Saturated Transformers are Constant-Depth Threshold Circuits

Opening the “black box” (Alishahi et al., 2020) of the representations within neural networks is an important step towards building systems with robust and interpretable behavior. In NLP, one part of this question is analyzing the languages that networks can model, and the mechanisms they use to represent linguistic structure and dependencies.

One path toward this goal is via formal analysis of specific network architectures (Merrill, 2021); for example, recurrent neural networks (RNNs). Due to their autoregressive formulation, formal linguistic analysis of RNNs has often characterized their power by relating them to automata- theoretic classes of formal languages (Weiss et al., 2018; Peng et al., 2018; Merrill, 2019, inter alia). Recently, however, RNNs have largely been overtaken in NLP by a new class of models: transformers (Vaswani et al., 2017). Transformers are not autoregressive, and therefore less naturally resemble automata, posing challenges to characterizing their linguistic capacity or inductive biases in the same terms as RNNs. Instead, some recent work has related them to circuit complexity classes, a direction that we continue to pursue in this paper. Drawing on classical circuit lower bound results, Hao et al. (2022) and Hahn (2020) derive theoretical limitations of transformers with hard attention, meaning the attention distributions focus all probability mass on one index. Together, their results show that AC⁰—the class of languages recognizable by constant-depth circuit families— upper bounds the formal languages hard-attention transformers can recognize.

However, hard attention is a strong assumption, making it unclear how these results transfer to practical transformers. For example, Bhattamishra et al. (2020) showed how transformers can solve synthetic counting tasks by using uniform attention patterns, which hard attention does not allow. Motivated by this potential disconnect between theory and practice, we aim to extend circuit- based analysis to transformers with saturated attention: a generalization of hard attention that has been argued to approximate attention patterns acquired through gradient descent (Merrill et al., 2021). Broadly speaking, saturated attention goes beyond hard attention in that it can “tie” across a subset of positions, rather than selecting just one position. The tied positions are then aggregated by averaging. Qualitatively, saturated attention heads can “count”: a capability observed in transformers in practice (Bhattamishra et al., 2020). Further, Merrill et al. (2021) show that transformer training dynamics lead attention heads in several pretrained transformers to approximate saturated attention. In summary, saturated attention strictly generalizes hard attention and should more closely reflect the attention patterns acquired in practical transformers.

Our main contributions are twofold. First, we show that saturated transformers can recognize languages outside AC⁰. Then, as depicted in Table 1, we prove that transformers with floating point activations and saturated attention can only recognize formal languages in the circuit complexity class TC⁰, constituting an upper bound for a more realistic model of transformers than past results with hard attention.

In §3, we formally define our model of the transformer, including defining saturated attention in contrast to hard attention. §4 introduces circuits in theoretical computer science and relevant complexity measures and classes for them.

In §5, we first briefly analyze saturated transformers with rational values where the embedding, scoring, and activation functions are allowed to be any size-preserving function. We find such transformers to be universally powerful. We also observe that when the positional embeddings are computed in time linear in the sequence length, saturated rational-valued transformers are exactly as powerful as the complexity class of their activation functions, because the full input sequence can be pooled to a single position, and an activation function can be used as an oracle over the full input sequence. However, this setup relies on the use of unrealistic embedding functions. To move to a more realistic model of computation, we then focus on saturated transformers whose values are restricted to be floats, which have a coarser granularity and, thus, cannot encode the full input sequence into a single position.

Building on results of Pérez et al. (2019), we demonstrate in §6 that saturated transformers with float activations transcend the theoretical limitations of hard-attention transformers. In particular, we will show that they can recognize the majority language, which lies outside AC⁰. We experimentally validate that transformers can learn to recognize the majority language. Taken together, these results suggest that the very weak characterization of hard-attention transformers does not hold in practice for saturated or soft attention.

In §7, we show that, on input sequences of length n, the size of each state vector in a transformer over floats is $O(logn)$ bits, similar to saturated LSTMs (cf. Merrill, 2019). Thus, the full transformer state at any given layer has size $O(nlogn)$⁠, although each feedforward block can only locally access a small, $O(logn)$ “piece”. Thus, while hierarchical representations can be implemented in a transformer (e.g., to process arbitrary-depth Dyck languages or reverse strings as in Weiss et al. [2021]), our result implies that they must be distributed in some way across n state vectors, rather than represented compactly within a single vector.

