Proc. Natl. Acad. Sci.

USA
Vol. 81, pp. 3088-3092, May 1984
Biophysics

Neurons with graded response have collective computational

properties like those of two-state neurons
(associative memory/neural network/stability/action potentials)

J. J. HOPFIELD
Divisions of Chemistry and Biology, California Institute of Technology, Pasadena, CA 91125; and Bell Laboratories, Murray Hill, NJ 07974
Contributed by J. J. Hopfield, February 13, 1984

ABSTRACT A model for a large network of "neurons" of the original model (1) but built of operational amplifiers
with a graded response (or sigmoid input-output relation) is and resistors will function.
studied. This deterministic system has collective properties in
very close correspondence with the earlier stochastic model Form of the Original Model
based on McCulloch-Pitts neurons. The content-addressable
memory and other emergent collective properties of the origi- The original model used two-state threshold "neurons" that
nal model also are present in the graded response model. The followed a stochastic algorithm. Each model neuron i had
idea that such collective properties are used in biological sys- two states, characterized by the output Vi of the neuron hav-
tems is given added credence by the continued presence of such ing the values V? or VI (which may often be taken as 0 and 1,
properties for more nearly biological "neurons." Collective respectively). The input of each neuron came from two
analog electrical circuits of the kind described will certainly sources, external inputs Ii and inputs from other neurons.
function. The collective states of the two models have a simple The total input to neuron i is then
correspondence. The original model will continue to be useful
for simulations, because its connection to graded response sys- Input to i = Hi = E TijVj + ii. [1]
tems is established. Equations that include the effect of action jsi
potentials in the graded response system are also developed.
The element Tij can be biologically viewed as a description
Recent papers (1-3) have explored the ability of a system of of the synaptic interconnection strength from neuron j to
highly interconnected "neurons" to have useful collective neuron i.
computational properties. These properties emerge sponta- CAM and other useful computations in this system involve
neously in a system having a large number of elementary the change of state of the system with time. The motion of
"neurons." Content-addressable memory (CAM) is one of the state of a system of N neurons in state space describes
the simplest collective properties of such a system. The the computation that the set of neurons is performing. A
mathematical modeling has been based on "neurons" that model therefore must describe how the state evolves in time,
are different both from real biological neurons and from the and the original model describes this in terms of a stochastic
realistic functioning of simple electronic circuits. Some of evolution. Each neuron samples its input at random times. It
these differences are major enough that neurobiologists and changes the value of its output or leaves it fixed according to
circuit engineers alike have questioned whether real neural a threshold rule with thresholds Ui.
or electrical circuits would actually exhibit the kind of be-
haviors found in the model system even if the "neurons" ViV ifE TiVj+ I, <U
were connected in the fashion envisioned. [2]
Two major divergences between the model and biological
or physical systems stand out. Real neurons (and real physi- Vi if E T11Vj + I, > U'.
cal devices such as operational amplifiers that might mimic iji
them) have continuous input-output relations. (Action po-
tentials are omitted until Discussion.) The original modeling The interrogation of each neuron is a stochastic process, tak-
used two-state McCulloch-Pitts (4) threshold devices having ing place at a mean rate W for each neuron. The times of
outputs of 0 or 1 only. Real neurons and real physical circuits interrogation of each neuron are independent of the times at
have integrative time delays due to capacitance, and the time which other neurons are interrogated. The algorithm is thus
evolution of the state of such systems should be represented asynchronous, in contrast to the usual kind of processing
by a differential equation (perhaps with added noise). The done with threshold devices. This asynchrony was deliber-
original modeling used a stochastic algorithm involving sud- ately introduced to represent a combination of propagation
den 0-1 or 1-0 changes of states of neurons at random times. delays, jitter, and noise in real neural systems. Synchronous
This paper shows that the important properties of the origi- systems might have additional collective properties (5, 6).
nal model remain intact when these two simplifications of The original model behaves as an associative memory (or
the modeling are eliminated. Although it is uncertain wheth- CAM) when the state space flow generated by the algorithm
er the properties of these new continuous "neurons" are yet is characterized by a set of stable fixed points. If these stable
close enough to the essential properties of real neurons points describe a simple flow in which nearby points in state
(and/or their dendritic arborization) to be directly applicable space tend to remain close during the flow (i.e., a nonmixing
to neurobiology, a major conceptual obstacle has been elimi- flow), then initial states that are close (in Hamming distance)
nated. It is certain that a CAM constructed on the basic ideas to a particular stable state and far from all others will tend to
terminate in that nearby stable state.
The publication costs of this article were defrayed in part by page charge
payment. This article must therefore be hereby marked "advertisement" Abbreviations: CAM, content-addressable memory; RC, resist-
Downloaded by guest on October 24, 2020

