- Home
- Documents
*Analog VLSI Implementation of Gradient .from our analog VLSI implementation. 1 Introduction ......*

prev

next

out of 8

View

212Download

0

Embed Size (px)

Analog VLSI Implementation of Multi-dimensional Gradient Descent

David B. Kirk, Douglas Kerns, Kurt Fleischer, Alan H. Barr California Institute of Technology

Beckman Institute 350-74 Pasadena, CA 91125

E-mail: dkIDegg.gg . cal tech. edu

Abstract

We describe an analog VLSI implementation of a multi-dimensional gradient estimation and descent technique for minimizing an on-chip scalar function fO. The implementation uses noise injec-tion and multiplicative correlation to estimate derivatives, as in [Anderson, Kerns 92]. One intended application of this technique is setting circuit parameters on-chip automatically, rather than manually [Kirk 91]. Gradient descent optimization may be used to adjust synapse weights for a backpropagation or other on-chip learning implementation. The approach combines the features of continuous multi-dimensional gradient descent and the potential for an annealing style of optimization. We present data measured from our analog VLSI implementation.

1 Introduction

This work is similar to [Anderson, Kerns 92], but represents two advances. First, we describe the extension of the technique to multiple dimensions. Second, we demon-strate an implementation of the multi-dimensional technique in analog VLSI, and provide results measured from the chip. Unlike previous work using noise sources in adaptive systems, we use the noise as a means of estimating the gradient of a function f(y), rather than performing an annealing process [Alspector 88]. We also estimate gr-;:dients continuously in position and time, in contrast to [Umminger 89] and [J abri 91], which utilize discrete position gradient estimates.

789

790 Kirk, Kerns, Fleischer, and Barr

It is interesting to note the existence of related algorithms, also presented in this volume [Cauwenberghs 93] [Alspector 93] [Flower 93]. The main difference is that our implementation operates in continuous time, with continuous differentiation and integration operators. The other approaches realize the integration and differ-entiation processes as discrete addition and subtraction operations, and use unit perturbations. [Cauwenberghs 93] provides a detailed derivation of the convergence and scaling properties of the discrete approach, and a simulation. [Alspector 93] provides a description of the use of the technique as part of a neural network hard-ware architecture, and provides a simulation. [Flower 93] derived a similar discrete algorithm from a node perturbation perspective in the context of multi-layered feed-forward networks. Our work is similar in spirit to [Dembo 90] in that we don't make any explicit assumptions about the "model" that is embodied in the function fO. The function may be implemented as a neural network. In that case, the gradient descent is on-chip learning of the parameters of the network.

We have fabricated a working chip containing the continuous-time multi-dimensional gradient descent circuits. This paper includes chip data for individ-ual circuit components, as well as the entire circuit performing multi-dimensional gradient descent and annealing.

2 The Gradient Estimation Technique

d/dt

d/dt

Figure 1: Gradient estimation technique from [Anderson, Kerns 92]

Anderson and Kerns [Anderson, Kerns 92] describe techniques for one-dimensional gradient estimation in analog hardware. The gradient is estimated by correlating (using a multiplier) the output of a scalar function f( v(t)) with a noise source n(t), as shown in Fig. 1. The function input y(t) is additively "contaminated" by the noise n(t) to produce v(t) = y(t) + n(t). A scale factor B is used to set the scale of the noise to match the function output, which improves the signal-to-noise ratio. The signals are "high-pass" filtered to approximate differentiation (shown as d/ dt operators in Fig. 1) directly before the multiplication. The results of the multiplication are "low-pass" filtered to approximate integration.

The gradient estimate is integrated over time, to smooth out some of the noise and to damp the response. This smoothed estimate is compared with a "zero" reference, using an amplifier A, and the result is fed back to the input, as shown in Fig. 2. Th~ contents of Fig. 1 are represented by the "Gradient Estimation" box in Fig. 2.

We have chosen to implement the multi-dimensional technique in analog VLSI. We

Analog VLSI Implementation of Multi-dimensional Gradient Descent 791

Gradient [ Estimation J dt

"zero"

Figure 2: Closing the loop: performing gradient descent using the gradient estimate.

will not reproduce here the one-dimensional analysis from [Anderson, Kerns 92]' but summarize some ofthe more important results, and provide a multi-dimensional derivation. [Anderson 92] provides a more detailed theoretical discussion.

