**PaLM** (**Pathways Language Model**) uses a standard Transformer model architecture (Vaswani et al., 2017) in a decoder-only setup (i.e., each timestep can only attend to itself and past timesteps), with several modifications. PaLM is trained as a 540 billion parameter, densely activated, autoregressive Transformer on 780 billion tokens. PaLM leverages Pathways (Barham et al., 2022), which enables highly efficient training of very large neural networks across thousands of accelerator chips.

Image credit: [PaLM: Scaling Language Modeling with Pathways](https://paperswithcode.com/paper/palm-scaling-language-modeling-with-pathways-1)

PaLM

PaLM: Scaling Language Modeling with Pathways

**PrivacyNet** is a [GAN](https://paperswithcode.com/method/gan)-based semi-adversarial network (SAN) that modifies an input face image such that it can be used by a face matcher for matching purposes but cannot be reliably used by an attribute classifier. PrivacyNet allows a person to choose specific attributes that have to be obfuscated in the input face images (e.g., age and race), while allowing for other types of attributes to be extracted (e.g., gender).

PrivacyNet

PrivacyNet: Semi-Adversarial Networks for Multi-attribute Face Privacy

CLIPort, a language-conditioned imitation-learning agent that combines the broad semantic understanding (what) of CLIP [1] with the spatial precision (where) of Transporter [2].

CLIPort

CLIPort: What and Where Pathways for Robotic Manipulation

**Wasserstein Gradient Penalty Loss**, or **WGAN-GP Loss**, is a loss used for generative adversarial networks that augments the Wasserstein loss with a gradient norm penalty for random samples $\mathbf{\hat{x}} \sim \mathbb{P}\_{\hat{\mathbf{x}}}$ to achieve Lipschitz continuity:

$$ L = \mathbb{E}\_{\mathbf{\hat{x}} \sim \mathbb{P}\_{g}}\left[D\left(\tilde{\mathbf{x}}\right)\right] - \mathbb{E}\_{\mathbf{x} \sim \mathbb{P}\_{r}}\left[D\left(\mathbf{x}\right)\right] + \lambda\mathbb{E}\_{\mathbf{\hat{x}} \sim \mathbb{P}\_{\hat{\mathbf{x}}}}\left[\left(||\nabla\_{\tilde{\mathbf{x}}}D\left(\mathbf{\tilde{x}}\right)||\_{2}-1\right)^{2}\right]$$

It was introduced as part of the [WGAN-GP](https://paperswithcode.com/method/wgan-gp) overall model.

WGAN-GP Loss

Improved Training of Wasserstein GANs

**VoTr** is a [Transformer](https://paperswithcode.com/method/transformer)-based 3D backbone for 3D object detection from point clouds. It contains a series of sparse and submanifold voxel modules. Submanifold voxel modules perform multi-head self-attention strictly on the non-empty voxels, while sparse voxel modules can extract voxel features at empty locations. Long-range relationships between voxels are captured via self-attention.

Given the fact that non-empty voxels are naturally sparse but numerous, directly applying standard Transformer on voxels is non-trivial. To this end, VoTr uses a sparse voxel module and a submanifold voxel module, which can operate on the empty and non-empty voxel positions effectively. To further enlarge the attention range while maintaining comparable computational overhead to the convolutional counterparts, two attention mechanisms are used for [multi-head attention](https://paperswithcode.com/method/multi-head-attention) in those two modules: Local Attention and Dilated Attention. Furthermore [Fast Voxel Query](https://paperswithcode.com/method/fast-voxel-query) is used to accelerate the querying process in multi-head attention.

VoTr

Voxel Transformer for 3D Object Detection

As the size and complexity of models and datasets grow, so does the need for communication-efficient variants of stochastic gradient descent that can be deployed to perform parallel model training. One popular communication-compression method for data-parallel [SGD](https://paperswithcode.com/method/sgd) is QSGD (Alistarh et al., 2017), which quantizes and encodes gradients to reduce communication costs. The baseline variant of QSGD provides strong theoretical guarantees, however, for practical purposes, the authors proposed a heuristic variant which we call QSGDinf, which demonstrated impressive empirical gains for distributed training of large neural networks. In this paper, we build on this work to propose a new gradient quantization scheme, and show that it has both stronger theoretical guarantees than QSGD, and matches and exceeds the empirical performance of the QSGDinf heuristic and of other compression methods.