in accordance with 18 U.S.C. §1734 solely to indicate this fact. ance-capacitance.

3088
 Biophysics: Hopfield Proc. NatL Acad. Sci. USA 81 (1984) 3089

If the location of a particular stable point in state space is

thought of as the information of a particular memory of the
system, states near to that particular stable point contain
partial information about' that memory. From an initial state
of partial information about a memory, a final stable state I
with all the information of the memory is found. The memo-
ry is reached not by knowing an address, but rather by sup-
plying in the initial state some subpart of the memory. Any
subpart of adequate size will do-the memory is truly ad-
dressable by content ratner than location. A given T matrix
contains many memories simultaneously, which are recon-
structed individually from partial information in an initial
state.
Convergent flow to stable states is the essential feature of
this CAM operation. There is a simple mathematical condi-
tion which guarantees that the state space flow algorithm
converges on stable states. Any symmetric T with zero diag-
onal elements (i.e., Ti = T1i, Tii = 0) will produce such a
flow. The proof of this property followed from the construc-
tion of an appropriate energy function that is always de-
creased by any state change produced by the algorithm.
Consider the function
FIG. 1. (a) The sigmoid input-output relation for a typical neu-
E = - >1Ej T11ViVj - E I1V, + UiVI. [3] ron. All the g(u) of this paper have such a form, with possible hori-
2 i~j zontal and vertical translations. (b) The input-output relation g(Xu)
for the "neurons" of the continuous model for three values of the
The change AE in E due to changing the state of neuron i by gain scaling parameter X. (c) The output-input relation u = g-' (V)
for the g shown in b. (d) The contribution of g to the energy Qf Eq. 5
Avi is as a function of V.

AE =-Z TiiVj + Ii - Ui AVi. [4]

TijVj represents the electrical current input to cell i due to
the present potential of cell j, and Tij is thus the synapse
But according to the algorithm, A Vi is positive only when the efficacy. Linear summing of inputs is assumed. Tij of both
bracket is positive, and similarly for the negative case. Thus signs should occur. Ii is any other (fixed) input current to
any change in E under the algorithm is negative. E is bound- neuron i.
ed, so the iteration of the algorithm must lead to stable states The same set of equations represents the resistively con-
that do not further change with time. nected network of electrical amplifiers sketched in Fig. 2. It
appears more complicated than the description of the neural
A Continuous, Deterministic Model system because the electrical problem of providing inhibition
and excitation requires an additional inverting amplifier and
We now construct a model that is based on continuous varia- a negative signal wire. The magnitude of Tij is 1/Rij, where
bles and responses but retains all the significant behaviors of R is the resistor connecting the output ofj to the input line i,
the original model. Let the output variable Vi for neuron i wf~ile the sign of Tij is determined by the choice of the posi-
have the range VP < Vi < V! and be a continuous and mono-
tone-increasing function of the instantaneous input ui to neu-
ron i. The typical input-output relation g1(u,) shown in Fig.
la is sigmoid with asymptotes V? and Vi. For neurons ex-
hibiting action potentials, ui could be thought of as the mean
soma potential of a neuron from the total effect of its excit-
atory and inhibitory inputs. Vi can be viewed as the short-
term average of the firing rate of the cell i. Other biological
interpretations are possible- for example, nonlinear pro-
cessing may be done at junctions in a dendritic arbor (7), and
the model "neurons" could represent such junctions. In
terms of electrical circuits, g,{u1) represents the input-output
characteristic of a nonlinear amplifier with negligible re-
sponse time. It is convenient also to define the inverse out-
put-input relation, g7-(V).
In a biological system, ui will lag behind the instantaneous neuron
outputs Vj of the other cells because of the input capacitance
C of the cell membranes' the transmembrane resistance R, 7 amplifier 7inverting amplifier
and the finite impedance TIJ between the output Vj and the * resistor in network
cell body of cell i. Thus there is a resistance-capacitance Tij
(RC) charging equation that determines the rate of change of FIG. 2. An electrical circuit that corresponds to Eq. 5 when the
ui. amplifiers are fast. The input capacitance and resistances are not
Ci(dui/dt) = IJ TijVj - ui/Ri + Ii drawn. A particularly simple special case can have'all positive Tij of
rci the same strength and no negative Tij and replaces the array of nega-
Lo]
Downloaded by guest on October 24, 2020