3 Multi-dimensional Derivation

The multi-dimensional gradient descent operation that we are approximating can be written as follows:

(1)

where y and y' are vectors, and the solution is obtained continuously in time t, rather than at discrete ti. The circuit described in the block diagram in Fig. 1 computes an approximation to the gradient:

(2)

We approximate the operations of differentiation and integration in time by realiz-able high-pass and low-pass filters, respectively. To see that Eq. 2 is valid, and that this result is useful for approximating Eq. 1, we sketch an N-dimensional extension of [Anderson 92]. Using the chain rule,

d ~ of dtf 0l.(t) + !let)) = L.-J (yj(t) + nj(t)) ~

. Y3 3

Assuming nj(t) ~ yj (t), the rhs is approximated to produce

dd f0l.(t)+n(t)) = Lnj(t)oo~ t . Y3 3

(3)

(4)

Multiplying both sides by ni(t), and taking the expectation integral operator E[ ] of each side,

E [nitt) ~ f 0t(t) + !!(t) 1 = E [n;(t) ~>;(t) :~ ] (5) If the noise sources ni(t) and nj (t) are unc.orrelated, nat) is independent of nj (t) when i =P j, and the sum on the right has a contribution only when i = j,

E [ni(t) :t f 0l.(t) + net)) 1 = E [n~(t)n~(t) :~ 1 (6)

792 Kirk, Kerns, Fleischer, and Barr

[ d 1 of E ni(t)- f ~(t) + n(t)) ~ an- = a\l f dt UYi

(7)

The expectation operator E[] can be used to smooth random variations of the noise nj(t). So, we have

Since the descent rate k is arbitrary, we can absorb a into k. can approximate the gradient descent technique as follows:

y~(t) ~ -k E [ni(t) ~ f (1t(t) + n(t)) 1

(8)

U sing equation 8, we

(9)

4 Elements of the Multi-dimensional Implementation

We have designed, fabricated, and tested a chip which allows us to test these ideas. The chip implementation can be decomposed into six distinct parts:

noise source(s): an analog VLSI circuit which produces a noise function. An in-dependent, correlation-free noise source is needed for each input dimension, designated ni(t). The noise circuit is described in [Alspector 91].

target function: a scalar function f(Yl , Y2, ... , YN) of N input variables, bounded below, which is to be minimized [Kirk 91]. The circuit in this case is a 4-dimensional variant of the bump circuit described in [Delbriick 91]. In the general case, this fO can be any scalar function or error metric, computed by some circuit. Specifically, the function may be a neural network.

input signal(s): the inputs Yi(t) to the function fO. These will typically be on-chip values, or real-world inputs.

multiplier circuit(s): the multiplier computes the correlation between the noise values and the function output. Offsets in the multiplication appear as systematic errors in the gradient estimate, so it is important to compensate for the offsets. Linearity is not especially important, although monotonicity is critical. Ideally, the multiplication will also have a "tanh-like" character, limiting the output range for extreme inputs.

integrator: an integration over time is approximated by a low-pass filter

differentiator: the time derivatives of the noise signals and the function are ap-proximated by a high-pass filter.

The N inputs, Yi(t), are additively "contaminated" with the noise signals, ni(t), by capacitive coupling, producing Vi(t) = Yi(t) + ni(t), the inputs to the function fO. The function output is differentiated, as are the noise functions. Each differentiated noise signal is correlated with the differentiated function output, using the multi-pliers. The results are low-pass filtered, providing N partial derivative estimates, for the N input dimensions, shown for 4 dimensions in Fig. 3.

The function fO is implemented as an 4-dimensional extension of Delbriick's [D~lbriick 91] bump circuit. Details ofthe N-dimensional bump circuit can be found in [Kirk 93]. For learning and other applications, the function fO can implement some other error metric to be minimized.

Analog VLSI Implementation of Multi-dimensional Gradient Descent 793

, ............................................................................................ : : : : ~' ........................................................................................... :

n1(t) - ...... ------------n: : , ......................................................................................... ;. .. n2(t) ! ! r .. .. .................................. .L.! ~ro : : n4(t) : d/dt !

y1(t)

y2(t)

y3(t)

y4(t)

v1(t)

v2(t) ----4 I ~-+---=-'::....j

f(.)

J dt d/dt .

. . . \0 , .......... ...

Figure 3: Block diagram for a 4-dimensional gradient estimation circuit.

5 Chi p Results

We have tested chips implementing the gradient estimation and gradient descent techniques described in this paper. Figure