NUQSGD

NUQSGD: Provably Communication-efficient Data-parallel SGD via Nonuniform Quantization

**SRU++** is a self-attentive recurrent unit that combines fast recurrence and attention for sequence modeling, extending the [SRU](https://www.paperswithcode.com/method/sru) unit. The key modification of SRU++ is to incorporate more expressive non-linear operations into the recurrent network. Specifically, given the input sequence represented as a matrix $\mathbf{X} \in \mathbb{R}^{L \times d}$, the attention component computes the query, key and value representations using the following multiplications,

$$
\mathbf{Q} =\mathbf{W}^{q} \mathbf{X}^{\top} 
$$

$$
\mathbf{K} =\mathbf{W}^{k} \mathbf{Q} \\
$$

$$
\mathbf{V} =\mathbf{W}^{v} \mathbf{Q}
$$

where $\mathbf{W}^{q} \in \mathbb{R}^{d^{\prime} \times d}, \mathbf{W}^{k}, \mathbf{W}^{v} \in \mathbb{R}^{d^{\prime} \times d^{\prime}}$ are model parameters. $d^{\prime}$ is the attention dimension that is typically much smaller than $d$. Note that the keys $\mathbf{K}$ and values $\mathbf{V}$ are computed using $\mathbf{Q}$ instead of $\mathbf{X}$ such that the weight matrices $\mathbf{W}^{k}$ and $\mathbf{W}^{v}$ are significantly smaller. 

Next, we compute a weighted average output $\mathbf{A} \in \mathbb{R}^{d^{\prime} \times L}$ using [scaled dot-product attention](https://paperswithcode.com/method/scaled):

$$
\mathbf{A}^{\top}=\operatorname{softmax}\left(\frac{\mathbf{Q}^{\top} \mathbf{K}}{\sqrt{d^{\prime}}}\right) \mathbf{V}^{\top}
$$

The final output $U$ required by the elementwise recurrence is obtained by another linear projection,

$$
\mathbf{U}^{\top}=\mathbf{W}^{o}(\mathbf{Q}+\alpha \cdot \mathbf{A})
$$

where $\alpha \in \mathbb{R}$ is a learned scalar and $\mathbf{W}\_{o} \in \mathbb{R}^{3 d \times d^{\prime}}$ is a parameter matrix. $\mathbf{Q}+\alpha \cdot \mathbf{A}$ is a [residual connection](https://paperswithcode.com/method/residual-connection) which improves gradient propagation and stabilizes training. We initialize $\alpha$ to zero and as a result,

$$
\mathbf{U}^{\top}=\mathbf{W}^{o} \mathbf{Q}=\left(\mathbf{W}^{o} \mathbf{W}^{q}\right) \mathbf{X}^{\top}
$$

initially falls back to a linear transformation of the input $X$ skipping the attention transformation. Intuitively, skipping attention encourages leveraging recurrence to capture sequential patterns during early stage of training. As $|\alpha|$ grows, the attention mechanism can learn long-range dependencies for the model. In addition, $\mathbf{W}^{o} \mathbf{W}^{q}$ can be interpreted as applying a matrix factorization trick with a small inner dimension $d^{\prime}<d$, reducing the total number of parameters. The Figure compares the differences of SRU, SRU with this factorization trick (but without attention), and SRU++.

The last modification is adding [layer normalization](https://paperswithcode.com/method/layer-normalization) to each SRU++ layer. We apply normalization after the attention operation and before the matrix multiplication with $\mathbf{W}^{o}$

$$
\mathbf{U}^{\top}=\mathbf{W}^{o} \operatorname{layernorm}(\mathbf{Q}+\alpha \cdot \mathbf{A})
$$

This implementation is post-layer normalization in which the normalization is added after the residual connection.