Finally, in §8, we use the bounded size of transformer representations to upper bound the formal languages that can be recognized by saturated transformers with floating-point values. In particular, we show that such transformers can be simulated by constant-depth threshold circuits, and thus only recognize languages in TC⁰. Informally, this suggests that moving from hard attention to saturated attention can be thought of as extending the implicit class of circuit gates available in the network to include threshold gates.

Our results make progress in the analysis of transformers by deriving upper bounds for a more realistic model of transformers than has previously been analyzed. RoBERTa, T5, and other pretrained transformers have been shown to be approximately saturated (Merrill et al., 2021), so our results imply that TC⁰ may be a meaningful upper bound on the computation expressible within such networks. Our analysis also motivates future work further refining the circuit characterization of saturated transformers, as well as comparing transformers with soft and saturated attention.

We will often use w to refer to a string over any generic alphabet Σ, that is, w ∈ Σ^*. Semantically, w corresponds to the string a transformer receives as input. In contrast, we use x and other symbols to refer to binary strings in {0,1}^*. These binary strings will represent intermediate values within the transformer computation, rather than the raw input to the transformer.

Under our model, all values in the transformer are binary strings. In order to compute self attention and other operations over binary strings, we need to define datatypes describing the semantics of these binary strings as numbers. We will describe a semantics for binary strings as integers, as often comes up in circuit complexity. We then extend this to rational numbers and floats, which are necessary for representing the division operations that occur in attention heads within transformers.

We define floats $F$ as the subset of the rationals where the denominator is constrained to be a power of 2.² Multiplication and addition are defined as for ℚ, and are guaranteed to produce another float. Notably, division for floats is implemented by multiplying by an approximate multiplicative inverse, so it may be that $(x/Fy)·ℚy≠x$⁠. See Appendix A for a more formal discussion.

In §5, we will study transformers over rational values. From §6 onwards, we will then take the values in transformers to be floats unless otherwise stated. Going forward, we will generally omit datatype subscripts from operations where they are clear from context. We will sometimes write $D$ as a set in function signatures, for example, $f:Dk→Dk$⁠. In this usage, it refers to the set {0,1}^*, but it is often more intuitive to write the datatype shorthand (rather than {0,1}^*) to hint at the intended semantics of the functional arguments.

Under our model, integers, rationals, and floats are all abstractions built out of binary strings. For any x ∈ {0,1}^* (which can be interpreted semantically as an integer, float, or rational), we define its size $|x|$ as the total length of x measured in bits. We imagine a tuple $〈p,q〉$ is encoded by padding p,q to the same length with leading 0’s, and interleaving bits from each sequence. This means the size of a rational is $2max(|p|,|q|)+1$⁠. For example, the integer 2 takes 2 bits to specify, while the float $12$ takes 5 bits (1 for the sign, 2 for the numerator, 2 for the denominator).

We say that a function f : {0,1}^* →{0,1}^* is size-preserving iff there exist constants c,n such that for all inputs x with $n≤|x|$⁠, $|f(x)|≤c·|x|$⁠. Let 𝒫 be the set of size-preserving functions. While size-preserving functions are defined here over binary strings, they can be equivalently applied over integers, rationals, and floats, since these datatypes, as we have defined them, are just binary strings.

We define the following general transformer model, which can be parameterized to use different types of attention patterns and whose internal functions (e.g., feedforward blocks) can be computed by different function classes.

