Add cian version

backend/app.py

@@ -39,8 +39,8 @@ def load_document(filename):
 # Load prompts at startup
 SYSTEM_PROMPT_TEMPLATE = load_prompt("system_prompt.txt")
 TRANSITION_PROMPT_TEMPLATE = load_prompt("transition_prompt.txt")
-DOCUMENT = load_document("endy.txt")
-USER_GOAL = "Understand the new in situ doping strategy introduced here, how it changes precursor formation and cathode performance compared to conventional methods, and what generalizable materials insights and electrochemical improvements result from it."
+DOCUMENT = load_document("cian.txt")
+USER_GOAL = "More specifically, I want to understand the architecture of a transformer, why it works, and why it was designed that way. Pages 3, 4, 5, 6 are quite mysterious to me."
 app = FastAPI()
 
 # Enable CORS

backend/documents/cian.txt

@@ -0,0 +1,324 @@
+  
+
+{0}------------------------------------------------
+
+Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or scholarly works.
+
+# **Attention Is All You Need**
+
+Ashish Vaswani<sup>\*</sup> Google Brain avaswani@google.com
+
+ $Llion Jones^*$ Google Research llion@google.com noam@google.com Aidan N. Gomez<sup>\*†</sup>
+
+University of Toronto
+
+aidan@cs.toronto.edu
+
+Noam Shazeer<sup>\*</sup>
+
+Google Brain
+
+Niki Parmar<sup>\*</sup> Google Research nikip@google.com
+
+Jakob Uszkoreit $^*$ Google Research usz@google.com
+
+Łukasz Kaiser<sup>\*</sup> Google Brain lukaszkaiser@google.com
+
+Illia Polosukhin<sup>\* ‡</sup> illia.polosukhin@gmail.com
+
+### Abstract
+
+The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 Englishto-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.
+
+<sup>\*</sup> Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention and the parameter-free position representation and became the other person involved in nearly every detail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and tensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.
+
+<sup>&</sup>lt;sup>†</sup>Work performed while at Google Brain.
+
+<sup>&</sup>lt;sup>‡</sup>Work performed while at Google Research.
+
+{1}------------------------------------------------
+
+## 1 Introduction
+
+Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15].
+
+Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states  $h_t$ , as a function of the previous hidden state  $h_{t-1}$  and the input for position t. This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. Recent work has achieved significant improvements in computational efficiency through factorization tricks [21] and conditional computation [32], while also improving model performance in case of the latter. The fundamental constraint of sequential computation, however, remains.
+
+Attention mechanisms have become an integral part of compelling sequence modeling and transduction models in various tasks, allowing modeling of dependencies without regard to their distance in the input or output sequences  $[2, 19]$ . In all but a few cases  $[27]$ , however, such attention mechanisms are used in conjunction with a recurrent network.
+
+In this work we propose the Transformer, a model architecture eschewing recurrence and instead relying entirely on an attention mechanism to draw global dependencies between input and output. The Transformer allows for significantly more parallelization and can reach a new state of the art in translation quality after being trained for as little as twelve hours on eight P100 GPUs.
+
+#### Background 2
+
+The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU [16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building block, computing hidden representations in parallel for all input and output positions. In these models, the number of operations required to relate signals from two arbitrary input or output positions grows in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes it more difficult to learn dependencies between distant positions [12]. In the Transformer this is reduced to a constant number of operations, albeit at the cost of reduced effective resolution due to averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as described in section  $3.2$ .
+
+Self-attention, sometimes called intra-attention is an attention mechanism relating different positions of a single sequence in order to compute a representation of the sequence. Self-attention has been used successfully in a variety of tasks including reading comprehension, abstractive summarization, textual entailment and learning task-independent sentence representations [4, 27, 28, 22].
+
+End-to-end memory networks are based on a recurrent attention mechanism instead of sequencealigned recurrence and have been shown to perform well on simple-language question answering and language modeling tasks [34].
+
+To the best of our knowledge, however, the Transformer is the first transduction model relying entirely on self-attention to compute representations of its input and output without using sequencealigned RNNs or convolution. In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9].
+
+## 3 Model Architecture
+
+Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35]. Here, the encoder maps an input sequence of symbol representations  $(x_1,...,x_n)$  to a sequence of continuous representations  $\mathbf{z} = (z_1, ..., z_n)$ . Given  $\mathbf{z}$ , the decoder then generates an output sequence  $(y_1, \ldots, y_m)$  of symbols one element at a time. At each step the model is auto-regressive [10], consuming the previously generated symbols as additional input when generating the next.
+
+{2}------------------------------------------------
+
+Image /page/2/Figure/0 description: The image shows a diagram of the Transformer model architecture. The diagram consists of two main parts: the encoder on the left and the decoder on the right. The encoder takes an input embedding and adds positional encoding to it. The result is then passed through a stack of N encoder layers. Each encoder layer consists of a multi-head attention sub-layer followed by a feed forward sub-layer. Both sub-layers have an add & norm operation around them. The decoder takes an output embedding and adds positional encoding to it. The result is then passed through a stack of N decoder layers. Each decoder layer consists of a masked multi-head attention sub-layer, a multi-head attention sub-layer, and a feed forward sub-layer. All sub-layers have an add & norm operation around them. The output of the decoder is passed through a linear layer and a softmax layer to produce the output probabilities.
+
+Figure 1: The Transformer - model architecture.
+
+The Transformer follows this overall architecture using stacked self-attention and point-wise, fully connected layers for both the encoder and decoder, shown in the left and right halves of Figure 1, respectively.
+
+#### 3.1 Encoder and Decoder Stacks
+
+**Encoder:** The encoder is composed of a stack of  $N = 6$  identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, positionwise fully connected feed-forward network. We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is LayerNorm $(x + \text{Sublayer}(x))$ , where Sublayer $(x)$  is the function implemented by the sub-layer itself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension  $d_{\text{model}} = 512$ .