tive wires with a single negative feedback amplifier sending a com-

ui =
gi (i) mon output to each "neuron."
 3090 Biophysics: Hopfield Proc. NatL. Acad. SeL USA 81 (1984)

tive or negative output of amplifier j at the connection site. Consider the case in which V? < 0 < V! for all i. Then the
Ri is now zero of voltage for each Vi can be chosen such that gi(O) = 0
for all i. Because the values of asymptotes are totally unim-
1/Ri = 1/pi + 1/Rij, [6] portant in all that follows, we will simplify notation by taking
them as ± 1 for all i. The second simplification is to treat the
where Pi is the input resistance of amplifier i. Ci is the total case in which Ii = 0 for all i. Finally, while the continuous
input capacitance of the amplifier i and its associated input case has an energy function with self-connections Tij, the
lead. We presume the output impedance of the amplifiers is discrete case need not, so Ti = 0 will be assumed for the
negligible. These simplifications result in Eq. 5 being appro- following analysis.
priate also for the network of Fig. 2. This continuous system has for symmetric T the underly-
Consider the quantity ing energy function

E =- E!
2
E i~ E = -2 E E
2 jj TijViVj + 2E lR| g['(V)dV. [11]
jj
Vj
Where are the maxima and minima of the first term of Eq.
+>E (1/R1) g7 (V)dV + E IiVi. [7] 11 in the domain of the hypercube -1 c Vi c 1 for all i? In
the usual case, all extrema lie at corners of the N-dimension-
Its time derivative for a symmetric T is al hypercube space. [In the pathological case that T is a posi-
tive or negative definite matrix, an extrermum is also possible
in the interior of the space. This is not the case for informa-
dE/dt = -a dVi/dt TijV/ - ui/Ri + Ii)* [8] tion storage matrices of the usual type '(1).]
The discrete, stochastic algorithm searches for minimal
states at the corners of the hypercube-corners that are low-
The parenthesis is the right-hand side of Eq. 5, so er than adjacent corners. Since E is a linear function of a
single Vi along any cube edge, the energy minima (or maxi-
dE/dT = -a Ci(dVi/dt)(dui/dt) ma) of
[9]
E =-- E Z T, V V [12]
2 isj
= _-E Cigg"(Vj)(dVi/dt)2.
for the discrete space Vi = + 1 are exactly the same corners
Since g7'(V1) is a monotone increasing function and Ci is as the energy maxima and minima for the continuous case
positive, each term in this sum is nQnnegative. Therefore -1 ' Vi . 1.
The second term in Eq. 11 alters the overall picture some-
dE/dt ' 0, dE/dt = 0 -* dVi/dt = 0 for all i. [10] what. To understand that alteration most easily, the gain g
can be scaled, replacing
Together with the boundedness of E, Eq. 10 shows that the
time evolution of the system is a motion in state space that Vi = gi(ui) by Vi = g1(Aud)
seeks out minima in E and comes to a stop at such points. E and
is a Liapunov function for the system.
This deterministic model has the same flow properties in -i= g l(Vi) by ui = (1/x)gi7(V1). [13]
its continuous space that the stochastic model does in its dis-
crete space. It can therefore be used in CAM or any other This scaling changes the steepness of the sigmoid gain curve
computational task for which an energy function is essential without altering the output asymptotes, as indicated in Fig.
(3). We expect that the qualitative effects of disorganized or lb. gi(x) now represents a standard form in which the scale
organized anti-symmetric parts of Tij should have similar ef- factor X = 1 corresponds to a standard gain, X >> 1 to a
fects on the CAM operation of the new' and old system. The system with very high gain and step-like gain curve, and X
new computational behaviors (such as learning sequences) small corresponds to a low gain and flat sigmoid curve (Fig.
that can be produced by antisymmetric contributions to Tij lb). The second term in E is now
within the stochastic model will also hold for the determinis-
1IRi I- gi-'(V)dV.
tic continuous model. Anecdotal support for these assertions 1
comes from unpublished work of John Platt (California Insti-
+
x 1
i
[14]
tute of Technology) solving Eq. 5 on a computer with some
random Tij removed from an otherwise symmetric T, and The integral is zero for Vi = 0 and positive otherwise, getting
from experimental work of John Lambe (Jet Propulsion Lab- very large as Vi approaches + 1 because of the slowness with
oratory), David Feinstein (California Institute of Technolo- which g(V) approaches its asymptotes (Fig. 1d). However,
gy), and Platt generating sequences of states by using an in the high-gain limit A-X00 this second term becomes negli-
antisymmetric part of T in a real circuit of a six "neurons" gible, and the locations of the maxima and minima of the full
(personal communications). energy expression become the same as that of Eq. 12 or Eq.
3 in the absence of inputs and zero thresholds. The only sta-
Relation Between the Stable States of the Two Models ble points of the very high gain, continuous, deterministic
system therefore correspond to the stable points of the sto-
For a given T, the stable states of the continuous system chastic system.
have a simple correspondence with the stable states of the For large but finite X, the second term in Eq. 11 begins to
stochastic system. We will work with a slightly simplified contribute. The form of gi(Yi) leads to a large positive contri-
instance of the general equations to put a minimum of mathe- bution near all surfaces, edges, and corners of the hypercube
matics in the way of seeing the correspondence. The same while it still contributes negligibly far from the surfaces. This
basic idea carries over, with more arithmetic, to the general leads to an energy surface that still has its maxima at corners
Downloaded by guest on October 24, 2020