On an input string w ∈ Σⁿ, a transformer computes L layers of output sequences v_ℓ,1,⋯ ,v_ℓ,n (for ℓ ≤ L), where each $vℓ,i∈Dm$⁠. In the 0th layer, each token w_i and its position i are embedded into a value v_0,i. Subsequent layers aggregate information from the previous value sequence v_ℓ using a multi-head attention mechanism, and output a new value sequence v_{ℓ +1}. More formally, these layers are structured as follows:

1. Embedding Layer:v_0,i = ϕ(w_i,i).

2. Attention Head: Each of the H attention heads in layer ℓ maps the full previous sequence into a new value via s_ℓ,h and then applies the attention function α:

   $aℓ,h,i,j=sℓ,h(vℓ,i,vℓ,j)bℓ+1,h,i=∑j=1nα(aℓ,h,i,:)j·vℓ,j.$

   Crucially, the semantics for addition and multiplication here (as well as in the computation of α) come from the datatype $D$⁠.
3. Activation Block:³

   $vℓ+1,i=fℓ+1(vℓ,i,bℓ,:,i).$

An attention function α maps a vector of scores $a∈Dn$ to a probability distribution over 1,⋯ ,n. Specifically, we consider two attention functions: hard attention η(a) and soft attention ζ(a).

Hard attention collapses the attention scores to a one-hot distribution with all mass concentrated at one index. Let $M(a)={i∣ai=maxjaj}$⁠.

In contrast, saturated attention spreads probability mass evenly across “tied” scores.

Merrill (2019) shows how this form of attention can be derived by taking a large-norm limit of the network weights; a derivation can be found there. Saturated attention reduces to hard attention when $|M(a)|=1$⁠, and attends uniformly when $|M(a)|=n$⁠. Both hard and uniform attention can be implemented with numerical stability, motivating weak saturated (or, “uniform”) attention:

In general, we will use “saturated attention” to refer to strong saturated attention and provide upper bounds for this setting. On the other hand, our lower bounds only use weak saturated attention, thereby showing that even weak saturated attention is more powerful than hard attention.

Finally, we define language recognition for transformers.

This says that the decision problem of recognizing ℒ must be linearly separable using the first value in the last layer of the transformer. In practice, the first token in a transformer is often set to CLS, and its output can be passed to a classifier during finetuning (Devlin et al., 2019). This inspires Definition 5. There are other potential ways to define language recognition and generation for transformers (Hewitt et al., 2020; Yao et al., 2021), but they do not lead to meaningful differences for our purposes.

Finally, we define $AHAT(D)$ as the set of languages recognizable by some saturated transformer over $D$⁠, where the internal functions can be any size-preserving function.⁴

We note that size preservation is a weak condition to assume about the internal functions in practical transformers: Because any linear-time- computable function is size-preserving, it is strictly weaker than assuming that the internal functions can be computed in linear time. To further justify this condition, we explicitly show in Appendix B that the component functions within transformers are size-preserving.

Circuit complexity is a branch of computational complexity theory that studies circuit families as a model of computation.⁶ Intuitively, circuits are useful for formally studying the types of computational problems that can be efficiently solved with parallelism, as the depth of a circuit corresponds to the runtime of a program on an idealized, fully parallel computer. We review background on circuits, circuit families, and relevant complexity measures and classes.

For a fixed n, a circuit is a computation graph, where leaves correspond to input bits x_i and their negations ¬x_i, and the internal nodes are logic gates (typically ∧ and ∨), with one labeled as the output node. The gates can conventionally be taken to have either binary or unbounded fan-in. The circuit computes a function f : {0,1}ⁿ →{0,1} by substituting the input values into the leaf nodes, propagating the computation through the graph, and returning the value of the output node. Figure 1 shows an example circuit that takes inputs of length 5, and returns whether they contain the bigram 11.

Two important notions of complexity for a circuit are its size and depth. The size of a circuit is the number of gates. The depth is the longest path from an input node to the output node. For a circuit family, both quantities can be expressed as functions of the input size n. A circuit complexity class is a set of formal languages that can be recognized by circuit families of a certain size, depth, and set of gates. In particular, we will discuss the classes AC⁰ and TC⁰.

Intuitively, AC⁰ represents the class of problems that are highly parallelizable when the computational primitives are standard logic gates. In contrast, TC⁰ will also represent highly parallelizable computation, but when the gates are expanded to include threshold gates.

For a bitstring x ∈ {0,1}^*, define the threshold gate θ_≥k(x) to return 1 iff ≥ k bits in x are 1, and equivalently for ≤ k. For example, θ_≥3(110011) = 1.