+
+**Decoder:** The decoder is also composed of a stack of  $N = 6$  identical layers. In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization. We also modify the self-attention sub-layer in the decoder stack to prevent positions from attending to subsequent positions. This masking, combined with fact that the output embeddings are offset by one position, ensures that the predictions for position  $i$  can depend only on the known outputs at positions less than  $i$ .
+
+### 3.2 Attention
+
+An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum
+
+{3}------------------------------------------------
+
+Image /page/3/Figure/0 description: The image shows two diagrams, one titled "Scaled Dot-Product Attention" and the other "Multi-Head Attention". The "Scaled Dot-Product Attention" diagram shows a series of operations: MatMul, Scale, Mask (opt.), SoftMax, and MatMul, with inputs Q, K, and V. The "Multi-Head Attention" diagram shows multiple "Scaled Dot-Product Attention" blocks, each preceded by a "Linear" layer, and their outputs are concatenated using a "Concat" operation, followed by a "Linear" layer. The inputs to the "Linear" layers are V, K, and Q, respectively, and the number of "Scaled Dot-Product Attention" blocks is denoted by 'h'.
+
+Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel.
+
+of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.
+
+#### 3.2.1 Scaled Dot-Product Attention
+
+We call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of queries and keys of dimension  $d_k$ , and values of dimension  $d_v$ . We compute the dot products of the query with all keys, divide each by  $\sqrt{d_k}$ , and apply a softmax function to obtain the weights on the values.
+
+In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix  $Q$ . The keys and values are also packed together into matrices  $K$  and  $V$ . We compute the matrix of outputs as:
+
+$$Attention(Q, K, V) = softmax(\frac{QK^{T}}{\sqrt{d_{k}}})V$$
+(1)
+
+The two most commonly used attention functions are additive attention [2], and dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor of  $\frac{1}{\sqrt{d_k}}$ . Additive attention computes the compatibility function using a feed-forward network with a single hidden layer. While the two are similar in theoretical complexity, dot-product attention is much faster and more space-efficient in practice, since it can be implemented using highly optimized matrix multiplication code.
+
+While for small values of  $d_k$  the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of  $d_k$  [3]. We suspect that for large values of  $d_k$ , the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients<sup>4</sup>. To counteract this effect, we scale the dot products by  $\frac{1}{\sqrt{d}}$ .
+
+#### 3.2.2 Multi-Head Attention
+
+Instead of performing a single attention function with  $d_{\text{model}}$ -dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values  $h$  times with different, learned linear projections to  $d_k$ ,  $d_k$  and  $d_v$  dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding  $d_v$ -dimensional
+
+ ${}^{4}$ To illustrate why the dot products get large, assume that the components of  $q$  and  $k$  are independent random variables with mean 0 and variance 1. Then their dot product,  $q \cdot k = \sum_{i=1}^{d_k} q_i k_i$ , has mean 0 and variance  $d_k$ .
+
+{4}------------------------------------------------
+
+output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.
+
+Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this.
+
+$$\text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1, ..., \text{head}_h)W^O$$
+$$\text{where } \text{head}_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)$$
+
+Where the projections are parameter matrices  $W_i^Q \in \mathbb{R}^{d_{\text{model}} \times d_k}$ ,  $W_i^K \in \mathbb{R}^{d_{\text{model}} \times d_k}$ ,  $W_i^V \in \mathbb{R}^{d_{\text{model}} \times d_v}$ and  $W^O \in \mathbb{R}^{\check{h}d_v \times d_{\text{model}}}$ .
+
+In this work we employ  $h = 8$  parallel attention layers, or heads. For each of these we use  $d_k = d_v = d_{\text{model}}/h = 64$ . Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.
+
+#### 3.2.3 Applications of Attention in our Model
+
+The Transformer uses multi-head attention in three different ways:
+
+- In "encoder-decoder attention" layers, the queries come from the previous decoder layer, and the memory keys and values come from the output of the encoder. This allows every position in the decoder to attend over all positions in the input sequence. This mimics the typical encoder-decoder attention mechanisms in sequence-to-sequence models such as [38, 2, 9].
+- The encoder contains self-attention layers. In a self-attention layer all of the keys, values and queries come from the same place, in this case, the output of the previous layer in the encoder. Each position in the encoder can attend to all positions in the previous layer of the encoder.
+- Similarly, self-attention layers in the decoder allow each position in the decoder to attend to all positions in the decoder up to and including that position. We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to  $-\infty$ ) all values in the input of the softmax which correspond to illegal connections. See Figure 2.
+
+#### 3.3 Position-wise Feed-Forward Networks
+
+In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.
+
+$$FFN(x) = \max(0, xW_1 + b_1)W_2 + b_2 \tag{2}$$
+
+While the linear transformations are the same across different positions, they use different parameters from layer to layer. Another way of describing this is as two convolutions with kernel size 1. The dimensionality of input and output is  $d_{\text{model}} = 512$ , and the inner-layer has dimensionality  $d_{ff} = 2048.$ 
+
+#### 3.4 Embeddings and Softmax
+
+Similarly to other sequence transduction models, we use learned embeddings to convert the input tokens and output tokens to vectors of dimension  $d_{\text{model}}$ . We also use the usual learned linear transformation and softmax function to convert the decoder output to predicted next-token probabilities. In our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation, similar to [30]. In the embedding layers, we multiply those weights by  $\sqrt{d_{\text{model}}}$ .
+
+{5}------------------------------------------------
+
+Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types.  $n$  is the sequence length,  $d$  is the representation dimension,  $k$  is the kernel size of convolutions and  $r$  the size of the neighborhood in restricted self-attention.
