Figure 3 - uploaded by Joy Bose
Content may be subject to copyright.
Source publication
For a number of years, artificial neural networks have been used for a variety of applications to automate tasks not suitable for the conventional computing model, such as patternrecognition. Their inherent nonlinearity and parallelism makes them suitable for approximating a variety of functions and graphs. This work is an attempt to develop a sc...
Similar publications
In fact, the reservoir simulation is an essential step while developing the optimal operation policy for dam and reservoir. Generally, an accurate simulation for the reservoir characteristics should lead more reliable and robust optimization model for certain reservoir. The major challenge in the reservoir simulation is the non-linearity behavior o...
Citations
... The disadvantage is that it greatly increases the dimensionality of the input vector. Bose 16 uses an additional neural network, to store a measure of the context, instead of adding folds to the memory. ...
... However, for production systems, it should be implemented in hardware, thus achieving much better performance at lower cost. 16 ...
Robot navigation is a large area of research, where many different approaches have already been tried, including navigation based on visual memories. The Sparse Distributed Memory (SDM) is a kind of associative memory based on the properties of high-dimensional binary spaces. It exhibits characteristics, such as tolerance to noise and incomplete data, ability to work with sequences and the possibility of one-shot learning. Those characteristics make it appealing to use for robot navigation. The approach followed here was to navigate a robot using sequences of visual memories stored into a SDM. The robot makes intelligent decisions, such as selecting only relevant images to store, adjusting memory parameters to the level of noise and inferring new paths from the learnt trajectories. The method of encoding the information may influence the tolerance of the SDM to noise and saturation. This paper reports novel results of the limits of the model under different typical navigation problems. The SDM showed to be very robust to illumination and scenario changes, occlusion and saturation. An algorithm to build a topological map of the environment based on the visual memories is also described.
... The disadvantage is that it greatly increases the dimensionality of the input vector. J. Bose [20] uses an additional neural network, to store a measure of the context, instead of adding folds to the memory. For the present work, a solution inspired by Jaeckel and Karlsson's proposal of segmenting the addressing space [21] seemed more appropriate. ...
... Additionally, storage can be as low as 0.1 bits per bit of traditional memory. However, for production systems it should be implemented in hardware, thus achieving much better performance at a lower cost [20]. ...
Different approaches have been tried to navigate robots, including those based on visual memories. The Sparse Distributed
Memory (SDM) is a kind of associative memory based on the properties of high dimensional binary spaces. It exhibits characteristics
such as tolerance to noise and incomplete data, ability to work with sequences and the possibility of one-shot learning. Those
characteristics make it appealing to use for robot navigation. The approach presented in this work was to navigate a robot
using sequences of visual memories stored into a SDM. The robot makes intelligent decisions, such as selecting only relevant
images to store during path learning, adjusting memory parameters to the level of noise and inferring new paths from learnt
trajectories. The method of encoding the information may influence the tolerance of the SDM to noise and saturation. The present
paper reports novel results of the limits of the model under different typical navigation problems. An algorithm to build
a topological map of the environment based on the visual memories is also described.
... The disadvantage is that it greatly increases the dimensionality of the input vector. J. Bose [13] uses an additional neural network, to store a measure of the context, instead of adding folds to the memory. In the present work, it seemed more appropriate a solution inspired by Jaeckel and Karlsson's proposal of segmenting the addressing space [14] . ...
Navigation based on visual memories is very common among humans. However, planning long trips requires also a more sophisticated representation of the environment, such as a topological map. This paper describes a system that learns paths by storing sequences of images and image information in a Sparse Distributed Memory. Connections between paths are detected by exploring similarities in the images , and a topological representation of the connections is created. The robot is then able to plan paths and skip from one path to another at the connection points. The system was tested under reconstitutions of country and urban environments, and it was able to successfully plan paths and navigate.