It is known that AC⁰ ⊂TC⁰ ⊆NC¹, where NC¹ denotes the languages recognizable by $O(logn)$- depth circuits with bounded gate arity. Whether or not the latter containment between TC⁰ and NC¹ is strict is an open question. Whereas parity and other basic regular languages are outside AC⁰ (Furst et al., 1981), TC⁰ properly contains parity, although it is unknown whether it contains all the regular languages. Between AC⁰ and TC⁰ lies the class ACC⁰ (Yao, 1990).

We now begin our analysis of saturated transformers. Hao et al. (2022) and Hahn (2020) were able to give upper bounds on the power of hard attention without imposing any constraints on the embedding, scoring, and activation functions. The same will not be the case with saturated attention: any bounds on transformers will require leveraging some properties constraining their internal functions. One property we use will be size preservation. We will first show, though, that size preservation is not enough on its own: Deriving a nontrivial upper bound will depend on subtle assumptions about the transformer’s datatype.

With rational values and size-preserving internal functions, we will show saturated transformers can recognize any formal language, namely, the class ALL = {ℒ∣ℒ ⊆{0,1}^*}. Our construction resembles the universal approximation construction of Yun et al. (2020), which relies on the ability of the transformer to uniquely encode the full input string into a single value vector. After the full sequence is encoded locally into a single vector, the activation block can be used as a black box to recognize any language.

Theorem 1 says that our transformer architecture parameterized with a rational datatype can recognize any formal language. But a construction of this form feels unrealistic for two reasons. First, it requires the embedding layer to implement an unconventional prime encoding scheme in the embedding layer. Second, we are using the activation layer as a black box to recognize any language—even uncomputable ones! On the other hand, the feedforward subnetworks used in practice in transformers cannot even implement all computable functions when the weights are fixed independent of the sequence length n. We can get around both these issues by instead restricting the datatype to floats, which is the direction we will pursue in the remaining sections.⁸

In Appendix C, we develop an alternate perspective on the universality of transformers, showing that, if the embedding function is allowed to be computed in time linear in the sequence length, then the transformer’s complexity is equivalent to its activation functions’ complexity.

Appendix C contains a formal statement and proof. For example, allowing polynomial-time functions inside the transformer implies that the transformer will recognize exactly the complexity class P. A major unrealism about this setup is the assumption that ϕ can be an arbitrary function computable in time linear in n, motivating our main results in a more constrained setting in §8.

We are not stating the results in this section as evidence that practical transformers are capable of universal or arbitrary polynomial computation. Rather, the unnaturalness of these constructions (specifically, the prime numbers based position encoding) motivates us to slightly constrain our model of the transformer in a realistic way: We will switch the datatype from rationals to floats, because even using only simple uniform attention, a model with rationals and unconstrained internal functions is universal. We will soon see that this realistic constraint prevents universal simulation, and in fact bounds the capacity of the saturated transformer within TC⁰.

We now move to the setting of saturated transformers over floats. Hao et al. (2022) identified that hard-attention transformers can only recognize languages within AC⁰. In contrast, saturated transformers over floats can recognize the “majority” language maj, which is known to lie outside AC⁰ (Furst et al., 1981). Pérez et al. (2019, Prop. 3.3) show how maj can be recognized by transformers. In Theorem 3, we offer a simpler construction that leverages only a single uniform attention head, as opposed to the model of transformers they were considering. Thus, this construction is achievable with saturated attention.

In Figure 3, we show empirically that a 1-layer transformer can learn and generalize maj. This supports our argument that the theoretical limitations of hard-attention transformers do not apply to practical transformers. We train with three different types of positional encoding: none, meaning no positional information; learned, where each position gets a trainable embedding vector, and the sinusoidal scheme of Vaswani et al. (2017). The model with no positional embeddings generalizes the best, followed by the learned embeddings. It appears that while maj is in the capacity of the transformer, the standard sinusoidal positional embedding scheme provides the wrong inductive bias for learning it. This recalls the finding of Yao et al. (2021) that the choice of positional encodings seems to greatly impact the transformer’s ability to generalize formal language tasks to longer sequences.