+
+| Layer Type                  | Complexity per Layer     | Sequential<br>Operations | Maximum Path Length |
+|-----------------------------|--------------------------|--------------------------|---------------------|
+| Self-Attention              | $O(n^2 \cdot d)$         | $O(1)$                   | $O(1)$              |
+| Recurrent                   | $O(n \cdot d^2)$         | $O(n)$                   | $O(n)$              |
+| Convolutional               | $O(k \cdot n \cdot d^2)$ | $O(1)$                   | $O(log_k(n))$       |
+| Self-Attention (restricted) | $O(r \cdot n \cdot d)$   | O(1)                     | $O(n/r)$            |
+
+#### 3.5 Positional Encoding
+
+Since our model contains no recurrence and no convolution, in order for the model to make use of the order of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence. To this end, we add "positional encodings" to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension  $d_{\text{model}}$ as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].
+
+In this work, we use sine and cosine functions of different frequencies:
+
+$$PE_{(pos,2i)} = sin(pos/10000^{2i/d_{\text{model}}})$$
+$$PE_{(pos,2i+1)} = cos(pos/10000^{2i/d_{\text{model}}})$$
+
+where  $pos$  is the position and i is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from  $2\pi$  to  $10000 \cdot 2\pi$ . We chose this function because we hypothesized it would allow the model to easily learn to attend by relative positions, since for any fixed offset  $k$ ,  $PE_{pos+k}$  can be represented as a linear function of  $PE_{pos}$ .
+
+We also experimented with using learned positional embeddings [9] instead, and found that the two versions produced nearly identical results (see Table 3 row  $(\bar{E})$ ). We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training.
+
+#### Why Self-Attention
+
+In this section we compare various aspects of self-attention layers to the recurrent and convolutional layers commonly used for mapping one variable-length sequence of symbol representations  $(x_1,...,x_n)$  to another sequence of equal length  $(z_1,...,z_n)$ , with  $x_i,z_i\in\mathbb{R}^d$ , such as a hidden layer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we consider three desiderata.
+
+One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required.
+
+The third is the path length between long-range dependencies in the network. Learning long-range dependencies is a key challenge in many sequence transduction tasks. One key factor affecting the ability to learn such dependencies is the length of the paths forward and backward signals have to traverse in the network. The shorter these paths between any combination of positions in the input and output sequences, the easier it is to learn long-range dependencies [12]. Hence we also compare the maximum path length between any two input and output positions in networks composed of the different layer types.
+
+As noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially executed operations, whereas a recurrent layer requires  $O(n)$  sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence
+
+{6}------------------------------------------------
+
+length  $n$  is smaller than the representation dimensionality  $d$ , which is most often the case with sentence representations used by state-of-the-art models in machine translations, such as word-piece [38] and byte-pair [31] representations. To improve computational performance for tasks involving very long sequences, self-attention could be restricted to considering only a neighborhood of size  $r$  in the input sequence centered around the respective output position. This would increase the maximum path length to  $O(n/r)$ . We plan to investigate this approach further in future work.
+
+A single convolutional layer with kernel width  $k < n$  does not connect all pairs of input and output positions. Doing so requires a stack of  $O(n/k)$  convolutional layers in the case of contiguous kernels, or  $O(log_k(n))$  in the case of dilated convolutions [18], increasing the length of the longest paths between any two positions in the network. Convolutional layers are generally more expensive than recurrent layers, by a factor of  $k$ . Separable convolutions [6], however, decrease the complexity considerably, to  $O(k \cdot n \cdot d + n \cdot d^2)$ . Even with  $k = n$ , however, the complexity of a separable convolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer, the approach we take in our model.
+
+As side benefit, self-attention could yield more interpretable models. We inspect attention distributions from our models and present and discuss examples in the appendix. Not only do individual attention heads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic and semantic structure of the sentences.
+
+#### Training
+
+This section describes the training regime for our models.
+
+### 5.1 Training Data and Batching
+
+We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs. Sentences were encoded using byte-pair encoding [3], which has a shared sourcetarget vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT 2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece vocabulary [38]. Sentence pairs were batched together by approximate sequence length. Each training batch contained a set of sentence pairs containing approximately 25000 source tokens and 25000 target tokens.
+
+### 5.2 Hardware and Schedule
+
+We trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using the hyperparameters described throughout the paper, each training step took about  $0.4$  seconds. We trained the base models for a total of 100,000 steps or 12 hours. For our big models, (described on the bottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps  $(3.5 \text{ days}).$ 
+
+### 5.3 Optimizer
+
+We used the Adam optimizer [20] with  $\beta_1 = 0.9$ ,  $\beta_2 = 0.98$  and  $\epsilon = 10^{-9}$ . We varied the learning rate over the course of training, according to the formula:
+
+$$lrate = d_{\text{model}}^{-0.5} \cdot \min(\text{step\_num}^{-0.5}, \text{step\_num} \cdot \text{warmup\_steps}^{-1.5})$$
+ (3)
+
+This corresponds to increasing the learning rate linearly for the first  $warmup\_steps$  training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used  $warmup\_steps = 4000.$ 
+
+#### 5.4 Regularization
+
+We employ three types of regularization during training:
+
+{7}------------------------------------------------
+
+| Model                           | BLEU  |       | Training Cost (FLOPs) |                     |
+|---------------------------------|-------|-------|-----------------------|---------------------|
+|                                 | EN-DE | EN-FR | EN-DE                 | EN-FR               |
+| ByteNet [18]                    | 23.75 |       |                       |                     |
+| Deep-Att + PosUnk [39]          |       | 39.2  |                       | $1.0 \cdot 10^{20}$ |
+| GNMT + RL [38]                  | 24.6  | 39.92 | $2.3 \cdot 10^{19}$   | $1.4 \cdot 10^{20}$ |
+| ConvS2S [9]                     | 25.16 | 40.46 | $9.6 \cdot 10^{18}$   | $1.5 \cdot 10^{20}$ |
+| MoE [32]                        | 26.03 | 40.56 | $2.0 \cdot 10^{19}$   | $1.2 \cdot 10^{20}$ |
+| Deep-Att + PosUnk Ensemble [39] |       | 40.4  |                       | $8.0 \cdot 10^{20}$ |
+| GNMT + RL Ensemble [38]         | 26.30 | 41.16 | $1.8 \cdot 10^{20}$   | $1.1 \cdot 10^{21}$ |
+| ConvS2S Ensemble [9]            | 26.36 | 41.29 | $7.7 \cdot 10^{19}$   | $1.2 \cdot 10^{21}$ |
+| Transformer (base model)        | 27.3  | 38.1  | $3.3 \cdot 10^{18}$   |                     |
+| Transformer (big)               | 28.4  | 41.8  | $2.3 \cdot 10^{19}$   |                     |
+
+Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the English-to-German and English-to-French newstest2014 tests at a fraction of the training cost.
