Yannick Deville

A unified stability analysis of the Hérault-Jutten source separation neural network

Abstract This paper presents an analysis of the stability of the equilibrium points of the Herault-Jutten neural network. We show that a previously reported numerical analysis method only yields a sufficient stability condition. By extending this method, we provide an analytical necessary and sufficient condition, and we bridge the gap with another reported method.

A class of replacement policies for medium and high-associativity structures

The content of set associative and fully associative structures (such as cache memories, TLBs and main memories) is controlled by a replacement algorithm. Replacing the elements that have not been accessed for a long period yields high performance. This property has been used in the LRU policy, it is also used in this paper, in order to define a new class of replacement policies that provide two improvements over LRU : 1) they exhibit higher performance, 2) they have a lower complexity of implementation : when the associativity increases, their complexity remains unchanged while it increases very rapidly for LRU. Trace-driven simulations show that, in 4-way caches, the performance improvement of one of the proposed policies over LRU is similar to the improvement of LRU over random and FIFO (i.e. about 10 %). For higher associativities, the proposed policies yield much higher performance in a particular range of cache sizes. In fully associative TLBs, the performance improvement is smaller but they yield much simpler implementations.

Analysis of the stability of time-domain source separation algorithms for convolutively mixed signals

In this paper, we investigate the self-adaptive source separation problem for convolutively mixed signals. The proposed approach uses a recurrent structure adapted by a generic rule involving arbitrary separating functions. We first analyze the stability of this class of algorithms. We then apply these results to some classical rules for instantaneous and convolutive mixtures that were proposed in the literature but only partly analyzed. This provides a better understanding of the conditions of operation of these rules. Eventually, we define and analyze a normalized version of the proposed type of algorithms, which yields several attractive features.

A low-cost usage-based replacement algorithm for cache memories

Mainly three replacement policies have been used in cache memories : LRU, FIFO and random. LRU achieves higher performance by being usage-based, i.e., by storing the order of the accesses to the cache blocks in order to remove the blocks that have remained unused for the longest period, However, storing this status results in a complex implementation. On the opposite, FIFO and random are non-usage-based ; their implementation is simpler, but their performance is lower. This paper presents the SIDE policy, which achieves a trade-off between the above features : by being usage-based, it yields almost the same performance as LRU. However, it is implemented almost in the same way as FIFO which yields simplicity and makes it easy to design a cache chip performing SIDE or FIFO replacement, depending on the mode chosen.

A high-selectivity continuous-time GaAs balanced filter

A second-order bandpasslike RC active filter implemented in 1- mu m-gate-length GaAs technology is presented. This filter is designed to achieve low sensitivity and high selectivity at the highest possible center frequency. Its quality factor can be tuned up to 20 or 30, using low-accuracy tuning biases, while its center frequency is tuned up to 1.6 GHz. >

A convolutive source separation method with self-optimizing non-linearities

This paper deals with the separation of two convolutively mixed signals. The proposed approach uses a recurrent structure adapted by a generic rule involving arbitrary separating functions. These functions should ideally be set so as to minimize the asymptotic error variance of the structure. However, these optimal functions are often unknown in practice. The proposed alternative is based on a self-adaptive (sub-)optimization of the separating functions, performed by estimating the projection of the optimal functions on a predefined set of elementary functions. The equilibrium and stability conditions of this rule and its asymptotic error variance are studied. Simulations are performed for real mixtures of speech signals. They show that the proposed approach yields much better performance than classical rules.

A process-dependent partitioning strategy for cache memories

This paper presents a method for partitioning cache memories which contain both instruction and data blocks. This partitioning scheme limits the total number of instruction blocks, while no conditions are set on the total number of data blocks. The partitioning is reconfigured by changing the maximum allowed number of instruction blocks, called a “threshold”. A method for choosing this threshold is provided, based on the observation that its optimum value depends on the executable code of the processes considered but almost not on its input data. Trace-driven simulations show that, when medium cache sizes are considered, the proposed structure often yields a much lower miss ratio than the corresponding classical mixed cache (typically 20% to 40% and 50% to 80% lower, respectively if the caches have 4 and 16 ways).

An automatic TV tuner alignment system

A new approach to automatic TV tuner alignment is presented. It contains two aspects, i.e. hardware (consisting of piecewise-linear transcoders) and software algorithms (which optimize the transfer functions of these transcoders with respect to measured data points). This approach allows one to choose the desired tuner performance and it yields a major improvement of the performance/price ratio compared to the conventional solution. More precisely, the performance is defined by the maximum frequency mistuning of each RF filter. For given transcoder memory sizes, the proposed approach improves the performance (i.e. decreases mistunings) by a factor of 2.7 to 4.5. Or, for a given performance, it reduces the transcoder memory sizes by a factor of 2. The target performance considered is a maximum frequency mistuning of 600 kHz in the overall frequency band of operation. This target is reached with a memory containing only 90 bits per RF filter. The proposed approach is also a good candidate for future high-definition TV standards, which will require lower frequency mistunings. Besides its use in TV tuner alignment, it is attractive in all systems requiring very simple and/or fast adaptive transcoders. It may be applied to the automatic electronic tuning of integrated active filters and phase shifters.

GaAs integrated RC active filters with high-selectivity bandpass and low-selectivity bandstop responses

This paper presents bandpass and bandstop second-order RC active filters integrated on GaAs and operating above 1 GHz. In particular, high-selectivity bandpass filters have been investigated and two structures are proposed. These structures combine high selectivity and low sensitivity more easily than classical filters when only low-gain elements are available, and they may therefore also be useful at high frequencies. The quality factor of the proposed GaAs implementations of these structures can be tuned up to 50 typically when low-accuracy tuning biases are used.

Optimum piecewise-linear transcoders Part 1. Weighted minimax piecewise-linear approximation and minimax decomposition of piecewise functions

This paper mainly concerns the following mathematical problem: an initial single-argument single-valued function F is known only through a set of points Pi(Xi; Yi Wi, with Yi = F(Xi), and with application-dependent weights Wi. An optimum (or almost-optimum in some cases) approximating function/should be derived from these points Pi f searched within a predefined class of functions. Various such classes are successively considered. They consist of subsets of piecewise-linear functions. The approximation criterion used to derivef from points Pi consists of determining an approximating function which minimizes an overall error. This error is typically defined as the maximum among local weighted errors associated with each point Pi Beyond piecewise-linear approximation, this paper also presents algorithms for optimizing the domains of operation of the subfunctions of any type of piecewise function acccording to a possibly-weighted minimax criterion. This investigation is motivated by an industrial application...