The title of this paper, "On Seeing Robots", leaves substantial scope for playful exploration. The simple ambiguity is, of course, between describing robots that see their worlds and systems that see robots. These categories are not exclusive: I also combine them and discuss robots that see robots and even robots that see themselves. Furthermore, the title is designed to echo, and pay homage to, a classic vision paper entitled "On Seeing Things" by Max Clowes [1] as I have done once before [2]. But the context, the arguments and the conclusions are new; the comparison is used explicitly here to show the difference between the classical approach and an emerging situated approach to robotic perception. The most important reading of the title is that the paper is about how we see robots; it is about the computational paradigms, the assumptions, the architectures and the tools we use to design and build robots.

The development of projections from the retinas to the cortex is mathematically analyzed according to the previously proposed thermodynamic formulation of the self-organization of neural networks. Three types of submodality included in the visual afferent pathways are assumed in two models: model (A), in which the ocularity and retinotopy are considered separately, and model (B), in which on-center/off-center pathways are considered in addition to ocularity and retinotopy. Model (A) shows striped ocular dominance spatial patterns and, in ocular dominance histograms, reveals a dip in the binocular bin. Model (B) displays spatially modulated irregular patterns and shows single-peak behavior in the histograms. When we compare the simulated results with the observed results, it is evident that the ocular dominance spatial patterns and histograms for models (A) and (B) agree very closely with those seen in monkeys and cats.

This paper explores the effect of initial weight selection on feed-forward networks learning simple functions with the back-propagation technique.

In this paper, we will describe several extensions to our earlier work, utilizing a segment-based approach. We will formulate our segmental framework and report our study on the use of multi-layer perceptrons for detection and classification of phonemes. We will also examine the outputs of the network, and compare the network performance with other classifiers. Our investigation is performed within a set of experiments that attempts to recognize 38 vowels and consonants in American English independent of speaker.

This is a summary of results with Dyna, a class of architectures for intelligent systems based on approximating dynamic programming methods. Dyna architectures integrate trial-and-error (reinforcement) learning and execution-time planning into a single process operating alternately on the world and on a learned forward model of the world. We describe and show results for two Dyna architectures, Dyna-AHC and Dyna-Q. Using a navigation task, results are shown for a simple Dyna-AHC system which simultaneously learns by trial and error, learns a world model, and plans optimal routes using the evolving world model. We show that Dyna-Q architectures (based on Watkins's Q-Iearning) are easy to adapt for use in changing environments.

The neocognitron is a neural network for pattern recognition and feature extraction. An analog CCD parallel processing architecture developed at Lincoln Laboratory is particularly well suited to the computational requirements of shared-weight networks such as the neocognitron, and implementation of the neocognitron using the CCD architecture was simulated. A modification to the neocognitron training procedure, which improves network performance under the limited arithmetic precision that would be imposed by the CCD architecture, is presented.

We are concerned with the problem of the number of nodes needed in a feedforward neural network in order to represent a fUllction to within a specified accuracy.

The learning time of a simple neural network model is obtained through an analytic computation of the eigenvalue spectrum for the Hessian matrix, which describes the second order properties of the cost function in the space of coupling coefficients. The form of the eigenvalue distribution suggests new techniques for accelerating the learning process, and provides a theoretical justification for the choice of centered versus biased state variables.

We study the evolution of the generalization ability of a simple linear perceptron with N inputs which learns to imitate a "teacher perceptron". The system is trained on p aN binary example inputs and the generalization ability measured by testing for agreement with the teacher on all 2N possible binary input patterns. The dynamics may be solved analytically and exhibits a phase transition from imperfect to perfect generalization at a 1. Except at this point the generalization ability approaches its asymptotic value exponentially, with critical slowing down near the transition; the relaxation time is ex (1 - y'a)-2.

We compare the performance of the modular architecture, composed of competing expert networks, suggested by Jacobs, Jordan, Nowlan and Hinton (1991) to the performance of a single back-propagation network on a complex, but low-dimensional, vowel recognition task. Simulations reveal that this system is capable of uncovering interesting decompositions in a complex task. The type of decomposition is strongly influenced by the nature of the input to the gating network that decides which expert to use for each case. The modular architecture also exhibits consistently better generalization on many variations of the task.