+
+**Residual Dropout** We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. For the base model, we use a rate of  $P_{drop} = 0.1.$ 
+
+**Label Smoothing** During training, we employed label smoothing of value  $\epsilon_{ls} = 0.1$  [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.
+
+## 6 Results
+
+### 6.1 Machine Translation
+
+On the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big) in Table 2) outperforms the best previously reported models (including ensembles) by more than  $2.0$ BLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is listed in the bottom line of Table 3. Training took 3.5 days on 8 P100 GPUs. Even our base model surpasses all previously published models and ensembles, at a fraction of the training cost of any of the competitive models.
+
+On the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of  $41.0$ , outperforming all of the previously published single models, at less than  $1/4$  the training cost of the previous state-of-the-art model. The Transformer (big) model trained for English-to-French used dropout rate  $P_{drop} = 0.1$ , instead of 0.3.
+
+For the base models, we used a single model obtained by averaging the last 5 checkpoints, which were written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We used beam search with a beam size of 4 and length penalty  $\alpha = 0.6$  [38]. These hyperparameters were chosen after experimentation on the development set. We set the maximum output length during inference to input length  $+ 50$ , but terminate early when possible [38].
+
+Table 2 summarizes our results and compares our translation quality and training costs to other model architectures from the literature. We estimate the number of floating point operations used to train a model by multiplying the training time, the number of GPUs used, and an estimate of the sustained single-precision floating-point capacity of each GPU $^5$ .
+
+#### 6.2 Model Variations
+
+To evaluate the importance of different components of the Transformer, we varied our base model in different ways, measuring the change in performance on English-to-German translation on the
+
+<sup>&</sup>lt;sup>5</sup>We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.
+
+{8}------------------------------------------------
+
+Table 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base model. All metrics are on the English-to-German translation development set, newstest2013. Listed perplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to per-word perplexities.
+
+|      | N | $d_{model}$ | $d_{ff}$                                  | h  | $d_k$ | $d_v$ | $P_{drop}$ | $\epsilon_{ls}$ | train steps | PPL (dev) | BLEU (dev) | params $\times 10^6$ |
+|------|---|-------------|-------------------------------------------|----|-------|-------|------------|-----------------|-------------|-----------|------------|----------------------|
+| base | 6 | 512         | 2048                                      | 8  | 64    | 64    | 0.1        | 0.1             | 100K        | 4.92      | 25.8       | 65                   |
+| (A)  |   |             |                                           | 1  | 512   | 512   |            |                 |             | 5.29      | 24.9       |                      |
+|      |   |             |                                           | 4  | 128   | 128   |            |                 |             | 5.00      | 25.5       |                      |
+|      |   |             |                                           | 16 | 32    | 32    |            |                 |             | 4.91      | 25.8       |                      |
+|      |   |             |                                           | 32 | 16    | 16    |            |                 |             | 5.01      | 25.4       |                      |
+| (B)  |   |             |                                           |    | 16    |       |            |                 |             | 5.16      | 25.1       | 58                   |
+|      |   |             |                                           |    | 32    |       |            |                 |             | 5.01      | 25.4       | 60                   |
+| (C)  | 2 |             |                                           |    |       |       |            |                 |             | 6.11      | 23.7       | 36                   |
+|      | 4 |             |                                           |    |       |       |            |                 |             | 5.19      | 25.3       | 50                   |
+|      | 8 |             |                                           |    |       |       |            |                 |             | 4.88      | 25.5       | 80                   |
+|      |   | 256         |                                           |    | 32    | 32    |            |                 |             | 5.75      | 24.5       | 28                   |
+|      |   | 1024        |                                           |    | 128   | 128   |            |                 |             | 4.66      | 26.0       | 168                  |
+|      |   |             | 1024                                      |    |       |       |            |                 |             | 5.12      | 25.4       | 53                   |
+|      |   |             | 4096                                      |    |       |       |            |                 |             | 4.75      | 26.2       | 90                   |
+| (D)  |   |             |                                           |    |       |       | 0.0        |                 |             | 5.77      | 24.6       |                      |
+|      |   |             |                                           |    |       |       | 0.2        |                 |             | 4.95      | 25.5       |                      |
+|      |   |             |                                           |    |       |       |            | 0.0             |             | 4.67      | 25.3       |                      |
+|      |   |             |                                           |    |       |       |            | 0.2             |             | 5.47      | 25.7       |                      |
+| (E)  |   |             | positional embedding instead of sinusoids |    |       |       |            |                 |             | 4.92      | 25.7       |                      |
+| big  | 6 | 1024        | 4096                                      | 16 |       |       | 0.3        |                 | 300K        | 4.33      | 26.4       | 213                  |
+
+development set, newstest2013. We used beam search as described in the previous section, but no checkpoint averaging. We present these results in Table 3.
+
+In Table 3 rows  $(A)$ , we vary the number of attention heads and the attention key and value dimensions, keeping the amount of computation constant, as described in Section 3.2.2. While single-head attention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.
+
+In Table 3 rows (B), we observe that reducing the attention key size  $d_k$  hurts model quality. This suggests that determining compatibility is not easy and that a more sophisticated compatibility function than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected, bigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our sinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical results to the base model.