The theoretical limits on hard-attention transformers were derived by Hao et al. (2022) by bounding the size in bits of the representation v_ℓ,i at each layer ℓ and position i. Specifically, they show that the value v_ℓ,i is representable in $O(logn)$ bits on input sequences of length n. Thus, each value can only contain limited information about the input sequence, intuitively explaining their upper bound on hard-attention transformers. Inspired by their analysis, this section will show that, in a saturated transformer, each v_ℓ,i also has a size of $O(logn)$ bits. Later in §8, we will use this property to show that saturated attention transformers are limited in the formal languages they can recognize.

How many bits does it take to represent the value of an attention head within a saturated transformer? As a naive bound, the output of a saturated attention head is specified by a float for each of n values attended over from the last layer, which would take at least linearly many bits in n. However, this upper bound on its size is not tight. Instead, we will show that all head and activation values can be represented in $O(logn)$ bits. Our analysis will rely heavily on the following lemma:

In particular, we will use Lemma 1 to show that, when each of a sequence of n values has size $O(logn)$⁠, the sum will also have size $O(logn)$⁠.

We will now leverage Lemma 1 to show that the values are of bounded size in any transformer over floats with an elementwise-size-preserving attention function.

Note that saturated attention satisfies this definition. We are ready to prove a theorem bounding the size of the representations in transformers with elementwise-size-preserving attention.

w_i ∈ Σ has size O(1), and i ∈ [n] has size $O(logn)$⁠. Since ϕ ∈ 𝒫, v_0,i = ϕ(w_i,i) has size $O(logn)$ for all i.

Corollary 4.1 follows because saturated attention is elementwise-size-preserving. Softmax attention, on the other hand, is not guaranteed to fulfill this property, because it requires computing the exponential function. This technical challenge prevents generalizing our technique to soft attention.

Similar to hard-attention transformers (Hao et al., 2022), the size of each vector representation in a saturated transformer over floats is $O(logn)$⁠. This is enough memory for individual vectors to “count”, a behavior that has been observed in both LSTMs (Weiss et al., 2018) and transformers (Bhattamishra et al., 2020). On the other hand, $O(logn)$ space is not enough memory for individual vectors (for example, the CLS output) to encode arbitrarily large combinatorial objects like trees. However, transformers are not limited to computing in an “online” fashion where tokens are consumed sequentially, meaning that their effective state is n values of size $O(logn)$⁠. Notably, trees with n leaves can be encoded in a distributed fashion across n values of size $O(logn)$⁠. One construction for this is, at index i, to store w_i and i, along with a pointer j to the parent. Since i,j can both be represented in $logn$ bits, each vector uses only $O(logn)$ bits.

Additionally, the $O(logn)$ space bound has implications from the perspective of circuit complexity. While saturated attention cannot be simulated in AC⁰, we will show in §8 that saturated transformers over $F$can be simulated by TC⁰ circuits.

We have proved that each value vector in a saturated transformer over floats has $O(logn)$ size. Now, we show how this implies saturated transformers can be simulated by TC⁰ circuits. Our results heavily leverage the following lemmas:

So that our results are self-contained, we reproduce a proof of this lemma in Appendix D. Applying Lemma 2 to a size-preserving function with at most $clogn$ input bits immediately yields:

In other words, such functions can be computed with AC⁰ circuits. In addition, we will show that the sum of n floats of size at most $clogn$ can be computed by TC⁰ circuits.

We now construct a TC⁰ circuit that simulates a saturated transformer over floats.

As input, f receives v_i as well as H head outputs, all of which have size $O(logn)$⁠. As the total size of the input is $O(logn)$⁠, we can use Corollary 2.1 to compute the output of f with an AC⁰ circuit. The size of the output is $O(logn)$ by size preservation of f. The same idea holds for ϕ as well as the linear classification head.

We have simulated each transformer component with a TC⁰ subcircuit, completing the proof.

