E FFICIENT, PROBABILISTIC ANALYSIS OF

COMBINATORIAL NEURAL CODES

Thomas F. Burns Irwansyah

Neural Coding and Brain Computing Unit Department of Mathematics
OIST Graduate University, Okinawa, Japan University of Mataram
[email protected] West Nusa Tenggara, Indonesia
[email protected]
arXiv:2210.10492v1 [cs.NE] 19 Oct 2022

A BSTRACT
Artificial and biological neural networks (ANNs and BNNs) can encode inputs
in the form of combinations of individual neurons’ activities. These combinato-
rial neural codes present a computational challenge for direct and efficient anal-
ysis due to their high dimensionality and often large volumes of data. Here we
improve the computational complexity – from factorial to quadratic time – of di-
rect algebraic methods previously applied to small examples and apply them to
large neural codes generated by experiments. These methods provide a novel and
efficient way of probing algebraic, geometric, and topological characteristics of
combinatorial neural codes and provide insights into how such characteristics are
related to learning and experience in neural networks. We introduce a procedure
to perform hypothesis testing on the intrinsic features of neural codes using infor-
mation geometry. We then apply these methods to neural activities from an ANN
for image classification and a BNN for 2D navigation to, without observing any
inputs or outputs, estimate the structure and dimensionality of the stimulus or task
space. Additionally, we demonstrate how an ANN varies its internal representa-
tions across network depth and during learning.

1 I NTRODUCTION
To understand the world around them, organisms’ biological neural networks (BNNs) encode infor-
mation about their environment in the dynamics of spikes varying over time and space. Artificial
neural networks (ANNs) use similar principles, except instead of transmitting spikes they usually
transmit a real-valued number in the range of [0, 1] and their dynamics are typically advanced in a
step-wise, discrete manner. Both BNNs and ANNs adjust their internal structures, e.g., connection
strengths between neurons, to improve their performance in learned tasks. This leads to encoding
input data into internal representations, which they then transform into task-relevant outputs, e.g.,
motor commands. Combinatorial neural coding schemes, i.e., encoding information in the collective
activity of neurons (also called ‘population coding’), is widespread in BNNs (Averbeck et al., 2006;
Osborne et al., 2008; Schneidman et al., 2011; Froudarakis et al., 2014; Bush et al., 2015; Stevens,
2018; Beyeler et al., 2019; Villafranca-Faus et al., 2021; Burns et al., 2022; Hannagan et al., 2021)
and long-utilized in ANNs, e.g., in associative memory networks (Little, 1974; Hopfield, 1982;
Tsodyks & Feigel'man, 1988; Adachi & Aihara, 1997; Krotov & Hopfield, 2016).
Advances in mathematical neuroscience (Curto & Itskov, 2008; Curto et al., 2019) has led to the
development of analyses designed to understand the combinatorial properties of neural codes and
their mapping to the stimulus space. Such analyses were initially inspired by the combinatorial
coding seen in place cells (Moser et al., 2008), where neurons represent physical space in the form
of ensemble and individual activity (Brown & Alex, 2006; Fenton et al., 2008). Place fields, the
physical spatial areas encoded by place cells, can be arranged such that they span multiple spatial
dimensions, e.g., 3D navigation space in bats (Yartsev & Ulanovsky, 2013). They can also encode
for ‘social place’ (Omer et al., 2018), the location of conspecifics. Just as these spatial and social
dimensions of place (external stimuli) may be represented by combinatorial coding, so too may
other dimensions in external stimuli, such as in vision (Fujii & Ito, 1996; Panzeri & Schultz, 2001;
Averbeck et al., 2006; Froudarakis et al., 2014; Fetz, 1997).

1
In place cells, the term receptive field (RF) or place field may intuitively be thought of as a physical
place. In the context of vision, for example, we may think of RFs less spatially and more abstractly as
representing stimuli features or dimensions along which neurons may respond more or less strongly,
e.g., features such as orientation, spatial frequency, or motion (Niell & Stryker, 2008; Juavinett &
Callaway, 2015). Two neurons which become activated simultaneously upon visual stimuli moving
to the right of the visual field may be said to share the RF of general rightward motion, for example.
We may also think of RFs even more abstractly as dimensions in general conceptual spaces, such
as the reward–action space of a task (Constantinescu et al., 2016), visual attributes of characters or
icons (Aronov et al., 2017), olfactory space (Bao et al., 2019), the relative positions people occupy
in a social hierarchy (Park et al., 2021), and even cognition and behaviour more generally (Bellmund
et al., 2018).
In the method described in Curto et al. (2019), tools from algebra are used to extract the combina-
torial structure of neural codes. The types of neural codes under study are sets of binary vectors
C ⊂ Fn2 , where there are n neurons in states 0 (off) and 1 (on). The ultimate structure of this method
is the canonical form of a neural code CF (C). The canonical form may be analysed topologically,
geometrically, and algebraically to infer features such as the potential convexity of the receptive
fields (RFs) which gave rise to the code, or the minimum number of dimensions those RFs must
span in real space. Such analyses are possible because CF (C) captures the minimal essential set
of combinatorial descriptions which describe all existing RF relationships implied by C. RF rela-
tionships (whether and how RFs intersect or are contained by one-another in stimulus space) are
considered to be implied by C by assuming that if two neurons become activated or spike simul-
taneously, they likely receive common external input in the form of common stimulus features or
common RFs. Given sufficient exploration of the stimulus space, it is possible to infer topolog-
ical features of the global stimulus space by only observing C (Curto & Itskov, 2008; Mulas &
Tran, 2020). To the best of our knowledge, these methods have only been developed and used for
small examples of BNNs. Here we apply them to larger BNNs and to ANNs (by considering the
co-activation of neurons during single stimulus trials).
Despite the power and broad applicability of these methods (Curto & Itskov, 2008; Curto et al.,
2019; Mulas & Tran, 2020), two major problems impede their usefulness: (1) the computational time
complexity of the algorithms to generate CF (C) is factorial in the number of codewords O(nm!)1 ,
limiting their use in large, real-world datasets; and (2) there is no tolerance for noise in C, nor
consideration given towards the stochastic or probabilistic natures of neural firing. We address these
problems by: (1) introducing a novel method for improving the time complexity to quadratic in the
number of neurons O(n2 ) by computing the generators of CF (C) and using these to answer the
same questions; and (2) using information geometry (Nakahara & Amari, 2002; Amari, 2016) to
perform hypothesis testing on the presence/absence of inferred geometric or topological properties
of the stimulus or task space. As a proof of concept, we apply these new methods to data from a
simulated BNN for spatial navigation and a simple ANN for visual classification, both of which may
contain thousands of codewords.