+
+## 6.3 English Constituency Parsing
+
+To evaluate if the Transformer can generalize to other tasks we performed experiments on English constituency parsing. This task presents specific challenges: the output is subject to strong structural constraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence models have not been able to attain state-of-the-art results in small-data regimes [37].
+
+We trained a 4-layer transformer with  $d_{model} = 1024$  on the Wall Street Journal (WSJ) portion of the Penn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting, using the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences [37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens for the semi-supervised setting.
+
+We performed only a small number of experiments to select the dropout, both attention and residual (section 5.4), learning rates and beam size on the Section 22 development set, all other parameters remained unchanged from the English-to-German base translation model. During inference, we
+
+{9}------------------------------------------------
+
+| Parser                              | Training                 | WSJ 23 F1 |
+|-------------------------------------|--------------------------|-----------|
+| Vinyals & Kaiser el al. (2014) [37] | WSJ only, discriminative | 88.3      |
+| Petrov et al. (2006) [29]           | WSJ only, discriminative | 90.4      |
+| Zhu et al. (2013) [40]              | WSJ only, discriminative | 90.4      |
+| Dyer et al. (2016) [8]              | WSJ only, discriminative | 91.7      |
+| Transformer (4 layers)              | WSJ only, discriminative | 91.3      |
+| Zhu et al. (2013) [40]              | semi-supervised          | 91.3      |
+| Huang & Harper (2009) [14]          | semi-supervised          | 91.3      |
+| McClosky et al. (2006) [26]         | semi-supervised          | 92.1      |
+| Vinyals & Kaiser el al. (2014) [37] | semi-supervised          | 92.1      |
+| Transformer (4 layers)              | semi-supervised          | 92.7      |
+| Luong et al. (2015) [23]            | multi-task               | 93.0      |
+| Dyer et al. (2016) [8]              | generative               | 93.3      |
+
+Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)
+
+increased the maximum output length to input length + 300. We used a beam size of 21 and  $\alpha = 0.3$ for both WSJ only and the semi-supervised setting.
+
+Our results in Table 4 show that despite the lack of task-specific tuning our model performs surprisingly well, yielding better results than all previously reported models with the exception of the Recurrent Neural Network Grammar [8].
+
+In contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the Berkeley-Parser [29] even when training only on the WSJ training set of 40K sentences.
+
+## 7 Conclusion
+
+In this work, we presented the Transformer, the first sequence transduction model based entirely on attention, replacing the recurrent layers most commonly used in encoder-decoder architectures with multi-headed self-attention.
+
+For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014 English-to-French translation tasks, we achieve a new state of the art. In the former task our best model outperforms even all previously reported ensembles.
+
+We are excited about the future of attention-based models and plan to apply them to other tasks. We plan to extend the Transformer to problems involving input and output modalities other than text and to investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs such as images, audio and video. Making generation less sequential is another research goals of ours.
+
+The code we used to train and evaluate our models is available at https://github.com/ tensorflow/tensor2tensor.

frontend/src/hooks/useDocumentProcessor.js

@@ -47,36 +47,36 @@ export const useDocumentProcessor = () => {
             // Use hardcoded chunks for the document
             const hardcodedChunks = [
   {
-    "topic": "The Core Problem: Instability in Advanced Single-Crystal Cathodes",
-    "text": `INTRODUCTION\nWith the aim to establish a low-carbon and environmentally\nfriendly society, next-generation lithium-ion batteries (LIBs)\nare urgently required to promote the large-scale application of\nelectric vehicles. 1,2 Since Ni element dominates the reversible\ncapacity, Ni-rich layered oxides receive great expectations due\nto their exceptional advantages in energy density and\nproduction cost, especially for ultrahigh-nickel Li-\nNixCo,Mn1-x-yO2 (NCM, x ≥ 0.9).3−8 However, the conven-\ntional spherical NCM cathodes formed by agglomeration of\nprimary particles always suffer from severe structural\ndegradation with the generation of intergranular microcracks\nand detrimental phase transitions during cycling.9–12 Recently,\nthe NCM fabrication with quasi single-crystal morphology\n(SNCM) has attracted enormous research attention both in industry and laboratory, which is widely considered as a\npromising modification strategy. 13,14 The SNCM cathode\nconsists of single dispersed particles of 2−5 μm with an\nenhanced crystalline structure and internal boundary-free\nconfiguration. 15,16 Benefiting from its exceptional morphology,\nSNCM can significantly alleviate the accumulation of\nanisotropic stresses and intergranular microcrack formation,\nleading to a considerable enhancement in capacity retention and cycled particle integrity.17,18 Meanwhile, its excellent\nmechanical strength is beneficial in further increasing the\ncompaction density of cathodes to achieve higher specific\nenergy density with relative low cost.19,20\nWhile the transformation of agglomerated nanoparticles into\nseparate micron-sized single-crystals presents numerous\nadvantages, it also unveils some inherent challenges.21 The\nlower specific surface area and prolonged Li+ diffusion distance\nin single-crystal particles can contribute to the sluggish Li+\ndiffusion kinetics, resulting in poor capacity and inferior rate\nperformance of SNCM.22−24 This will further trigger\ninhomogeneous distribution of Li+ concentration during\ncycling, causing the coexistence of multiple phases with severe\nlattice strain within single-crystals.25 As a result, lattice strain\nexacerbates planar gliding and intragranular cracks generation\nof microsized single-crystals, which has been considered as the\nroot cause of surface degradation and structural damage.26,27\nTherefore, the poor Li+ diffusion kinetics is the origin of various degradation mechanisms in SNCM and urgently needs\nto be improved to advance its application potential.`
+    "topic": "The Fundamental Motivation: Dispensing with Recurrence",
+    "text": "Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5].\n\nRecurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states  $h_t$ , as a function of the previous hidden state  $h_{t-1}$  and the input for position t. This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. The fundamental constraint of sequential computation, however, remains.\n\nIn this work we propose the Transformer, a model architecture eschewing recurrence and instead relying entirely on an attention mechanism to draw global dependencies between input and output."