case. but the minima become displaced slightly toward the interior
 Biophysics: Hopfield Proc. NatL Acad. ScL USA 81 (1984) 3091
can be nonlocally stored in a matrix of synaptic (or resistive)
interconnections ih such a way that particular memories can
be reconstructed from a starting state'that gives partial infor-
mation about one of them.
The convergence of the neuronal state of the continuous,
deterministic model to its stable states (memories) is based
on the existence of an energy function that directs the flow in
state space. Such a function can be constructed in the con-
tinuous, deterministic model when T is symmetric, just as
was the case for the original stochastic model with two-state
neurons. Other interesting uses and interpretations of the be-
haviors of the original model based on the existence of an
underlying energy function will also hold for the continuous
("graded response") model (3).
A direct correspondence between the stable states of the
two models was shown. Ior steep response curves (high
gain) there is a 1:1 correspondence between the memories of
the two models. When the response is less steep (lower gain)
the continuous-response model can have fewer stable states
than the stochastic model with the same T matrix, but the
existing stable states will still correspond to particular stable
states of the stochastic model. This simple correspondence
is possible because of the quadratic form of the interaction
between different neurons in the'energy function. More
complicated energy functions, which have occasionally been
FIG. 3. An energy contour map for a two-neuron, two-stable-
used in constraint satisfaction problems (13, 14), may have in
state system. The ordinate and abscissa are the outputs of the two addition stable states within the interior of the domain of
neurons. Stable states are located near the lower left and upper right state space in the continuous model which have no corre-
corners, and unstable extrema at the other two corners. The arrows spondence within the discrete two-state model.
show the motion of the state from Eq. 5. This motion is not in gener- This analysis indicates that a real circuit of operational
al perpendicular to the energy contours. The system parameters are amplifiers, capacitors, and resistors should be able to oper-
T12 = T21 = 1, X = 1.4, and g(u) = (2/IT)tan-1 (irTu/2). Energy con- ate as a CAM, reconstructing the stable states that have been
tours are 0.449, 0.156, 0.017, -0.003, -0.023, and -0.041. designed into T. As long as T is symmetric and the amplifiers
are fast compared with the characteristic RC time of the in-
of the space._As X decreases, each minimum moves further put network, the system will converge to stable states and
inward. As X is further decreased minima disappear one at a cannot oscillate or display chaotic behavior. While the symn-
time, when the topology of the energy surface makes a mini- metry of the network is essential to the mathematics, a prag-
mum and a saddle point coalesce. Ultimately, for very small matic view indicates that approximate symmetry will suf-
X, the second term in Eq. 11 dominates, and the only mini- fice, as was experimentally shown in the stochastic model.
mum is at Vi = 0. When the gain is large enough that there
Equivalence of the gain curves and input capacitance of the
