 7a, the response properties of certain units in the network closely resemble neurons found in area 7a. The back-propagation procedure requires the use of a teacher. Here we examine the effect of using different kinds of teachers. As long as the teacher represents information about stimulus location in head-centered coordinates, the trained network contains units of the kind found in area 7a. Differences in teacher format only effect the quantitative distribution of the different unit types. When the teacher does not represent stimulus location explicitly, the network does not contain units of the required kind.

Learning the hidden structure of speech.

In the work described here, the backpropagation neural network learning procedure is applied to the analysis and recognition of speech. This procedure takes a set of input/output pattern pairs and attempts to learn their functional relationship; it develops the necessary representational features during the course of learning. A series of computer simulation studies was carried out to assess the ability of these networks to accurately label sounds, to learn to recognize sounds without labels, and to learn feature representations of continuous speech. These studies demonstrated that the networks can learn to label presegmented test tokens with accuracies of up to 95%. Networks trained on segmented sounds using a strategy that requires no external labels were able to recognize and delineate sounds in continuous speech. These networks developed rich internal representations that included units which corresponded to such traditional distinctions as vowels and consonants, as well as units that were sensitive to novel and nonstandard features. Networks trained on a large corpus of unsegmented, continuous speech without labels also developed interesting feature representations, which may be useful in both segmentation and label learning. The results of these studies, while preliminary, demonstrate that backpropagation learning can be used with complex, natural data to identify a feature structure that can serve as the basis for both analysis and nontrivial pattern recognition.

The role of the posterior parietal cortex in coordinate transformations for visual-motor integration.

Lesion to the posterior parietal cortex in monkeys and humans produces spatial deficits in movement and perception. In recording experiments from area 7a, a cortical subdivision in the posterior parietal cortex in monkeys, we have found neurons whose responses are a function of both the retinal location of visual stimuli and the position of the eyes in the orbits. By combining these signals area 7 a neurons code the location of visual stimuli with respect to the head. However, these cells respond over only limited ranges of eye positions (eye-position-dependent coding). To code location in craniotopic space at all eye positions (eye-position-independent coding) an additional step in neural processing is required that uses information distributed across populations of area 7a neurons. We describe here a neural network model, based on back-propagation learning, that both demonstrates how spatial location could be derived from the population response of area 7a neurons and accurately accounts for the observed response properties of these neurons.

A back-propagation programmed network that simulates response properties of a subset of posterior parietal neurons.

Neurons in area 7a of the posterior parietal cortex of monkeys respond to both the retinal location of a visual stimulus and the position of the eyes and by combining these signals represent the spatial location of external objects. A neural network model, programmed using back-propagation learning, can decode this spatial information from area 7a neurons and accounts for their observed response properties.

A model of hippocampal learning during classical conditioning.

During classical conditioning of the rabbit blink reflex, a large fraction of the hippocampal pyramidal neurons are recruited to respond to the conditioned stimulus. The summed response of these neurons comes to match the whole time-amplitude profile of nictitating membrane movement, whereas some individual neurons respond only to part of this profile. A model is described which accounts for these data. The model proposes a learning rule to explain neuron recruitment and uses a tapped delay line to account for the temporal learning properties of the system and the spectrum of individual unit responses.

A computational model of hippocampal place fields.

There are neurons in the hippocampus that become active only when an animal is near a particular location in a specific environment. The activity of some of these units is known to be governed by the configuration of a small set of discrete landmarks. In order to respond in this fashion, these neurons must, in effect, be able to recognize particular locations. A model of this recognition process is described which is able to make quantitative predictions about how the response of these place-field units varies as properties of the environmental landmarks are manipulated. Computer simulations of the model show that it is consistent with the available quantitative data. These simulations also predict large, characteristic changes in place-field location and size with manipulations of the environmental landmarks. Comparison of this kind of prediction with actual experiments will serve as a test of the validity of the model.

Isolation and preliminary characterization of the Chinese hamster thymidine kinase gene.