... During a learning stage the robot stores images it can grab, and during the autonomous run it manages to follow the same path by correcting view matching errors which may occur [2, 3]. Kanerva proposes that the SDM must be ideal to store sequences of binary vectors, and J. Bose [4, 5] has extensively described this possibility. Kanerva demonstrates that the characteristics of the model hold for random binary vectors. ...
Abstract—A Sparse Distributed Memory (SDM) is a kind of associative memory,suitable to work with high-dimensional vectors of random data. This mem- ory model is attractive for Robotics and Artificial In- telligence, for it is able to mimic many characteristics of the human long-term memory. However, sensorial data is not always random: most of the times it is based on the Natural Binary Code (NBC) and tends to cluster around some specific points. This means that the SDM performs poorer than expected. As part of an ongoing project, in which we intend to navigate a robot using a sparse distributed memory to store sequences of sensorial information, we tested different methods of encoding the data. Some meth- ods perform better than others, though some may fade particular characteristics present in the original SDM model. Keywords: SDM, Sparse Distributed Memory, Data Encoding, Robotics
... It is possible that a good model arises from the study of the human brain, which is not, itself, very well understood. According to recent evidence, though, it is believed that the human brain is essentially a large memory system [1], [2], [3]. J. Hawkins [1] proposes the Memory Prediction Framework, a model which states that the brain is continuously making predictions about the world. ...
... On the other hand, there's a sound mathematical model available that, in theory, offers many of the characteristics that a human memory exhibits: Kanerva [2] created a Sparse Distributed Memory (SDM) model, and developed the mathematical support to implement it. D. Rogers [5], A. Anwar et al [6], R. Rao and D. Ballard [7], Furber et al [8], [3], among others, have implemented SDMs and improved the original model, but the SDM has never been pushed farther, despite some preliminary results. The lack of interest in this idea is not clear. ...
Despite all the progress that has been made in Robotics and Artificial Intelligence, traditional approaches seem unsuitable to build truly intelligent robots, exhibiting human-like behaviours. Many authors agree that the source of intelligence is, to a great extent, the use of a huge memory, where sequences of events that guide our later behaviour are stored. Inspired by that idea, our approach is to navigate a robot using sequences of images stored in a Sparse Distributed Memory-a kind of associative memory based on the properties of high dimensional binary spaces, which, in theory, exhibits some human-like behaviours. The robot showed good ability to correctly follow most of the sequences learnt, with small errors and good immunity to the kidnapped robot problem.
... Probabilistic or random allotment is most widely used in models; Dendrites-Choose and Axons-Choose are logical variants. Hypergeometric connectivity may at first seem difficult to achieve, but may approximate the actual, surprisingly regular, distributions of synapses in real brain circuits (see, e.g., Braitenberg and Schüz, 1998); close approximations to hypergeometric distributions may readily be achieved via initial overgeneration of synaptic contacts followed by selective die-off, concordant with what is actually observed during brain development (Purves and Lichtman, 1980; Oppenheim, 1991; Oppenheim et al., 1992; Schutze, 1993; Bose, 2003). Each of these four synaptic distribution patterns was implemented, and four different metrics were employed to measure the functional characteristics of each class of network. ...