SRU++

When Attention Meets Fast Recurrence: Training Language Models with Reduced Compute

One difficulty that arises with optimization of deep neural networks is that large parameter gradients can lead an [SGD](https://paperswithcode.com/method/sgd) optimizer to update the parameters strongly into a region where the loss function is much greater, effectively undoing much of the work that was needed to get to the current solution.

**Gradient Clipping** clips the size of the gradients to ensure optimization performs more reasonably near sharp areas of the loss surface. It can be performed in a number of ways. One option is to simply clip the parameter gradient element-wise before a parameter update. Another option is to clip the norm ||$\textbf{g}$|| of the gradient $\textbf{g}$ before a parameter update:

$$\text{ if } ||\textbf{g}||  > v \text{ then } \textbf{g} \leftarrow \frac{\textbf{g}{v}}{||\textbf{g}||}$$

where $v$ is a norm threshold.

Source: Deep Learning, Goodfellow et al

Image Source: [Pascanu et al](https://arxiv.org/pdf/1211.5063.pdf)

Gradient Clipping

**TSDAE** is an unsupervised sentence embedding method. During training, TSDAE encodes corrupted sentences into fixed-sized vectors and requires the decoder to reconstruct the original sentences from this sentence embedding. For good reconstruction quality, the semantics must be captured well in the sentence embedding from the encoder. Later, at inference, we only use the encoder for creating sentence embeddings.

The model architecture of TSDAE is a modified [encoder-decoder Transformer](https://paperswithcode.com/methods/category/autoencoding-transformers) where the key and value of the cross-attention are both confined to the sentence embedding only. Formally, the formulation of the modified cross-attention is:

$$
H^{(k)}=\text { Attention }\left(H^{(k-1)},\left[s^{T}\right],\left[s^{T}\right]\right)
$$

$$
\operatorname{Attention}(Q, K, V)=\operatorname{softmax}\left(\frac{Q K^{T}}{\sqrt{d}}\right) V
$$

where $H^{(k)} \in \mathbb{R}^{t \times d}$ is the decoder hidden states within $t$ decoding steps at the $k$-th layer, $d$ is the size of the sentence embedding, $\left[s^{T}\right] \in \mathbb{R}^{1 \times d}$ is a one-row matrix including the sentence embedding vector and $Q, K$ and $V$ are the query, key and value, respectively. By exploring different configurations on the STS benchmark dataset, the authors discover that the best combination is: (1) adopting deletion as the input noise and setting the deletion ratio to $0.6,(2)$ using the output of the [CLS] token as fixed-sized sentence representation (3) tying the encoder and decoder parameters during training.

TSDAE

TSDAE: Using Transformer-based Sequential Denoising Auto-Encoder for Unsupervised Sentence Embedding Learning

**LeCun's tanh** is an activation function of the form $f\left(x\right) = 1.7159\tanh\left(\frac{2}{3}x\right)$. The constants were chosen to keep the variance of the output close to 1.

Lecun's Tanh

**Switchable Atrous Convolution (SAC)** softly switches the convolutional computation between different atrous rates and gathers the results using switch functions. The switch functions are spatially dependent, i.e., each location of the feature map might have different switches to control the outputs of SAC. To use SAC in a detector, we convert all the standard 3x3 convolutional layers in the bottom-up backbone to SAC.

DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution

**1-bit Adam** is a [stochastic optimization](https://paperswithcode.com/methods/category/stochastic-optimization) technique that is a variant of [ADAM](https://paperswithcode.com/method/adam) with error-compensated 1-bit compression, based on finding that Adam's variance term becomes stable at an early stage. First vanilla Adam is used for a few epochs as a warm-up. After the warm-up stage, the compression stage starts and we stop updating the variance term $\mathbf{v}$ and use it as a fixed precondition. At the compression stage, we communicate based on the momentum applied with error-compensated 1-bit compression. The momentums are quantized into 1-bit representation (the sign of each element). Accompanying the vector, a scaling factor is computed as $\frac{\text { magnitude of compensated gradient }}{\text { magnitude of quantized gradient }}$. This scaling factor ensures that the compressed momentum has the same magnitude as the uncompressed momentum. This 1-bit compression could reduce the communication cost by $97 \%$ and $94 \%$ compared to the original float 32 and float 16 training, respectively.

