• Yu et al 2019 - A Review of Recurrent Neural Networks: LSTM Cells and Network Architectures
• Wu et al 2019 - A Comprehensive Survey on Graph Neural Networks
• Narang et al 2021 - Do Transformer Modifications Transfer Across Implementations and Applications? Comparison of many Transformer model variants

• Highway networks
• GLU (also considered a kind of activation, but it's more like a FF architecture). Variants: Shazeer 2020 ReGLU and SwiGLU work well.
• Capsule networks (also used in a CNN-type architecture)
• Sparsely-Gated Mixture-of-Experts. Used to greatly scale-up the number of parameters with (sub-linear? check this) increase in computation. Uses many overlapping feedforward networks that are gated by another network. 1000x improvements in model capacity.
• Dou et al 2019 - Dynamic Layer Aggregation for Neural Machine Translation with Routing-by-Agreement
• Dao et al 2022 - Monarch: Expressive Structured Matrices for Efficient and Accurate Training

• Residual connections
• ReZero Similar to residual connections, but with a trainable parameter that controls the strength of the nonlinearity (which is initialized to zero).

See also State-Space Models.

• RNNs: Elman networks, Jordan networks
• LSTMs, GRUs (see RNN Cells)
• Neural Turing Machines Cool idea, but this paper has a drawback because in practice they limit the size of the external memory, which makes it more like a neural finite-state machine (see p. 11, footnote 2). Not necessarily the be-all and end-all architecture for NTMs.
• Differentiable Neural Computer An extension to Neural Turing Machines
• Pointer Networks
• StackRNNs, StackLSTMs
• HyperNetworks Uses one network to generate the weights for another network.
• Memory networks (i.e. End-to-end memory networks)
• Associative LSTM
• Recurrent Additive Networks An early state space model
• Pointer-Generator Networks
• Convolutional Seq2seq
• Long Short-Term Memory-Network (LSTMN) Augments the LSTM cell with a memory network
• ByteNet Dilated convolution network for seq2seq that stacks the encoder and decoder, and doesn't use attention. Operates in linear time.
• Transformers
• Latent Transformer Non-autoregressive Transformer using latent variables
• Simple Self-Attention Network (SSAN) Single-layer transformer with 1 attention head
• RNMT+ Hybrid RNN/Transformer archecture. Outperforms the Transformer by half a BLEU point
• Universal Transformers A recurrent (across layers) Transformer with dynamic halting at each position
• Lightweight Recurrent Networks Related to the Transformer, LRNs are a drop-in replacement to other RNNs, which remove the sequential natural of RNN processing. Essentially uses a Key-Query-Value attention mechanism instead of the recurrence.
• Feedback Transformer Makes the Transformer recurrent by allowing each timestep to look back at all layers. Improves performance but makes training much slower because of the recurrence.
• ∞-former (Infinite former) Infinite Memory Transformer
• FNet A faster, attention-free Transformer architecture based on Fourier transforms
• Anthe: Less is More! A slim architecture for optimal language translation
• RWKV (Receptance Weighted Key Value) Network Information is passed across positions using a positional weight decay which gates the information. Allows parallel training like the transformer, but more efficient inference like the RNN
• RetNet (Retentive Network)

• TreeLSTM, also S-LSTMs

See also Wu et al 2019 - A Comprehensive Survey on Graph Neural Networks and Graph Neural Networks.

• Graph convolution networks
• Graph transformers

See also the table in Wikipedia's Activation functions.

• Sigmoid, Tahn, etc
• Softmax
• Maxout (explanation)
• Softsign
• HardTanh (from Collobert 2004)
• ReLU (history: also popularized here and earlier)
• Leaky ReLU
• Parametric ReLU (PReLU) Leaky ReLU with learned parameters.
• GLU and variants
• Gaussian error linear units (GELU) Roughly xσ(1.702x). Used in GTP-2 and BERT.
• Swish f(x) = xσ(βx). β=1.702 is GELU, β=1 is Sigmoid weighted Linear Unit (SiL)
• STL Signed Truncated Logarithm. Very cool activation function with great properties.

Comparisons:

• Narang et al 2021 Compares activation functions in the Transformer

Various representations of matrices, such as sparse, or low-dimensional ones.

• Tensor networks
• LoRA
• Monarch Matrices

• Max, average pooling
• Attention
• Transformer (it is actually a set network) and Simple Self-Attention Network (SSAN) which is a single-layer transformer with 1 attention head
• Deep sets
• Deep averaging networks (DAN) aka the neural bag-of-words model (NBOW)
• Weighted deep averaging networks. (A natural extension would be to predict the vector “a” from a pooling operation over vectors. Not sure if anyone has done this yet.)
• Weighted Multiset Automata
• See also Vinyals et al 2015 - Order Matters: Sequence to sequence for sets
• BiLSTM Aggregation
• Attentive Pooling and described in Attentive Pooling with Learnable Norms

• Neural Stacks, Queues, and DeQues (see also Probabilistic Neural Programs)
• Associative Memories
  • Memory networks A simple key-value associative memory
  • Holographic Reduced Representations An associative memory that compresses a collection of key-value vectors into a fixed-size representation using an approximation
• Continuous unbounded memory (see sections 3.2-3.3)

See also Wikipedia - Recurrent Neural Networks and Yu et al 2019 - A Review of Recurrent Neural Networks: LSTM Cells and Network Architectures

• Feedforward network (Elman network)
• Feedforward network with residual connections (with careful tuning, has been shown to perform as well as LSTMs I believe)
• LSTM
  • Forget gate
  • Peephole connections
• Associative LSTM
• GRU (has been shown not to perform as well as the LSTM cell, for example here)
• Minimal Gated Unit (MGU)

See Position Embeddings.

See also the Attention Mechanisms page.

• Feedforward attention (the original one)
• Dot product attention (aka Luong attention)
• Structural attention
• Structured Attention Networks
• Label Attention Layer
• Linear Attention (Faster to compute - makes the Transformer O(n))
• Random Feature Attention Uses random features to approximate the softmax, making it O(1).
• Continuous Attention Mechanism, used here
• Single-Headed Gated Attention Can simulate multi-head attention, and is more expressive (see Sect 3.3 and Theorem 1).

See also Neurosymbolic Methods

• PossibleWorldNet
• Neural Module Networks (see Neural Module Networks)
• Probabilistic Neural Programs

See also Conditional Computation.

• Infinite Neural Networks (GPNNs and Neural Tangent Kernel)
• Binarized Neural Networks
• Auto-Sizing Neural Networks