2 P RELIMINARIES

Before describing our own technical developments and improvements, we first outline some of the
key mathematical concepts and objects which we use and expand upon in later sections. For more
detailed information, we recommend referring to Curto & Itskov (2008); Curto et al. (2019).

2.1 C OMBINATORIAL NEURAL CODES

Let F2 = {0, 1}, [n] = {1, 2, . . . , n}, and Fn2 = {a1 a2 · · · an |ai ∈ F2 , for all i}. A codeword is
an element of Fn2 . For a given codeword c = c1 c2 · · · cn ,, we define its support as supp(c) = {i ∈
[n]|ci 6= 0}, which can be interpreted as the unique set of active neurons in a discrete time bin which
correspond to that codeword. A combinatorial neural code, or a code, is a subset of Fn2 . The support
of a code C is defined as supp(C) = {S ⊆ [n]|S = supp(c) for some c ∈ C}, which can be
interpreted as all sets of active neurons represented by all corresponding codewords in C.

1
n is the number of neurons and m is the number of codewords. In most datasets of interest n  m.

2
Let ∆ be a subset of 2[n] . The subset ∆ is an abstract simplicial complex if for any S ∈ ∆, the
condition S 0 ⊆ S gives S 0 ∈ ∆, for any S 0 ⊆ S. In other words, ∆ ⊆ 2[n] is an abstract simplicial
complex if it is closed under inclusion. So, the simplicial complex for a code C can be defined as
∆(C) = {S ⊆ [n]|S ⊆ supp(c), for some c ∈ C} .
A set S in a simplicial complex ∆ is referred to as an (|S| − 1)-simplex. For instance, a set with
cardinality 1 is called 0-simplex (geometrically, a point), a set with cardinality 2 is called a 1-simplex
(geometrically, an edge), and so on. Let S be an m-simplex in ∆. Any S 0 ⊆ S is called a face of S.

2.2 S IMPLICIAL COMPLEXES AND TOPOLOGY

Let C ⊆ Fn2 be a code and ∆(C) be the corresponding simplicial complex of C. From now on,
we will use ∆ to denote the corresponding simplicial complex of a code C. Define ∆m as a set of
m-simplices in ∆. Define
( )
X
Cm = αS S | αS ∈ F2 , ∀S ∈ ∆m .
S∈∆m
The set Cm forms a vector space over F2 whose basis elements are all the m-simplicies in ∆m . Now,
define the chain complex C∗ (∆, F2 ) to be the sequence {Cm }m≥0 . For any m ≥ 1, define a linear
Pm
transformation ∂m : Cm → Cm−1 , where for any σ ∈ ∆m , ∂m (σ) = i=0 σ i , with σ i ∈ ∆m−1
as a face of σ, for all i = 0, . . . , m. Moreover, the map ∂m can be extended linearly to all elements
in Cm as follows !
X X
∂m αS S = αS ∂m (S).
S∈∆m S∈∆m
Define the m-th mod-2 homology group of ∆ as
Ker (∂m )
Hm (∆, F2 ) =
Im (∂m+1 )
for all m ≥ 1 and
C0
H0 (∆, F2 ) = .
Im (∂1 )
Note that Hm (∆, F2 ) is also a vector space over F2 , for all m ≥ 0. So, the mod-2 m-th Betti number
βm (∆) of a simplicial complex ∆ is the dimension of Hm (∆, F2 ). The βm (∆, F2 ) gives the number
of m-dimensional holes in the geometric realisation of ∆.

2.3 C ANONICAL FORM

Q Q
Let σ and τ be subsets of [n], where σ∩τ = ∅. The polynomial of the form i∈σ xi j∈τ (1−xj ) ∈
F2 [x1 m . . . , xn ] is called a pseudo-monomial. In a given ideal J ⊆ F2 [x1 , . . . , xn ], a pseudo-
monomial f in J is said to be minimal if there is no pseudo-monomial g in J with deg(g) < deg(f )
such that f = gh for some h ∈ F2 [x1 , . . . , xn ]. For a given code C ⊆ Fn2 , we can define a neu-
0 n
ral ideal related
Q Q to C as JC = hρc0 |c F2 − Ci, where ρc0 is a pseudo-monomial of the form
i∈supp(c0 ) xi j6∈supp(c0 ) (1 − xj ) . A set of all minimal pseudo-monomials in JC , denoted by
CF (JC ) or simply CF (C), is called the canonical form of JC . Moreover, it can be shown that
JC = hCF (C)i. Therefore, the canonical form CF (C) gives a simple way to infer the RF rela-
tionships implied by all codewords in C. One way to calculate the CF (C) is by using a recursive
algorithm described in Curto et al. (2019). For a code C = {c1 , . . . , c|C| }, the aforementioned al-
gorithm works by constructing canonical forms CF (∅), CF ({c1 }) , CF ({c1 , c2 }) , . . . , CF (C) ,
respectively. In each stage, the algorithm evaluates polynomials, checks divisibility conditions, and
adds or removes polynomials from a related canonical form.