1-bit Adam

1-bit Adam: Communication Efficient Large-Scale Training with Adam's Convergence Speed

**Boost-GNN** is an architecture that trains GBDT and GNN jointly to get the best of both worlds: the GBDT model deals with heterogeneous features, while GNN accounts for the graph structure. The model benefits from end-to-end optimization by allowing new trees to fit the gradient updates of GNN.

Boost-GNN

Boost then Convolve: Gradient Boosting Meets Graph Neural Networks

RMSNorm regularizes the summed inputs to a neuron in one layer according to root mean square (RMS), giving the model re-scaling invariance property and implicit learning rate adaptation ability. RMSNorm is computationally simpler and thus more efficient than LayerNorm.

RMSNorm

Root Mean Square Layer Normalization

A **Noisy Linear Layer** is a [linear layer](https://paperswithcode.com/method/linear-layer) with parametric noise added to the weights. This induced stochasticity can be used in reinforcement learning networks for the agent's policy to aid efficient exploration. The parameters of the noise are learned with gradient descent along with any other remaining network weights. Factorized Gaussian noise is the type of noise usually employed.

The noisy linear layer takes the form:

$$y = \left(b + Wx\right) + \left(b\_{noisy}\odot\epsilon^{b}+\left(W\_{noisy}\odot\epsilon^{w}\right)x\right) $$

where $\epsilon^{b}$ and $\epsilon^{w}$ are random variables.

Noisy Linear Layer

Noisy Networks for Exploration

**GAN Least Squares Loss** is a least squares loss function for generative adversarial networks. Minimizing this objective function is equivalent to minimizing the Pearson $\chi^{2}$ divergence. The objective function (here for [LSGAN](https://paperswithcode.com/method/lsgan)) can be defined as:

$$ \min\_{D}V\_{LS}\left(D\right) = \frac{1}{2}\mathbb{E}\_{\mathbf{x} \sim p\_{data}\left(\mathbf{x}\right)}\left[\left(D\left(\mathbf{x}\right) - b\right)^{2}\right] + \frac{1}{2}\mathbb{E}\_{\mathbf{z}\sim p\_{data}\left(\mathbf{z}\right)}\left[\left(D\left(G\left(\mathbf{z}\right)\right) - a\right)^{2}\right] $$

$$ \min\_{G}V\_{LS}\left(G\right) = \frac{1}{2}\mathbb{E}\_{\mathbf{z} \sim p\_{\mathbf{z}}\left(\mathbf{z}\right)}\left[\left(D\left(G\left(\mathbf{z}\right)\right) - c\right)^{2}\right] $$

where $a$ and $b$ are the labels for fake data and real data and $c$ denotes the value that $G$ wants $D$ to believe for fake data.

GAN Least Squares Loss

Least Squares Generative Adversarial Networks

**OPT** is a suite of decoder-only pre-trained transformers ranging from 125M to 175B parameters. The model uses an AdamW optimizer and weight decay of 0.1. It follows a linear learning rate schedule, warming up from 0 to the maximum learning rate over the first 2000 steps in OPT-175B, or over 375M tokens in the smaller models, and decaying down to 10% of the maximum LR over 300B tokens. The batch sizes range from 0.5M to 4M depending on the model size and is kept constant throughout the course of training.

OPT: Open Pre-trained Transformer Language Models

**BARThez** is a self-supervised transfer learning model for the French language based on [BART](https://paperswithcode.com/method/bart). Compared to existing [BERT](https://paperswithcode.com/method/bert)-based French language models such as [CamemBERT](https://paperswithcode.com/paper/camembert-a-tasty-french-language-model) and [FlauBERT](https://paperswithcode.com/paper/flaubert-unsupervised-language-model-pre), BARThez is well-suited for generative tasks, since not only its encoder but also its decoder is pretrained.

Viet-Anh on Software

AI Glossary

Tech Terminology

Pathways Language Model