are many minima, each is associated with a well-defined amplifiers is not needed. For high-gain systems, the stable
minimum of the infinite gain case-as the gain is increased, states of the real circuit will be exactly those predicted by
each minimum will move until it reaches a particular cube the stochastic model.
corner when X -* oo. The same kind of mapping relation Neuronal and electromagnetic signals have finite propaga-
holds in general between the continuous deterministic sys- tion velocities. A neural circuit that is to operate in the mode
tem with sigmoid response curves and the stochastic model. described must have propagation delays that are considera-
An energy contour map for a two-neuron (or two opera- bly shorter than the RC or chemical integration time of the
tional amplifier) system with two stable states is illustrated network. The same must be true for the slowness of amplifi-
in Fig. 3. The two axes are the outputs of the two amplifiers. er response in the case of the electrical circuit.
The upper left and lower right corners are stable minima for The continuous model supplements, rather than replaces,
infinite gain, and the minima are displaced inward by the the original stochastic description. The important properties
finite gain. of the original model are not due to its simplifications, but
There are many general theorems about stability in net- come from the general structure lying behind the model. Be-
works of differential equations representing chemistry, cir- cause the original model is very efficient to simulate on a
cuits, and biology (8-12). The importance of this simple sym- digital computer, it will often be more practical to develop
metric system is not merely its stability, but the fact that the ideas and simulations on that model even when use on bio-
correspondence with a discrete system lends it a special rela- logical neurons or analog circuits is intended. The interesting
tion to elementary computational devices and concepts. collective properties transcend the 0-1 stochastic simplifica-
tions.
DISCUSSION Neurons often communicate through action potentials.
Real neurons and real amplifiers have graded, continuous The output of such neurons consists of a series of sharp
outputs as a function of their inputs (or sigmoid input-output spikes having a mean frequency (when averaged over a short
curves of finite steepness) rather than steplike, two-state re- time) that is described by the input-output relation of Fig.
sponse curves. Our original stochastic model of CAM and la. In addition, the delivery of transmitter at a synapse is
other collective properties of assemblies of neurons was quantized in vesicles. Thus Eq. 5 can be only an equation for
based on two-state neurons. A continuous, deterministic the behavior of a neural network neglecting the quantal noise
neuron network of interconnected neurons with graded re- due to action potentials and the releases of discrete vesicles.
sponses has been analyzed in the previous two sections. It Because the system operates by moving downhill on an ener-
Downloaded by guest on October 24, 2020

functions as a CAM in precisely the same collective way as gy surface, the injection of a small amount of quantal noise
did the original stochastic model of CAM. A set of memories will not greatly change the minimum-seeking behavior.
 3092 Biophysics: Ho'pfield Proc. NatL Acad. Sci. USA 81 (1984)