3 M ETHODS
Our main methodological contributions are: (1) improving the computational complexity of the
analyses relying on computing CF (C) (see Algorithm 1); and (2) using information geometry to
identify whether identified algebraic or topological features are statistically significant.

3
3.1 C OMPUTING AND ANALYSING THE CANONICAL FORM ’ S GENERATORS

We may perform the same analyses as in Curto et al. (2019) in quadratic time by using Algorithm
1 to construct the generators of CF (C) rather than constructing CF (C) itself (as in Algorithm 2 of
Curto et al. (2019)). Illustrative of this efficiency, representative experimental data with 25 neurons
and 46 codewords took < 1 second to analyse on a high-end desktop PC (Intel i7 CPU and 64GB of
memory), compared to 2 minutes 57 seconds using Algorithm 2 from Curto et al. (2019).

Algorithm 1 Algorithm for computing generators of CF (C)

Input:
M = C ⊂ Fn2 as a patterns × neurons matrix
Initialize:
D ← empty list . Stores the monomials.
P ← empty list . Stores the mixed monomial constructor tuples (σ,τ ).
B ← empty list . Stores the mixed monomial constructor tuples (τ ,σ).
for each column i of M do
for each P column j of M do
s ← k (i · j)k
if s < 1 then . The pair i, j have disjoint receptive fields.
append {i, j} to D
else
j 0 ←Pj−1
b ← k=1 (i · j 0 )k
if b = 0 then . The receptive field of j is a subset of receptive field of i.
append (i, j) to P
append (j, i) to B
end if
end if
end for
end for

Generating desired elements of JC is then straightforward: monomials are supersets of disjoint pairs
(from D) where each pair set has one element shared with at least one other disjoint pair set in the
superset; mixed monomials are all possible combinations of first (σ set) and second (τ set) elements
in the tuples of P (or vice-versa for B) – we do not allocate all of these elements but instead store
the set constructors; and the negative monomial appears if and only if the all 1s codeword exists
(which involves a simple summing check on columns of M ).

3.2 I NFORMATION GEOMETRY FOR COMBINATORIAL NEURAL CODES

Let N be the finite number of time bins for data of the neural activity patterns on n neurons. For any
S ⊆ [n], let v(S) ∈ Fn2 , where supp(v(S)) = S and
#{v(S)}
Pv(S) = .
N


We would like to find the parameters θ = θS1 , θS2 , . . . , θS2n −1 , where Si ⊆ [n], Si 6= ∅, and
S2n −1 = [n], such that the following exponential function
 
X
P(x, θ) = exp  θ S xS − ψ  ,
S⊆[n],S6=∅
Q
where xS = i∈S xi and ψ = − log(Pv(∅) ), describes a neural activity pattern from the given
neural activity data. We can calculate θS using the following formula for any S ⊆ [n], where
S 6= ∅, !
Pv(S)
θS = log Q ,
Pv(∅) S 0 (S,S 0 6=∅ exp (θS 0 )

4
 X
ηW = Pv(S) .
S⊆[n],S⊇W

θ


Given a θ-coordinate, we can calculate the associated G(θ) = gA,B A,B⊆[n]
matrix using the
following formula,
θ
gA,B = Eθ (XA XB ) − ηA ηB
X Y
= e−ψ eθW 0 − ηA ηB ,
W ⊇A∪B W 0 ⊆W
W 0 6=∅ (1)
P
X W 0 ⊆W θW 0
= e−ψ e W 0 6=∅ − ηA ηB ,
W ⊇A∪B

where ψ = − log(Pv(∅) ).
Example 3.1. Let n = 4, A = {1, 2}, and B = {2, 4}, then
X Y
θ
gA,B = e−ψ eθW 0
W ⊇{1,2,4} W 0 ⊆W
W 0 6=∅

Y Y
= e−ψ eθW 0 + e−ψ eθW 0
W 0 ⊆{1,2,4} W 0 ⊆{1,2,3,4}
W 0 6=∅ W 0 6=∅

= eθ{1} +θ{2} +θ{4} +θ{1,2} +θ{1,4} +θ{2,4} +θ{1,2,4} −ψ

+eθ{1} +θ{2} +θ{3} +θ{4} +θ{1,2} +θ{1,3} +θ{1,4} +θ{2,3} +θ{2,4} +θ{3,4} +θ{1,2,3} +θ{1,2,4} +θ{1,2,3,4} −ψ


−η{1,2} η{2,4}

3.3 H YPOTHESIS TESTING FOR ALGEBRAIC AND TOPOLOGICAL FEATURES

Using the previous sections, we can now perform hypothesis testing on specific RF relationships or
topological features such as holes.
Given Pv(S) for all S ⊆ [n] as in
the previous subsection, we can calculate ηW , for all W ⊆ [n],
Q
where ηW is equal to E i∈W xi = P rob{xi = 1, ∀i ∈ W }, using the following formula.
X
ηW = Pv(S)
S⊆[n],S⊇W

Given a set of neurons A ⊆ [n], where |A| = k, we want to test whether there is a k-th order
interaction between neurons in A or not. We can do this by hypothesis testing as follows.