   },
   {
-    "topic": "The Conventional Solution and Its Limitations: Dry Doping",
-    "text": "Element doping has been widely confirmed as a promising\napproach to enhance the Li+ diffusion kinetics and structural\nstability of SNCM.28−34 However, most current research has\npredominantly focused on the dry-doping method, where\ndopants are added during the mixing stage between lithium\nand SNCM precursors.16,35 For this approach, certain\nsubstitution elements face challenges in incorporating into\nthe cathode lattice due to their large atomic mass and limited\nsolid solubility,31 resulting in an inhomogeneous distribution\nwithin the bulk and causing internal structural degradation in\nSNCM. Moreover, the dopants tend to react with the lithium\nsource at the surface to generate a nonuniformly distributed\nimpurity layer that obstructs Li+ diffusion channels. Overly\nincreasing the sintering time and temperature to promote the\ndiffusion of doping ions may lead to severe Li/Ni disorder-\ning.36"
+    "topic": "The Overall Architecture: Encoder and Decoder Stacks",
+    "text": "The Transformer follows this overall architecture using stacked self-attention and point-wise, fully connected layers for both the encoder and decoder, shown in the left and right halves of Figure 1, respectively.\n\n**Encoder:** The encoder is composed of a stack of  $N = 6$  identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, positionwise fully connected feed-forward network. We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is LayerNorm $(x + \\text{Sublayer}(x))$ , where Sublayer $(x)$  is the function implemented by the sub-layer itself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension  $d_{\\text{model}} = 512$ .\n\n**Decoder:** The decoder is also composed of a stack of  $N = 6$  identical layers. In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization. We also modify the self-attention sub-layer in the decoder stack to prevent positions from attending to subsequent positions. This masking, combined with fact that the output embeddings are offset by one position, ensures that the predictions for position  $i$  can depend only on the known outputs at positions less than  $i$ ."
   },
   {
-    "topic": "The Innovation: An 'Inside-Out' In Situ Doping Strategy",
-    "text": "In contrast, the in situ doping method, where dopants\nand the main elements (Ni, Co, and Mn) are simultaneously added during the coprecipitation reaction for precursor\npreparation, can result in a homogeneous modification effect\nfrom the inside out and effectively address aforementioned\nproblems.9,37 Although the in situ doping possesses obvious\nmerits for holistically enhancing the Li+ diffusion kinetics of\nSNCM, there is a high threshold for attaining the desired\neffect. The choice of dopant needs to take into account its\nsolubility and thermodynamic diffusion properties. Meanwhile,\nthe amount of dopant addition requires precise regulation to\nprevent affecting the supersaturation within the reaction\nsystem and thus producing low-quality precursors.\nIn this work, we thoroughly investigated the in situ doping\ntechnology to achieve the optimal dopant element (Nb, Zr, W,\netc), doping amount and reaction conditions, enabling precise\ncontrol over the preparation of in situ doped single-crystal\nLiNi0.92Co0.03Mn0.05O2 (SNCM). It is confirmed that in situ\ndoping offers distinct advantages in the successful incorpo-\nration and uniform dispersion of dopants into the precursor\nlattice, resulting in modulated precursor primary particle\nmorphology."
+    "topic": "Attention Mechanism: Scaled Dot-Product Attention and the Need for Scaling",
+    "text": "An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.\n\nWe call our particular attention \"Scaled Dot-Product Attention\" (Figure 2). The input consists of queries and keys of dimension  $d_k$ , and values of dimension  $d_v$ . We compute the matrix of outputs as:\n\n$$Attention(Q, K, V) = softmax(\\frac{QK^{T}}{\\sqrt{d_{k}}})V$$\n(1)\n\nDot-product attention is much faster and more space-efficient in practice, since it can be implemented using highly optimized matrix multiplication code.\n\nWhile for small values of  $d_k$  the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of  $d_k$  [3]. We suspect that for large values of  $d_k$ , the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients$^{4}$. To counteract this effect, we scale the dot products by  $\\frac{1}{\\sqrt{d}}$ . (To illustrate why the dot products get large, assume that the components of  $q$  and  $k$  are independent random variables with mean 0 and variance 1. Then their dot product,  $q \\cdot k = \\sum_{i=1}^{d_k} q_i k_i$ , has mean 0 and variance  $d_k$ .)"
   },
   {
-    "topic": "Mechanism: How In Situ Doping Engineers a Better Precursor",
-    "text": "A schematic diagram is provided to better understand the\nmechanism of coprecipitation reaction and the underlying\nreason for the morphological variation (Figure 1g). During the\npreparation of precursors, the main transition metal ions (Ni2+,\nMn2+, Co2+) in the system rapidly combine with hydroxide and\nprecipitate to form massive crystal nuclei.42,43 Subsequently,\nthe massive crystal nuclei promptly bind ions from the solution\nto form primary particles. In general, the {010} planes combine\nions faster than other planes due to its higher surface\nenergy,41,44 thus resulting in that the primary particles prefer\nto grow horizontally to obtain high exposure {001} planes.45\nMeanwhile, the {001} planes with lower surface energy can\nmerely agglomerate fewer stacked layers in the longitudinal\ndirection. However, when Nb-ions are continuously added into\nthe reaction system, the {010} planes will interact with doping\nions predominantly because Nb5+ are subjected to stronger\nelectrostatic adsorption force due to their more positive\ncharges than Ni2+, Co2+ and Mn2+. The substantial amounts of\nNb-ions accumulated on {010} planes exhibit an inhibiting\neffect on its horizontal growth, resulting in the morphology of\nhigh exposure {010} planes. Furthermore, the surface energy of\n{001} planes relatively get raised, thus tending to agglomerate\nmore longitudinal stacked layers. It is obvious that all these characterizations substantiate the controlled morphology\nmodulation and homogeneous Nb doping of NCMOH-Nb,\ndemonstrating the success of precursor engineering for in situ\ndoping. Notably, considering that the cathodes incline to well\ninherit the morphological structure of precursors, both the\nhigh exposure {010} planes and more longitudinal stacked\nlayers of NCMOH-Nb precursor are favorable to provide more\nLi+ channels and enhanced diffusion kinetics for SNCM-Nb-\nwet cathode.46,47"
+    "topic": "Multi-Head Attention: Capturing Diverse Subspaces",
+    "text": "Instead of performing a single attention function with  $d_{\\text{model}}$ -dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values  $h$  times with different, learned linear projections to  $d_k$ ,  $d_k$  and  $d_v$  dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding  $d_v$ -dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.\n\nMulti-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this.\n\n$$\\text{MultiHead}(Q, K, V) = \\text{Concat}(\\text{head}_1, ..., \\text{head}_h)W^O$$\n$$\\text{where } \\text{head}_i = \\text{Attention}(QW_i^Q, KW_i^K, VW_i^V)$$\n\nIn this work we employ  $h = 8$  parallel attention layers, or heads. For each of these we use  $d_k = d_v = d_{\\text{model}}/h = 64$ . Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality."