Eq; 5 has a generalization to include action potentials. Let recovered. The second derivative term in Eq. 16 produces
all neurons have the same gain curves g(u), input capaci- noise in the system in the same fashion that diffusion pro-
tance C, input impedance-R, and maximum firing rate F. Let duces broadening in mobility-diffusion equations. These
g(u) have asymptotes 0 and 1. When a neuron has an input u, equations permit the study of the effects of action potential
it is presumed to produce action potentials VOS(t - taring) in noise on the continuous, deterministic system. Questions
a stochastic fashion with a probability Fg(u) of producing an such as the duration of stability of nominal stable states of
action potential per unit time. This stochastic view preserves the continuous, deterministic model Eq. 5 in the presence of
the basic idea of the input signal being transformed into a action potential noise should be directly answerable from
firing rate but does not allow precise timing of individual analysis or simulations of Eq. 15 or 16. Unfortunately the
action potentials. A synapse with strength Tij will deliver a steady-state solution of this problem is not equivalent to a
quantal charge VoTj to the input capacitance of neuron i thermal distribution-while Eq. 15 is a master equation, it
when neuron j produces an action potential. Let P(u1, u2, does not have detailed balance even in the high-gain limit,
Ui, ..., UN, t)du1, du2, ..., duN be the probability that input and the quantal noise is not characterized by a temperature.
potential 1 has the value ul,. . . The evolution of the state of
the network is described by The author thanks David Feinstein, John Lambe, Carver Mead,
and John Platt for discussions and permission to mention unpub-
lished work. The work at California Institute of Technology was
aP/at= > (1/RC)(a(uiP)/au,) supported in part by National Science Foundation Grant DMR-
8107494. This is contribution no. 6975 from the Division of Chemis-
try and Chemical Engineering, California Institute of Technology.
+ X Fg(uj)[-P + P(ul - TVVo/C,... ji - TUjVo/C,...)]. [i1]
1. Hopfield, J. J. (1982) Proc. Natl. Acad. Sci. USA 79, 2554-
2558.
If VO is small, the term in brackets can be expanded in a 2. Hopfield, J. J. (1984) in Modeling and Analysis in Biomedi-
Taylor series, yielding cine, ed. Nicolini, C. (World Scientific Publishing, New York),
in press.
3. Hinton, G. E. & Sejnowski, T. J. (1983) in Proceedings of the
aPlat= X (1/RC)(d(u1P)/d3u) IEEE Computer Science Conference on Computer Vision and
Pattern Recognition (Washington, DC), pp. 448-453.
4. McCulloch, W. A. & Pitts, W. (1943) Bull. Math. Biophys. 5,
-E (aP/au,)(VoF/C) E Tij g(uj) 115-133.
5. Little, W. A. (1974) Math. Biosci. 19, 101-120.
6. Little, W. A. & Shaw, G. L. (1978) Math. Biosci. 39, 281-289.
7. Poggio, T. & Torre, V. (1981) in Theoretical Approaches to
+ VOF/2C2 > g(uk)TikTik (a2P/auiauj). [16] Neurobiology, eds. Reichardt, W. E. & Poggio, T. (MIT Press,
Qj,k Cambridge, MA), pp. 28-38.
8. Glansdorf, P. & Prigogine, R. (1971) in Thermodynamic Theory
In the limit as VYO O0, F X such that FVo = constant, of Structure, Stability, and Fluctuations (Wiley, New York),
the second derivative term can be omitted. This simplifica- pp. 61-67.
tion has the solutions that are identical to those of the contin- 9. Landauer, R. (1975) J. Stat. Phys. 13, 1-16.
tous, deterministic model, namely 10. Glass, L. & Kauffman, S. A. (1973) J. Theor. Biol. 39, 103-
129.
p= H 6(ui -UP% 11. Grossberg, S. (1973) Stud. Appl. Math. 52, 213-257.
12. Glass, L. (1975) J. Chem. Phys. 63, 1325-1335.
13. Kirkpatrick, S., Gelatt, C. D. & Vecchi, M. P. (1983) Science
where ui(t) obeys Eq. 5. 220, 671-680.
In the model, stochastic noise from the action potentials 14. Geman, S. & Geman, D. (1984) IEEE Transactions Pat. Anal.
disappears in this limit and the continuous model of Eq. 5 is Mech. Intell., in press.
Downloaded by guest on October 24, 2020