1. Calculate θS and ηW , for all S, W ⊆ [n].

2. Specify a coordinate for P(x; η, θ) based on A as
ζkA = ηk−
A
; θkA ,

A
where ηk− = (ηH )H⊆[n],|H|≤k and θkA = (θH )H⊆[n],|H|>k .
3. Set the corresponding null hypothesis coordinate as
ζk0 = ηk−
A
; θk0 ,

where ηA = 0, ηH is equal to the previous step except for H = A, and θk0 is equal to the
one in the previous step.
θ


4. Determine the corresponding G(θ) = gA,B A,B⊆[n]
matrix related to θ-coordinate using
equation 1. Arrange the rows and columns of G(θ) such that
 
Aθ B θ
G(θ) = ,
BθT Dθ

5
 where Aθ is the submatrix of G(θ) with row and column indices from all H ⊆ [n] with
|H| ≤ k and Dθ is the submatrix of G(θ) with row and column indices from all H ⊆ [n]
with |H| > k.
 
η
5. Determine the corresponding G(η) = gA,B matrix related to η-coordinate using
A,B⊆[n]
the equation G(η) = G(θ)−1 . We can write G(η) in the form
 
Aη B η
G(η) = ,
BηT Dη
where Aη is the submatrix of G(η) with row and column indices from all H ⊆ [n] with
|H| ≤ k and Dη is the submatrix of G(η) with row and column indices from all H ⊆ [n]
with |H| > k.
6. Determine the corresponding G(ζkA ) matrix related to the mixed coordinate ζkA with
 
A ζKA O
G(ζkA ) = ,
O D ζK A

−1 −1
where AζK
A = A
θ and DζK
A = D
η .

7. Calculate the test statistic as follows

N
!
A
X P(xi ; ηk− , θk0 )
λ = 2 log
i=1
A
P(xi ; ηk− , θkA )
!!
A
P(x; ηk− , θk0 )
≈ 2N Ẽ log A , θA )
P(x; ηk− k
 A

≈ 2N D P(x; ηk− , θk0 ); P(x; ηk−
A
, θkA )
ζ 0
≈ N gAA (ηA − ηA )

ζA0
≈ N gAA
k
(ηA − ηA )
ζA
where gAA
k
is the entry of the G(ζkA ) matrix.
8. Fix a level of significance α and find the value χ2α (1) (chi-square value with significance
level α and degree of freedom 1) from the χ2 look-up table.
9. Compare λ and χ2 = max{χ2α (1), 1 − χ2α (1)}
• If λ ≥ χ2 , there is a significant interaction between neurons in A (reject the null
hypothesis)
• Otherwise, there is no significant interaction between neurons in A (accept the null
hypothesis)

Since G scales at 2n , we use a subset M of all neurons, where A ⊂ M and |M | = 10. We pick a set
A relevant to the feature we want to test the significance of, and choose random neurons (without
replacement) not already in A for the remaining elements of M , repeating the test until we exhaust
all neurons. We then correct for multiple comparisons and use α = 0.05 to detect whether there is a
significant interaction in A.
The choice of A depends on which feature we wish to analyse. When analysing whether or not two
neurons are disjoint or not in their RFs (a monomial relationship in CF (C)), we set A as those two
neurons. When analysing whether the RF of i is contained within the RF of j (a mixed monomial
relationship in CF (C)), we first set A as those two neurons, and then set A as i with a random set
of neurons (repeating this at least 5 times, with different random sets, and correcting for multiple
comparisons). For every dimension m where βm (∆, F2 ) > 0, when analysing whether a hole is
significant, we test for all possible sets A ⊂ M ⊂ [n] which close the hole. If any test closes the
hole, there is no hole, whereas if no test closes the whole, there is a hole.

6
4 A PPLICATIONS
4.1 S PATIAL NAVIGATION IN BNN S

Using the RatInABox simulation package (George et al., 2022), we created simple 2D navigation
environments with 0, 1, 2, or 3 holes in the first dimension. We used a random cover of 40 place
cells modelled using Gaussians for the probability of firing and geodesic receptive field geometries.
Starting at a random position, we then simulated random walks governed by Ornstein-Uhlenbeck
processes for 30m, with parameters based on rat locomotion data in (Sargolini et al., 2006). We
constructed a combinatorial neural code C using a window size of 10ms, allowing for up to 3,000
unique codewords. We constructed ∆(C) up to dimension 2 and calculated β1 (∆, F2 ), with the
hypothesis that β1 would be equal to the respective number of holes in the environment. Figure 1
shows an example of a single place cell and part of a simulated trajectory for an environment with
β1 = 1 and a geometric realisation of ∆(C) constructed after a 30 minute random walk. Table 1
shows the number of statistically significant holes found after different durations of the trajectories
for environments with different topologies. Although after 10 minutes of a random walk some holes
were occasionally detected, in all cases after 20 minutes all holes in the environment were detected
consistently. There were a large number of of monomials across all conditions (all simulations had
> 1000) due to the covering nature of the RF arrangements. There were also a small number (all
simulations had < 5) of mixed monomials (RFs found to be subsets, significantly so, of other RFs).

Figure 1: Example of a single place cell’s receptive field and part of a simulated trajectory for an
environment with β1 = 1 (left) and a geometric realisation of ∆(C) constructed after a 30 minute
random walk in that environment (right).