   },
   {
-    "topic": "Result 1: Improved Structural Stability and Phase Transition",
-    "text": "Effects of In Situ Doping on Structural Stability and\nLi+ Diffusion Kinetics. To comprehensively investigate the\neffects of dry and wet Nb doping on the structural evolution\nand phase transition reversibility during Li+ (de)intercalation,\nin situ XRD measurements were conducted with a potential\nrange of 2.5−4.5 V at 0.2 C. [...] Figure 3a exhibits a\nmagnified view of the (003) peak and the corresponding lattice\nparameter variations during the initial cycle. When charging\nfrom 2.5 to 4.18 V, the (003) diffraction peaks of all cathodes\nare slightly shifted to a smaller 2θ angle. With further charging\nto 4.5 V, the (003) diffraction peaks shift significantly to a\nlarger 2θ angle, illustrating that the cathodes encounter a\nsudden decay of c-axis parameters during the H2−H3 phase\ntransition.53,54 Because it is difficult to fully recover, H2−H3 is\ngenerally considered to be an irreversible phase transition. Its\nshift angle is a sign of structural stability for the SNCM\ncathode.55,56 In comparison, it is revealed that SNCM-Nb-wet\nfeatures the minimum H2−H3 phase transition angle of 1.11°,\noutperforming the 1.23° of SNCM-Nb-dry and 1.30° of\nSNCM-pure, which indicates the advantages of in situ doping\nfor suppressing the dramatic contraction of the lattice\nparameters. [...] In comparison, SNCM-Nb-wet experiences a\nsmooth H2−H3 phase transition accompanied by a confined\nwidth and intensity change during the entire charging/\ndischarging process (Figure 3d). Simultaneously, its structural\ntransition is fully reversible without residual phases. This\nphenomenon is due to that the internal homogeneous Nb5+\nmodification formed by in situ doping can effectively build\nstronger metal-oxygen bonds and stabilize the crystal\nstructure,48,58 thus alleviating the Li+ migration hindrance\nand hysteresis caused by the long transport distance inside the\nsingle-crystal particles. Thus, SNCM-Nb-wet obtained a\nsmoother phase transition and improved Li+ diffusion kinetics."
+    "topic": "Applications of Attention: The Three Uses in the Transformer",
+    "text": "The Transformer uses multi-head attention in three different ways:\n\n- In \"encoder-decoder attention\" layers, the queries come from the previous decoder layer, and the memory keys and values come from the output of the encoder. This allows every position in the decoder to attend over all positions in the input sequence. This mimics the typical encoder-decoder attention mechanisms in sequence-to-sequence models such as [38, 2, 9].\n- The encoder contains self-attention layers. In a self-attention layer all of the keys, values and queries come from the same place, in this case, the output of the previous layer in the encoder. Each position in the encoder can attend to all positions in the previous layer of the encoder.\n- Similarly, self-attention layers in the decoder allow each position in the decoder to attend to all positions in the decoder up to and including that position. We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to  $-\\infty$ ) all values in the input of the softmax which correspond to illegal connections."
   },
   {
-    "topic": "Result 2: Superior Electrochemical Performance and Cycling Stability",
-    "text": "Comparison of Electrochemical Performances of Dry/\nIn Situ Doped Cathodes. After demonstrating that in situ\ndoping is beneficial in providing excellent structural stability and lithium-ion diffusion kinetics, we performed electro-\nchemical measurements under different operating temperatures\nin order to verify the advantages it brings to lithium storage\nperformance. As illustrated in Figure 4a, the initial discharge\ncapacities of Nb-doped cathodes at 1 C (SNCM-Nb-dry: 184.1\nmAh g−1, SNCM-Nb-wet: 182.7 mAh g−1) appear slightly\nlower than that of pristine cathode (SNCM-pure: 187.6 mAh\ng−1) due to the doping of inert Nb-ions.62 Whereas, SNCM-\nNb-wet displays the best cycling stability among these three\nsingle-crystal cathodes, with an excellent cycling retention of\n76.1% after 400 cycles. In comparison, SNCM-pure and\nSNCM-Nb-dry display a consistent capacity decay, showing\ncycling retentions of 56.8% and 65.1%, respectively. [...] Benefiting from the\nimproved Li+ diffusion kinetics inside the micron-sized\nsingle-crystal particles by in situ doping, SNCM-Nb-wet still\ncan exhibit enhanced cycling stability at high rate of 5 C with\n76.5% after 200 cycles, significantly superior to 64.6% of\nSNCM-Nb-dry and 44.2% of SNCM-pure (Figure 4c).\nFurthermore, increasing the operating temperature of Ni-\nrich SNCM is commonly regarded as an effective method to\nimprove the discharge specific capacities, but it generally\ncauses thermal safety issues during the practical application.69\nConsequently, the prepared samples are further evaluated at\nincreased temperatures up to 45 °C to demonstrate the\nadvantages of in situ doping. Figure 4d displays the cycling\nperformances of pristine and Nb-doped cathodes at 1 C and 45\n°C. SNCM-Nb-wet exhibits an outstanding capacity stability of\n75.5% after 200 cycles, obviously better than 63.6% of SNCM-\nNb-dry and 51.2% of SNCM-pure."