Recall that, over rationals, we found that size- preserving saturated transformers could recognize any language. In contrast, we have now shown that using floating-point representations places such transformers within TC⁰. In this paper, we have only considered non-uniform AC⁰ and TC⁰, as opposed to the uniform variants of these classes, which are more closely connected to familiar formal language classes like the regular and context- free languages (cf. Cojocaru, 2016; Mahajan, 2007). As transformers satisfy some intuitive notion of uniformity, an open question is whether saturated transformers also fall into uniform TC⁰.

Compared with hard attention, saturated attention adds theoretical power to transformers. We showed that saturated attention lets transformers recognize languages outside AC⁰, which is the upper bound with hard attention. Further, while saturated transformers with rational values and size-preserving internal functions can recognize any language, we characterize the limits of size- preserving saturated transformers with floats. Specifically, saturated transformers with float values fall in TC⁰, a more powerful circuit class than AC⁰. Thus, going from hard to saturated attention can be understood as augmenting the model with threshold gates. This illustrates one way that the circuit complexity paradigm characterizes the power of transformers. Going forward, there are many interesting open questions that circuit analysis can answer, such as comparing the power of saturated and soft attention, and refining existing upper bounds for transformers in terms of uniform circuit families.

Thanks to Yiding Hao, Dana Angluin, and Robert Frank for sharing an early draft of their work. We also appreciate helpful feedback from Dana Angluin, Matt Gardner, Yoav Goldberg, Michael Hahn, Kyle Richardson, and Roy Schwartz.

Let / be truncated division between integers. We divide a float by an integer p by defining an approximate multiplicative inverse p⁻¹. The numerator is $2|p|/p$ and the denominator is $2|p|$⁠. For division by a float p,q, we simply apply the integer approach and then multiply by q. This yields numerator $2|p|/p·q$ and denominator $2|p|$⁠.

The fact that float division is defined in terms of integer multiplication and division implies that it is size-preserving and can be simulated in TC⁰, which we use in §8.

We justify that feedforward neural networks are size-preserving over floats. Feedforward neural networks are made up of a fixed (with respect to n) number of addition, multiplication, division, ReLU, and square root (for layer norm) operations. Therefore, it suffices to show that these operations are all in $S(F)$⁠.

For multiplication, the numerator is just p₁ · p₂, which has size $≤2|pmax|$⁠. Let the denominators be $q1=2k1$ and $q2=2k2$⁠. Then the denominator is $2k1+k2$⁠, which has size $≤2|qmax|$⁠.

Division can be analyzed in terms of the approximate multiplicative inverse (§9).⁹ Its numerator has size $≤|p|+1+|q|≤2(|p|+|q|)$ for large enough input size. The denominator has size $≤|p|+1≤2|p|$ for large enough input size.

Size preservation is trivially satisfied for ReLU, which cannot expand the size of the input.

To make layer norm work, we just need to analyze square root, which we will define in a truncated fashion over integers. The square root of a rational, then, simply takes the square root of p and q. We have that $|p|≤|p|$ and analogously for q.

Size preservation is one way to characterize the constraints on transformers’ internal functions; a slightly different perspective is to fix ϕ and analyze how the language recognition abilities of the transformer change depending on the computational resources allotted to each s_ℓ,h and f_ℓ. In this section, we derive an alternate universality theorem in terms of time complexity classes. We will show that as long as ϕ is powerful enough, such transformers have equivalent time complexity to their activation functions.

Recall that a transformer is a tuple $〈Σ,D,α,L,H,ϕ,sℓ,h,fℓ〉$⁠. In contrast to $AHAT(D)$ (cf. Definition 6), we will now work with a different class of transformer languages $AHAT(D,T(m))$ We will allow the embedding functions to be linear in the sequence length, and explore the effect of varying the complexity of the other internal functions. Let FTIME(T(m)) be the set of functions computable by a Turing machine in T(m) time.¹⁰

For any $T(m)≥m$⁠, we will show transformers $AHAT(D,T(m))$ have the complexity of their activation functions. Formally:

The proof for Lemma 2 largely follows the proof of a core lemma of Hao et al. (2022). We reproduce a slightly adapted version of their proof here, because their manuscript is not yet publicly available, and we wish for our paper to be self-contained.

Any function f : {0,1}^c →{0,1}^d can be computed by a Boolean circuit of depth 3 and size at most d(2^c + c + 1).