Table 1: Number of statistically significant holes found in 2D environments with 0, 1, 2, or 3 holes

after 10, 20, or 30 minutes of simulated time. Quoted are the means ± S.D.s across 10 simulations,
each with different random place cell coverings and trajectories. All simulations had β0 = 1.
Number of holes in the 2D environment
Time 0 1 2 3
10 minutes 0 ± 0 0 ± 0 0.8 ± 0.4 1.3 ± 0.46
20 minutes 0 ± 0 1 ± 0 2±0 3±0
30 minutes 0 ± 0 1 ± 0 2±0 3±0

4.2 V ISUAL CLASSIFICATION IN ANN S

We trained a multi-layer perceptron (MLP) to classify handwritten digits from the MNIST dataset
(LeCun et al., 2010) (see Figure 2, top, for examples). The model consisted of an input layer with
784 neurons (the digit pixel values), followed by two hidden layers, each with 50 neurons using
the rectified linear unit activation function and 20% dropout. The final output layer consisted of 10

7
Figure 2: Top: Example MNIST digits and their classes. Middle: Test loss and accuracy over
training epochs. Bottom: Histograms of the number of co-active neurons in each codeword, i.e.,
|supp(c)|, for C2 after 1 (left) and 10 (right) training epochs.

neurons (corresponding to the 10 digit class labels) and used a softmax activation function. The data
was split into 50,000 digits for training, 10,000 for validation, and 10,000 for testing, allowing for
up to 10,000 unique codewords in our analysis. The network was trained over 10 epochs with a
batch size of 32 samples. The optimiser was stochastic gradient descent (with learning rate 0.01 and
momentum 0.5) and the criterion was the cross-entropy loss between the one-hot vector of the true
class labels and the output layer’s activation for each sample. The MLP achieved > 96% accuracy
after 10 epochs (Figure 2, middle).
After each epoch, test samples which the network did not see during training were fed through the
network and the activity of all neurons in both hidden layers was recorded. The recorded activities
for each hidden layer corresponding to each sample were then binarized about their means (calcu-
lated over all samples) to create a code C of size 10, 000 × 50 for each layer, which we denote C1 for
layer one and C2 for layer two.

Table 2: Algebraic and geometric properties of RFs of the codes C1 and C2 from the first and second
hidden layers, respectively, of the MLP trained on MNIST classification. Numbers indicate counts
of polynomials of that type. Monomials are reported as a tuple over orders (first, second, ..., n-th),
e.g., ‘(5,3)’ means there were 8 monomials, 5 of order 1 and 3 of order 2.
Epoch 1 5 10
Layer code C1 C2 C1 C2 C1 C2
Monomials (25,9,1) (12,7) (5,1) (2) (1) -
Mixed monomials 559,171 266,322 2,530 66 10 3
Lower bound of dimension 4 3 3 2 2 1
Negative monomials 0 0 0 0 0 0
Local obstruction to convexity? No No No No No No
Intersection complete? No No No No No No

The codes C1 and C2 showed differences in their algebraic and geometric structures across training
epochs, and also differed between themselves (Table 2). In general, C1 had more overlapping RFs

8
and spanned a larger number of real dimensions (assuming convexity) than C2 . However, during
training, we find both codes lower their dimensionality and gradually spread out their RFs to cover
more of the space. This is also shown by the leftward shift between epoch 1 and 10 in the histograms
of the number of co-active neurons in C2 (Figure 2, bottom).

5 D ISCUSSION

We have shown it is possible to analyse the intrinsic geometry and topology of combinatorial neural
codes from biological and artificial networks in an efficient and probabilistic manner. With these
improved methods, we can now comfortably study codes with tens and even hundreds of thousands
of codewords. We have shown how these methods can be used to better understand (with some
statistically surety) how the internal representations of external inputs within these networks can
change through learning, experience, and network depth.
Neuroscientists have shown combinatorial neural codes can occupy low-dimensional subspaces
called neural manifolds in the covariance of their neural activities (Gallego et al., 2017; Feulner
& Clopath, 2021). Trajectories and regions in these subspaces can correspond to task cognition,
perceptual classification, and movement (Cohen et al., 2020; Chung & Abbott, 2021). For example,
Gardner et al. (2022) show the activity of populations of hundreds of grid cells within single mod-
ules of medial entorhinal cortex (a brain area partly responsible for navigation) occupy positions on
a toroidal manifold. Positions on this manifold correspond to positions in the 2D space which the
animal is navigating in.
These findings might lead us to believe combinatorial neural codes are intrinsically low-dimensional
despite being embedded in the high-dimensional combinatorial space of neural activity. However,
theoretical (Bartolo et al., 2020) and experimental (Rigotti et al., 2013) studies have shown that
the dimensionality of these neural manifolds is influenced and often directly corresponds to the
dimensionality of the task or learning under study. Indeed, the low-dimensional embeddings found
by Gardner et al. (2022) are predicted by the two-dimensionality of the navigation (the underlying
cause of the neural activity). Mathematically-optimal combinatorial neural codes and their RFs are
also related to the dimensionality of the inputs those codes are attempting to represent (Wang et al.,
2013). In more naturalistic and complex tasks, maintaining high-dimensional representations in the
neural code may allow for increased expressibility but lower generalisability, whereas reducing to
low-dimensional representations may allow for less expressibility but higher generalisability (Fusi
et al., 2016; Badre et al., 2021). High-dimensional codes are often found in recordings from BNNs
and are are often found when individual neurons encode for multiple input features, allowing linear
read-out of a large number of complex or simple features (Fusi et al., 2016). Such neurons, for
example in macaque inferotemporal cortex (Higgins et al., 2021), can also encode for very specific
and independent higher-dimensional features.
This implies combinatorial neural codes can include mixtures of coding strategies which are si-
multaneously low- and high-dimensional. One of the key advantages of the techniques developed
and applied in this study is that we can consider these different dimensionalities of coding at the
same time. We don’t reduce the embedding dimensionality to perform our analysis (which would
be equivalent to assuming a low-dimensional code). We also don’t try to map individual neuron re-
sponses to experimenter-known but network-unknown external, high-dimensional variables (which
would be equivalent to assuming a high-dimensional code). Instead, we keep the full, original
dimensionality of the data and can identify low- or high-dimensional features and response relation-
ships at local and global levels simultaneously, all without reference to information external to the
neural network. We also provide a method for testing the statistical significance of these features
and relationships, again while maintaining the original high embedding dimension of the data. This
allows us to avoid making any strong assumptions about dimensionality of the task, stimuli, or the
corresponding neural code – instead, we let the data speak for themselves.
We do carry over some limitations from prior work, most prominently: (a) we assume joint-activity
of neurons corresponds to common inputs or selectivity thereof; and (b) we binarize neural sig-
nals into ‘on’ and ‘off’ states. We suggest future work now focus on mitigating these limitations
by: (a) performing causal inference tests on neural co-activations; and (b) considering polynomials
over larger finite fields, e.g., F4 , or extending these methods to more ‘continuous’ structures, e.g.,
manifolds.