+    "topic": "The Position-Wise Feed-Forward Network Component",
+    "text": "In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.\n\n$$FFN(x) = \\max(0, xW_1 + b_1)W_2 + b_2 \\tag{2}$$\n\nWhile the linear transformations are the same across different positions, they use different parameters from layer to layer. Another way of describing this is as two convolutions with kernel size 1. The dimensionality of input and output is  $d_{\\text{model}} = 512$ , and the inner-layer has dimensionality  $d_{ff} = 2048.$"
   },
   {
-    "topic": "Visualizing the Mechanism: Uniform Li+ Distribution Prevents Cracks",
-    "text": "Simulation of Modified Mechanism in In Situ Doped\nCathodes. According to the aforementioned discussions, in\nsitu doped cathodes show a smoother phase transition process,\nwhich may lead to a more uniform internal Li+ concentration\ndistribution and superior Li+ diffusion kinetics. To visualize\nthis modification effect, COMSOL software is utilized to\nsimulate the internal Li+ concentration distribution [...]. Subsequently, along\nwith the detachment of Li+, SNCM-Nb-dry (Figure 5a) has an\naggravated Li+ concentration gradient distribution at all voltage\nstates. This corresponds to the severe two-phase coexistence\nobserved in the in situ XRD pattern, which will generate\ninhomogeneous intragranular stresses, eventually leading to\nstructural degradation. In contrast, SNCM-Nb-wet (Figure 5b)\ndisplays a relatively uniform and moderate Li+ concentration\nvariation during the whole charging/discharge process, which\nis beneficial for the single-crystal cathode to maintain the\nstructural stability throughout the long-term cycling. [...] To explore the modification mechanism of in\nsitu doping on structure stability, the morphology and\nmicrostructure evolution of SNCM-Nb-dry and SNCM-Nb-\nwet were detected by SEM and HRTEM after 100 cycles at 5\nC. As seen in Figure 6a, the internal cross-sectional structure of\nSNCM-Nb-dry is damaged after cycling and contains a\nmultitude of cracks and holes. [...] As for SNCM-Nb-\nwet, it can be discovered that after cycles the cross-sectional\nmorphology remains dense and integrated (Figure 6c). The\nHRTEM image of SNCM-Nb-wet (Figure 6d) also confirms\nthat the cathode maintains the intact R3m layered structure at\nthe particle surface[...]."
+    "topic": "Addressing Sequence Order: Positional Encoding",
+    "text": "Since our model contains no recurrence and no convolution, in order for the model to make use of the order of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence. To this end, we add \"positional encodings\" to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension  $d_{\\text{model}}$ as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].\n\nIn this work, we use sine and cosine functions of different frequencies:\n\n$$PE_{(pos,2i)} = sin(pos/10000^{2i/d_{\\text{model}}})$$\n$$PE_{(pos,2i+1)} = cos(pos/10000^{2i/d_{\\text{model}}})$$\n\nWe chose this fixed sinusoidal function because we hypothesized it would allow the model to easily learn to attend by relative positions, since for any fixed offset  $k$ ,  $PE_{pos+k}$  can be represented as a linear function of  $PE_{pos}$ . We also found that the sinusoidal version produced nearly identical results to learned positional embeddings, but may allow the model to extrapolate to sequence lengths longer than the ones encountered during training."
   },
   {
-    "topic": "Generalizing the Strategy and Summarizing the Mechanism",
-    "text": "In Situ Doping Promotion for Other Elements. More\nimportantly, we introduce dopants Zr4+ and W6+ in order to\nfurther demonstrate the advantages of in situ doping on the\ncathode properties. The dry and in situ doped cathodes are\nobtained by the same coprecipitation and two-stage calcination process. According to Figure 7a and Figure S24, the Zr-doped\nmaterials exhibit similar single-crystal particle morphology. [...] As\nseen in Figure 7g, SNCM-Zr-wet presents the best cycling\nretention with 90.8% after 100 cycles, obviously better than 80.5% of SNCM-Zr-dry and 69.9% of SNCM-pure. [...] Based\non the successful attempt of various doping ions, it is evident\nthat in situ doping achieves a uniform and consistent\nmodification effect within single-crystal particles and demon-\nstrates a certain applicability for different doping elements\n(Nb, Zr, W, etc).\nAnalysis of In Situ Doping Modification Mechanism.\n[...] In\ncomparison, the in situ doped cathodes benefited from the\nuniform distribution of doping ions and can substantially\naccelerate the Li+ diffusion kinetics and migration rate from the\ninside out. Furthermore, the fluent Li+ transportation results in\na uniform phase transition process and optimized Li+\nconcentration distribution, which can ensure the reversibility\nof the H2/H3 transition and alleviate the inner stress without\nthe formation of intragranular cracks. Hence, the in situ doped\ncathode maintains an integrated structure and excellent\nelectrochemical performances after long-term cycling."
+    "topic": "Why Self-Attention Works: Complexity and Path Length Advantages",
+    "text": "Motivating our use of self-attention we consider three desiderata. One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required. The third is the path length between long-range dependencies in the network. The shorter these paths between any combination of positions in the input and output sequences, the easier it is to learn long-range dependencies [12].\n\nAs noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially executed operations ($O(1)$), whereas a recurrent layer requires  $O(n)$  sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence length  $n$  is smaller than the representation dimensionality  $d$ , which is most often the case with sentence representations. Learning long-range dependencies is a key challenge, and Self-Attention reduces the maximum path length between any two positions to $O(1)$, compared to $O(n)$ for Recurrent Networks and $O(\log_k(n))$ for Convolutional Networks."
   }
 ];