A Theory of Sequence Indexing and Working Memory in Recurrent Neural Networks

Frady, E. Paxon; Kleyko, Denis; Sommer, Friedrich T.

doi:10.1162/neco_a_01084

Cited by 83 publications

(90 citation statements)

References 44 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…They serve as seed HD vectors, and they are used to make representations for more complex objects. To generate seed HD vectors, we use bipolar dense codes of equally probable +1s and −1s, i.e., {−1, +1} d where d = 10, 000; this dimensionality works particularly well for our applications, but it is essentially a hyperparameter that can be tuned [42]. In the following, we describe similarity measure and arithmetic operations using this code.…”

Section: Background In Hd Computingmentioning

confidence: 99%

Efficient Biosignal Processing Using Hyperdimensional Computing: Network Templates for Combined Learning and Classification of ExG Signals

et al. 2019

View full text Add to dashboard Cite

Recognizing the very size of the brain's circuits, hyperdimensional (HD) computing can model neural activity patterns with points in a HD space, that is, with HD vectors. Key examined properties of HD computing include: a versatile set of arithmetic operations on HD vectors, generality, scalability, analyzability, one-shot learning, and energy efficiency. These make it a prime candidate for efficient biosignal processing where signals are noisy and nonstationary, training data sets are not huge, individual variability is significant, and energy efficiency constraints are tight. Purely based on native HD computing operators, we describe a combined method for multiclass learning and classification of various ExG biosignals such as electromyography (EMG), electroencephalography (EEG), and electrocorticography (ECoG). We develop a full set of HD network templates that comprehensively encode body potentials and brain neural activity recorded from different electrodes into a single HD vector without requiring domain expert knowledge or ad-hoc electrode selection process. Such encoded HD vector is processed as a single unit for fast one-shot learning, and robust classification. It can be interpreted to identify the most useful features as well. Compared to state-of-the-art counterparts, HD computing enables online, incremental, and fast learning as it demands less than a third as much training data as well as less preprocessing.

show abstract

Section: Background In Hd Computingmentioning

confidence: 99%

Efficient Biosignal Processing Using Hyperdimensional Computing: Network Templates for Combined Learning and Classification of ExG Signals

et al. 2019

View full text Add to dashboard Cite

show abstract

“…For a long time it was believed that the main drawback of distributed representations is the inability to represent structure. Recently, however, DRs have been developed for complexly structured representations of objects, such as sequences [60], [61], [62], [63], [64] or graphs of situations (episodes) of knowledge bases, etc., e.g. [65], [66], [7], [8], [3].…”

Section: Discussionmentioning

confidence: 99%

Neural Distributed Representations of Vector Data in Intelligent Information Technologies

Gritsenko¹,

Revunova²,

Rachkovskij³

2018

Kibern. vyčisl. teh.

View full text Add to dashboard Cite

Introduction. Distributed representation (DR) of data is a form of a vector representation,where each object is represented by a set of vector components, and each vector component can belong to representations of many objects. In ordinary vector representations, the meaning of each component is defined, which cannot be said about DR. However, the similarity of RP vectors reflects the similarity of the objects they represent.DR is a neural network approach based on modeling the representation of information in the brain, resulted from ideas about a "distributed" or "holographic" representations. DRs have a large information capacity, allow the use of a rich arsenal of methods developed for vector data, scale well for processing large amounts of data, and have a number of other advantages. Methods for data transformation to DRs have been developed for data of various types -from scalar and vector to graphs.The purpose of the article is to provide an overview of a part of the work of the Department of Neural Information Processing Technologies (International Center) in the field of neural network distributed representations. The approach is a development of the ideas of Nikolai Mikhailovich Amosov and his scientific school of modeling the structure and functions of the brain.Scope. The formation of distributed representations from the original vector representations of objects using random projection is considered. With the help of the DR, it is possible to efficiently estimate the similarity of the original objects represented by numerical vectors. Gritsenko V.I., Rachkovskij D.A., Revunova E.G. ISSN 2519-2205 (Online), ISSN 0454-9910 (Print). Киб. и выч. техн. 2018. № 4 (194) 8The use of DR allows developing regularization methods for obtaining a stable solution of discrete ill-posed inverse problems, increasing the computational efficiency and accuracy of their solution, analyzing analytically the accuracy of the solution. Thus DRs allow for increasing the efficiency of information technologies applying them.Conclusions. DRs of various data types can be used to improve the efficiency and intelligence level of information technologies. DRs have been developed for both weakly structured data, such as vectors, and for complex structured representations of objects, such as sequences, graphs of knowledge-base situations (episodes), etc. Transformation of different types of data into the DR vector format allows unifying the basic information technologies of their processing and achieving good scalability with an increase in the amount of data processed.In future, distributed representations will naturally combine information on structure and semantics to create computationally efficient and qualitatively new information technologies in which the processing of relational structures from knowledge bases is performed by the similarity of their DRs. The neurobiological relevance of distributed representations opens up the possibility of creating intelligent information technologies based on them that function similarly to...

show abstract

“…Note that distributed representations are closely related to associative memory, e.g. [49,50] as well as to human memory [51].…”

Section: Regularization Methods For Solving Dipmentioning

confidence: 99%

Random Projection and Truncated SVD for Estimating Direction of Arrival in Antenna Array

Revunova¹,

Rachkovskij²

2018

Kibern. vyčisl. teh.

View full text Add to dashboard Cite

Introduction. The need to solve inverse problems arises in many areas of science and technology in connection with the recovery of the object signal based on the results of indirect remote measurements. In the case where the transformation matrix has a high conditional number, the sequence of its singular numbers falls to zero, and the output of the measuring system contains noise, the problem of estimating the input vector is called discrete ill-posed problem (DIP). It is known that the DIP solution using pseudoinverse of the input-output transformation matrix is unstable. To overcome the instability and to improve the accuracy of the solution, regularization methods are used. Our approaches to ensuring the stability of the DIP solution (truncated singular decomposition (TSVD) and random projection (RP)) use the integer regularization parameter, which is the number of terms of the linear model. Regularization with an integer parameter makes it possible to provide a model close to the best in terms of the accuracy of the input vector recovery, and also to reduce the computational complexity by reducing the dimensionality of the problem. The purpose of the article is to develop an approach to estimating the direction of arrival of signals in the antenna array using the DIP solution, to compare the results with the well-known MUSIC method, to reveal the advantages and disadvantages of the methods. Results. Comparison of TSVD and MUSIC (implemented in real numbers) when working with correlated sources and five snapshots showed the advantage of TSVD in terms of the power of the useful signal P ratio by 2.2 times with the number of antenna elements K = 15 and by 4.7 times with K = 90. The advantage of TSVD in P ratio is by 3.7 times for K = 15 and by 4.2 times for K = 90. Comparison of RP and MUSIC (implemented in real numbers), when working with correlated sources and five snapshots, showed the advantage of RP in P ratio by 3

show abstract

A Theory of Sequence Indexing and Working Memory in Recurrent Neural Networks

Cited by 83 publications

References 44 publications

Efficient Biosignal Processing Using Hyperdimensional Computing: Network Templates for Combined Learning and Classification of ExG Signals

Efficient Biosignal Processing Using Hyperdimensional Computing: Network Templates for Combined Learning and Classification of ExG Signals

Neural Distributed Representations of Vector Data in Intelligent Information Technologies

Random Projection and Truncated SVD for Estimating Direction of Arrival in Antenna Array

Contact Info

Product

Resources

About