9
R EFERENCES
Masaharu Adachi and Kazuyuki Aihara. Associative dynamics in a chaotic neural net-
work. Neural Networks, 10(1):83–98, 1997. ISSN 0893-6080. doi: https://doi.org/10.
1016/S0893-6080(96)00061-5. URL https://www.sciencedirect.com/science/
article/pii/S0893608096000615.
Shun-ichi Amari. Information Geometry and Its Applications. Springer Publishing Company, In-
corporated, 1st edition, 2016. ISBN 4431559779.
Dmitriy Aronov, Rhino Nevers, and David W. Tank. Mapping of a non-spatial dimension by the
hippocampal–entorhinal circuit. Nature, 543(7647):719–722, Mar 2017. ISSN 1476-4687. doi:
10.1038/nature21692.
Bruno B Averbeck, Peter E Latham, and Alexandre Pouget. Neural correlations, population coding
and computation. Nature Reviews Neuroscience, 7:358––366, 2006. doi: 10.1038/nrn1888.
David Badre, Apoorva Bhandari, Haley Keglovits, and Atsushi Kikumoto. The dimensional-
ity of neural representations for control. Current Opinion in Behavioral Sciences, 38:20–28,
2021. ISSN 2352-1546. doi: https://doi.org/10.1016/j.cobeha.2020.07.002. URL https://
www.sciencedirect.com/science/article/pii/S2352154620301042. Com-
putational cognitive neuroscience.
Xiaojun Bao, Eva Gjorgieva, Laura K. Shanahan, James D. Howard, Thorsten Kahnt, and Jay A.
Gottfried. Grid-like neural representations support olfactory navigation of a two-dimensional
odor space. Neuron, 102(5):1066–1075.e5, Jun 2019. ISSN 0896-6273. doi: 10.1016/j.neuron.
2019.03.034. URL https://doi.org/10.1016/j.neuron.2019.03.034.
Ramon Bartolo, Richard C. Saunders, Andrew R. Mitz, and Bruno B. Averbeck. Dimensional-
ity, information and learning in prefrontal cortex. PLOS Computational Biology, 16:1–26, 04
2020. doi: 10.1371/journal.pcbi.1007514. URL https://doi.org/10.1371/journal.
pcbi.1007514.
Jacob L. S. Bellmund, Peter Gärdenfors, Edvard I. Moser, and Christian F. Doeller. Navigating cog-
nition: Spatial codes for human thinking. Science, 362(6415):eaat6766, 2018. doi: 10.1126/
science.aat6766. URL https://www.science.org/doi/abs/10.1126/science.
aat6766.
Michael Beyeler, Emily L. Rounds, Kristofor D. Carlson, Nikil Dutt, and Jeffrey L. Krichmar. Neu-
ral correlates of sparse coding and dimensionality reduction. PLOS Computational Biology, 15
(6):1–33, 06 2019. doi: 10.1371/journal.pcbi.1006908. URL https://doi.org/10.1371/
journal.pcbi.1006908.
W Michael Brown and B Alex. Optimal Neuronal Tuning for Finite Stimulus Spaces. Neural
Computation, 18(1999):1511–1526, 2006.
Thomas F. Burns, Tatsuya Haga, and Tomoki Fukai. Multiscale and extended retrieval of associative
memory structures in a cortical model of local-global inhibition balance. eNeuro, 9(3), 2022.
doi: 10.1523/ENEURO.0023-22.2022. URL https://www.eneuro.org/content/9/
3/ENEURO.0023-22.2022.
Daniel Bush, Caswell Barry, Daniel Manson, and Neil Burgess. Using Grid Cells for Navigation.
Neuron, 87(3):507–520, Aug 2015. ISSN 1097-4199. doi: 10.1016/j.neuron.2015.07.006. URL
https://pubmed.ncbi.nlm.nih.gov/26247860.
SueYeon Chung and L.F. Abbott. Neural population geometry: An approach for understand-
ing biological and artificial neural networks. Current Opinion in Neurobiology, 70:137–144,
2021. ISSN 0959-4388. doi: https://doi.org/10.1016/j.conb.2021.10.010. URL https://
www.sciencedirect.com/science/article/pii/S0959438821001227. Com-
putational Neuroscience.
Uri Cohen, SueYeon Chung, Daniel D. Lee, and Haim Sompolinsky. Separability and geometry of
object manifolds in deep neural networks. Nature Communications, 11(1):746, Feb 2020.