The Chinese hamster thymidine kinase (TK) gene has been isolated from a recombinant phage library constructed with genomic DNA from mouse Ltk- cells transformed to Tk+ by transfection with Chinese hamster genomic DNA. The phage library was screened by the Benton-Davis plaque hybridization technique, using as probes, subclones of recombinant phage that were isolated from mouse Ltk+ transformants by the tRNA suppressor rescue method. The Chinese hamster TK gene is contained within 13.2 kilobases of genomic DNA in the isolate designated lambda 34S4. This gene, defined by restriction enzyme sensitivity experiments, homology studies with the chicken TK gene, and mRNA blotting experiments, may extend over 8.5 kilobases. Subclones of the lambda 34S4 isolate used as hybridization probes identified a 1,400-nucleotide polyadenylated RNA as the hamster TK mRNA. The abundance of this mRNA varies dramatically in Chinese hamster cells cultured under various growth conditions, providing direct evidence that the growth dependence of TK activity may be regulated in an important way at the level of cytoplasmic TK mRNA.

Mutational analysis of the cloned chicken thymidine kinase gene.

We have analyzed the transcription and coding unit of the chicken thymidine kinase (tk) gene. We have constructed a library of mutant chicken tk genes by the in vitro linker insertion method of Heffron et al. A total of 125 mutations within a 3.0 kbp HindIII fragment containing the gene were isolated and mapped. The effect of each mutation upon the thymidine kinase gene was determined by measurement of the transfection efficiency in mouse Ltk- cells. The chicken tk mRNA is about 2 kb and polyadenylated. The direction of transcription was also determined. From these results, we propose a structure for the gene in which at least three small introns separate the amino acid coding region into at least four segments.

Mapping functional domains in the promoter region of the herpes thymidine kinase gene.

The cloned herpes simplex virus type 1 (HSV-1) thymidine kinase (TK; ATP:thymidine 5'-phosphotransferase, EC 2.7.1.21) gene can be used to transform TK- cells to a TK+ phenotype. Transformants generated in this way express TK at a basal constitutive level that is inducible to a higher level by infection with TK- herpes virus. We have studied the effect of mutations generated in vitro on both the constitutive and virus-induced expression of TK in transformants. Four Xho I linker insertions and two deletions in the 5' untranscribed region of the cloned HSV-1 TK gene were generated in vitro. A deletion that removed all but nine base pairs of the 5' untranscribed region virtually eliminated constitutive expression and completely prevented induction by herpes virus infection. Two of the insertions have particularly interesting properties. One, nine base pairs upstream from the cap site, inactivates constitutive expression without stopping induction. The other, 50 base pairs upstream from the cap site has the opposite effect (i.e., normal constitutive expression but no induction). Analysis of these results leads us to propose that the 5' untranscribed region of the HSV-1 TK gene is quite complex with several functional domains having differential roles in the constitutive and herpes-induced expression of the TK gene.

Stabilization of a degradable protein by its overexpression in Escherichia coli.

Synthesis of proteins in Escherichia coli using recombinant DNA methodology has become an important tool for isolating and studying proteins. However, the E. coli protein degradation systems can interfere with the expression of cloned genes. To examine the effect of protein degradation, we have cloned the X90 allele of the E. coli lacZ gene. The X90 allele, an ochre mutant, codes for beta-galactosidase lacking approx. 12 amino acids from the carboxyl terminus. The X90 protein is rapidly degraded in wild-type E. coli. Randomly sheared DNA fragments from lambda placZ-X90 were inserted into the EcoRI site of the plasmid pOP203-UV5-3, a derivative of pMB9 containing the lactose operator-promoter region. Recombinant plasmids that carry the lacZ-X90 gene were identified by the Lac+ phenotype of their transformants in an ochre-suppressor-containing host and the Lac- phenotype in Su degrees or supE hosts. One recombinant plasmid, p41, with an insert of 7.6 kb codes for the synthesis of the X90 promoter at a quantity equal to or greater than 50% of the total cellular protein of several strains. In contrast to the normal situation, the X90 molecules synthesized in great excess from the plasmid are stable in Su degrees hosts and can be recovered primarily from the 10 000 X g pellets of sonication lysates. The surprising stability of the overproduced X90 protein may be due to the formation of proteinaceous aggregates.