The development of connectivity among brain networks (e.g., thalamocortical, cortico-thalamic, cortico-cortical) proceeds via a combination of axon and dendrite growth followed by a later process of synaptic pruning [Purves, D., Lichtman, J.W., 1980. Elimination of synapses in the developing nervous system. Science, 210, 153-157; Oppenheim, R.W., 1991. Cell death during development of the nervous system. Annual Review of Neuroscience, 14(1), 453-501.; Oppenheim, R., Qin-Wei Y., Prevette D., Yan Q., 1992. Brain-derived neurotrophic factor rescues developing avian motoneurons from cell death. Nature, 360, 755-757]. Sparse synaptic distribution (i.e., the low probability (<0.1) of contact among neurons; [Braitenberg, V., Schüz, A., 1998. Cortex: Statistics and geometry of neuronal connectivity: Springer Berlin.] can conform to any of a range of connectivity patterns with different distributional characteristics; and different distribution patterns can yield networks with very different functional properties. We rigorously investigate a range of different connectivity characteristics, and show that different synaptic distributions can substantially affect the functional capabilities of the resulting networks. In particular, we provide formal measures of information loss in transmission from one set of neurons to another as a function of synaptic distribution, and show a set of empirical cases with different information-theoretic utility. We characterize the trade-offs among utility and costs, and their dependence on different classes of developmental strategies by which axons from one cell group are "assigned" to synapses on dendrites from a target cell group. It is shown that hypergeometric distributions minimize a range of measured costs, compared to competing synaptic distributions. It is also found that the divergent performance among differently organized brain circuits expands with brain size, rendering the effects increasingly consequential for big brains. In summary, we propose that the characteristics of hypergeometric connectivity provide a coherent explanatory hypothesis of a range of developmental and anatomical data.
... As for actual robots, they usually rely on limited amounts of memory and heavy processing. However, evidence seems to be that the human brain works exactly the other way: limited processing and huge amounts of memory, to store sequences of events [9], [10], [4] that will lead future actions. Therefore, it's reasonable to assume that a human-like robot should rely, to a great extent, on an intelligent system with similar characteristics. ...
... The Sparse Distributed Memory [10] is a mathematically sound model of a memory which intends to model some characteristics of the human brain. D. Rogers [17], A. Anwar et al [2], R. Rao and D. Ballard [15], Furber et al [7], [4], among others, have implemented SDMs and improved the original model, but the SDM has never been pushed too farther, despite some encouraging results. ...
Traditional approaches to Artificial Intelligence (AI) and Robotics seem only to provide advances at a very slow pace. Many researchers agree that new, different approaches are needed to provide a breakthrough and allow the construction of robots with human-like capacities. Our approach consists in navigating a robot using vision a Sparse Distributed Memory (SDM), a kind of associative memory based on the properties of high dimensional binary spaces, which, in theory, exhibits some human-like behaviours. During learning the robot will store sequences of images in the SDM. During execution the robot will follow the sequence of images that is closest to its current view. Preliminary results show that the memory can store and predict sequences of images with a small error tolerance.
Navigation based on visual memories is very common among humans. However, planning long trips requires a more sophisticated
representation of the environment, such as a topological map, where connections between paths are easily noted. The present
approach is a system that learns paths by storing sequences of images and image information in a sparse distributed memory
(SDM). Connections between paths are detected by exploring similarities in the images, using the same SDM, and a topological
representation of the paths is created. The robot is then able to plan paths and switch from one path to another at the connection
points. The system was tested under reconstitutions of country and urban environments, and it was able to successfully map,
plan paths and navigate autonomously.
KeywordsSDM-Sparse distributed memory-Robot navigation-Path planning
A Sparse Distributed Memory (SDM) is a kind of associative memory suitable to work with high-dimensional vectors of random
data. This memory model exhibits the characteristics of a large boolean space, which are to a great extent those of the human
long-term memory. Hence, this model is attractive for Robotics and Artificial Intelligence, since it can possibly grant artificial
machines those same characteristics. However, the original SDM model is appropriate to work with random data. Sensorial data
is not always random: most of the times it is based on the Natural Binary Code and tends to cluster around some specific points.
This means that the SDM performs poorer than expected. As part of an ongoing project, in which the goal is to navigate a robot
using a SDM to store and retrieve sequences of images and associated path information, different methods of encoding the data
were tested. Some methods perform better than others, and one method is presented that can offer the best performance and
still maintain the characteristics of the original model.
KeywordsAssociative memory-sparse distributed memory-SDM-data encoding
Thesis (Ph. D., Information and Computer Science)--University of California, Irvine, 2006. Includes bibliographical references (leaves 166-187).