10
Alexandra O. Constantinescu, Jill X. O’Reilly, and Timothy E. J. Behrens. Organizing con-
ceptual knowledge in humans with a gridlike code. Science, 352(6292):1464–1468, 2016.
doi: 10.1126/science.aaf0941. URL https://www.science.org/doi/abs/10.1126/
science.aaf0941.
Carina Curto and Vladimir Itskov. Cell groups reveal structure of stimulus space. PLOS Com-
putational Biology, 4:1–13, 10 2008. doi: 10.1371/journal.pcbi.1000205. URL https:
//doi.org/10.1371/journal.pcbi.1000205.
Carina Curto, Alan Veliz-Cuba, and Nora Youngs. Chapter 7 - Analysis of Combinatorial Neu-
ral Codes: An Algebraic Approach. In Raina Robeva and Matthew Macauley (eds.), Alge-
braic and Combinatorial Computational Biology, MSE/Mathematics in Science and Engineering,
pp. 213–240. Academic Press, 2019. ISBN 978-0-12-814066-6. doi: https://doi.org/10.1016/
B978-0-12-814066-6.00007-6. URL https://www.sciencedirect.com/science/
article/pii/B9780128140666000076.
A Fenton, Hsin-yi Kao, Samuel A Neymotin, Andrey Olypher, Yevgeniy Vayntrub, William W
Lytton, and Nandor Ludvig. Unmasking the CA1 Ensemble Place Code by Exposures to Small
and Large Environments: More Place Cells and Multiple, Irregularly Arranged, and Expanded
Place Fields in the Larger Space. Journal of Neuroscience, 28:11250–11262, 2008. doi: 10.1523/
JNEUROSCI.2862-08.2008.
Eberhard E Fetz. Temporal Coding in Neural Populations? Science, 278(5345):1901–1902, 1997.
Barbara Feulner and Claudia Clopath. Neural manifold under plasticity in a goal driven learning
behaviour. PLOS Computational Biology, 17:1–27, 02 2021. doi: 10.1371/journal.pcbi.1008621.
URL https://doi.org/10.1371/journal.pcbi.1008621.
Emmanouil Froudarakis, Philipp Berens, Alexander S Ecker, R James Cotton, Fabian H Sinz, Dim-
itri Yatsenko, Peter Saggau, Matthias Bethge, and Andreas S Tolias. Population code in mouse
V1 facilitates readout of natural scenes through increased sparseness. Nature Neuroscience, 17:
851—-857, 2014. doi: 10.1038/nn.3707.
Hiroshi Fujii and Hiroyuki Ito. Dynamical Cell Assembly Hypothesis — Theoretical Possibility of
Spatio-temporal Coding in the Cortex. Neural Networks, 9:1303–1350, 1996.
Stefano Fusi, Earl K Miller, and Mattia Rigotti. Why neurons mix: high dimensionality for higher
cognition. Current Opinion in Neurobiology, 37:66–74, 2016. ISSN 0959-4388. doi: https://doi.
org/10.1016/j.conb.2016.01.010. URL https://www.sciencedirect.com/science/
article/pii/S0959438816000118. Neurobiology of cognitive behavior.
Juan A. Gallego, Matthew G. Perich, Lee E. Miller, and Sara A. Solla. Neural manifolds for the con-
trol of movement. Neuron, 94(5):978–984, 2017. ISSN 0896-6273. doi: https://doi.org/10.1016/j.
neuron.2017.05.025. URL https://www.sciencedirect.com/science/article/
pii/S0896627317304634.
Richard J. Gardner, Erik Hermansen, Marius Pachitariu, Yoram Burak, Nils A. Baas, Benjamin A.
Dunn, May-Britt Moser, and Edvard I. Moser. Toroidal topology of population activity in grid
cells. Nature, 602(7895):123–128, Feb 2022.
Tom M George, William de Cothi, Claudia Clopath, Kimberly Stachenfeld, and Caswell Barry.
RatInABox: A toolkit for modelling locomotion and neuronal activity in continuous environ-
ments. bioRxiv, 2022. doi: 10.1101/2022.08.10.503541. URL https://www.biorxiv.
org/content/early/2022/09/14/2022.08.10.503541.
T. Hannagan, A. Agrawal, L. Cohen, and S. Dehaene. Emergence of a compositional neural code
for written words: Recycling of a convolutional neural network for reading. Proceedings of the
National Academy of Sciences, 118(46), 2021. ISSN 0027-8424. doi: 10.1073/pnas.2104779118.
URL https://www.pnas.org/content/118/46/e2104779118.
Irina Higgins, Le Chang, Victoria Langston, Demis Hassabis, Christopher Summerfield, Doris
Tsao, and Matthew Botvinick. Unsupervised deep learning identifies semantic disentanglement
in single inferotemporal face patch neurons. Nature Communications, 12(1):6456, Nov 2021.

11
 ISSN 2041-1723. doi: 10.1038/s41467-021-26751-5. URL https://doi.org/10.1038/
s41467-021-26751-5.
J J Hopfield. Neural networks and physical systems with emergent collective computational abilities.
Proceedings of the National Academy of Sciences, 79(8):2554–2558, 1982. doi: 10.1073/pnas.
79.8.2554.
Ashley L. Juavinett and Edward M. Callaway. Pattern and component motion responses in mouse
visual cortical areas. Current Biology, 25:1759–1764, 2015.
D Krotov and J J Hopfield. Dense associative memory for pattern recognition. In Neural Information
Processing Systems, volume abs/1606.01164, 2016. URL http://arxiv.org/abs/1606.
01164.
Yann LeCun, Corinna Cortes, and Christopher Burges. MNIST handwritten digit
database. http://yann.lecun.com/exdb/mnist/, 2010. URL http://yann.lecun.com/
exdb/mnist/.
W. A. Little. The existence of persistent states in the brain. Mathematical Biosciences, 19(1):101–
120, 1974. ISSN 0025-5564. doi: https://doi.org/10.1016/0025-5564(74)90031-5. URL https:
//www.sciencedirect.com/science/article/pii/0025556474900315.
Edvard I Moser, Emilio Kropff, and May-britt Moser. Place Cells, Grid Cells, and the Brain’s Spatial
Representation System. Annual Review of Neuroscience, 31:69–89, 2008. doi: 10.1146/annurev.
neuro.31.061307.090723.
Raffaella Mulas and Ngoc M. Tran. Minimal embedding dimensions of connected neural codes.
Algebraic Statistics, 11:99––106, 2020. doi: 10.2140/astat.2020.11.99.
Hiroyuki Nakahara and Shun-ichi Amari. Information-Geometric Measure for Neural Spikes. Neu-
ral Computation, 14(10):2269–2316, 10 2002.
Cristopher Niell and M. P. Stryker. Highly Selective Receptive Fields in Mouse Visual Cortex.
Journal of Neuroscience, 28:7520–7536, 2008. doi: 10.1523/JNEUROSCI.0623-08.2008.Highly.
David B Omer, Shir R Maimon, and Liora Las. Social place-cells in the bat hippocampus. Science,
359:218–224, 2018.
Leslie C. Osborne, Stephanie E. Palmer, Stephen G. Lisberger, and William Bialek. The neural
basis for combinatorial coding in a cortical population response. Journal of Neuroscience, 28
(50):13522–13531, Dec 2008. ISSN 1529-2401. doi: 10.1523/JNEUROSCI.4390-08.2008. URL
https://doi.org/10.1523/JNEUROSCI.4390-08.2008.
Stefano Panzeri and Simon R Schultz. A Unified Approach to the Study of Temporal, Correlational,
and Rate Coding. Neural Computation, 13:1311–1349, 2001.
Seongmin A. Park, Douglas S. Miller, and Erie D. Boorman. Inferences on a multidimen-
sional social hierarchy use a grid-like code. Nature Neuroscience, 24(9):1292–1301, Sep 2021.
ISSN 1546-1726. doi: 10.1038/s41593-021-00916-3. URL https://doi.org/10.1038/
s41593-021-00916-3.
Mattia Rigotti, Omri Barak, Melissa R. Warden, Xiao-Jing Wang, Nathaniel D. Daw, Earl K. Miller,
and Stefano Fusi. The importance of mixed selectivity in complex cognitive tasks. Nature, 497
(7451):585–590, May 2013.
Francesca Sargolini, Marianne Fyhn, Torkel Hafting, Bruce L. McNaughton, Menno P. Witter, May-
Britt Moser, and Edvard I. Moser. Conjunctive representation of position, direction, and velocity
in entorhinal cortex. Science, 312(5774):758–762, 2006. doi: 10.1126/science.1125572. URL
https://www.science.org/doi/abs/10.1126/science.1125572.
Elad Schneidman, Jason L. Puchalla, Ronen Segev, Robert A. Harris, William Bialek, and 2nd
Berry, Michael J. Synergy from silence in a combinatorial neural code. Journal of Neuroscience,
31(44):15732–15741, Nov 2011. ISSN 1529-2401. doi: 10.1523/JNEUROSCI.0301-09.2011.
URL https://doi.org/10.1523/JNEUROSCI.0301-09.2011.

12
Charles F. Stevens. Conserved features of the primate face code. Proceedings of the National
Academy of Sciences, 115(3):584–588, 2018. ISSN 0027-8424. doi: 10.1073/pnas.1716341115.
URL https://www.pnas.org/content/115/3/584.
M. V Tsodyks and M. V Feigel'man. The enhanced storage capacity in neural networks with low
activity level. Europhysics Letters (EPL), 6(2):101–105, may 1988. doi: 10.1209/0295-5075/6/
2/002. URL https://doi.org/10.1209/0295-5075/6/2/002.
Marı́a Villafranca-Faus, Manuel Esteban Vila-Martı́n, Daniel Esteve, Esteban Merino, Anna Teruel-
Sanchis, Ana Cervera-Ferri, Joana Martı́nez-Ricós, Ana Lloret, Enrique Lanuza, and Vicent
Teruel-Martı́. Integrating pheromonal and spatial information in the amygdalo-hippocampal
network. Nature Communications, 12(1):5286, Sep 2021. ISSN 2041-1723. doi: 10.1038/
s41467-021-25442-5. URL https://doi.org/10.1038/s41467-021-25442-5.
Zhuo Wang, Alan A Stocker, and Daniel D Lee. Optimal neural population codes for high-
dimensional stimulus variables. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and
K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems, volume 26. Cur-
ran Associates, Inc., 2013. URL https://proceedings.neurips.cc/paper/2013/
file/6da9003b743b65f4c0ccd295cc484e57-Paper.pdf.
Michael M Yartsev and Nachum Ulanovsky. Representation of Three-Dimensional Space in the
Hippocampus of Flying Bats. Science, 340:367–373, 2013.

13