Add new terms to the glossary (#20051)

* add new terms

* apply review

docs/source/en/glossary.mdx

@@ -17,7 +17,7 @@ documentation.
 
 ## A
 
-### Attention mask
+### attention mask
 
 The attention mask is an optional argument used when batching sequences together.
 
@@ -75,23 +75,40 @@ by the tokenizer under the key "attention_mask":
 
 ### autoencoding models 
 
-see [MLM](#mlm)
+see [masked language modeling](#masked-language-modeling)
 
 ### autoregressive models
 
-see [CLM](#clm)
+see [causal language modeling](#causal-language-modeling)
+
+## B
+
+### backbone
+
+The backbone is the network (embeddings and layers) that outputs the raw hidden states or features. It is usually connected to a [head](#head) which accepts the features as its input to make a prediction. For example, [`ViTModel`] is a backbone without a specific head on top. Other models can also use [`VitModel`] as a backbone such as [DPT](model_doc/dpt).
 
 ## C
 
-### CLM
+### channel
+
+Color images are made up of some combination of values in three channels - red, green, and blue (RGB) - and grayscale images only have one channel. In 🤗 Transformers, the channel can be the first or last dimension of an image's tensor: [`n_channels`, `height`, `width`] or [`height`, `width`, `n_channels`].
+
+### causal language modeling
+
+A pretraining task where the model reads the texts in order and has to predict the next word. It's usually done by
+reading the whole sentence but using a mask inside the model to hide the future tokens at a certain timestep.
+
+### connectionist temporal classification (CTC)
 
-Causal language modeling, a pretraining task where the model reads the texts in order and has to predict the next word.
-It's usually done by reading the whole sentence but using a mask inside the model to hide the future tokens at a
-certain timestep.
+An algorithm which allows a model to learn without knowing exactly how the input and output are aligned; CTC calculates the distribution of all possible outputs for a given input and chooses the most likely output from it. CTC is commonly used in speech recognition tasks because speech doesn't always cleanly align with the transcript for a variety of reasons such as a speaker's different speech rates.
+
+### convolution
+
+A type of layer in a neural network where the input matrix is multiplied element-wise by a smaller matrix (kernel or filter) and the values are summed up in a new matrix. This is known as a convolutional operation which is repeated over the entire input matrix. Each operation is applied to a different segment of the input matrix. Convolutional neural networks (CNNs) are commonly used in computer vision.
 
 ## D
 
-### Decoder input IDs
+### decoder input IDs
 
 This input is specific to encoder-decoder models, and contains the input IDs that will be fed to the decoder. These
 inputs should be used for sequence to sequence tasks, such as translation or summarization, and are usually built in a
@@ -108,7 +125,7 @@ Machine learning algorithms which uses neural networks with several layers.
 
 ## F
 
-### Feed Forward Chunking
+### feed forward chunking
 
 In each residual attention block in transformers the self-attention layer is usually followed by 2 feed forward layers.
 The intermediate embedding size of the feed forward layers is often bigger than the hidden size of the model (e.g., for
@@ -127,12 +144,26 @@ For models employing the function [`apply_chunking_to_forward`], the `chunk_size
 embeddings that are computed in parallel and thus defines the trade-off between memory and time complexity. If
 `chunk_size` is set to 0, no feed forward chunking is done.
 
+## H
+
+### head
+
+The model head refers to the last layer of a neural network that accepts the raw hidden states and projects them onto a different dimension. There is a different model head for each task. For example:
+
+  * [`GPT2ForSequenceClassification`] is a sequence classification head - a linear layer - on top of the base [`GPT2Model`].
+  * [`ViTForImageClassification`] is an image classification head - a linear layer on top of the final hidden state of the `CLS` token - on top of the base [`ViTModel`].
+  * [`Wav2Vec2ForCTC`] ia a language modeling head with [CTC](#connectionist-temporal-classification-(CTC)) on top of the base [`Wav2Vec2Model`].
+
 ## I
 
-### Input IDs
+### image patch
+
+Vision-based Transformers models split an image into smaller patches which are linearly embedded, and then passed as a sequence to the model. You can find the `patch_size` - or resolution - of the model in it's configuration.
 
-The input ids are often the only required parameters to be passed to the model as input. *They are token indices,
-numerical representations of tokens building the sequences that will be used as input by the model*.
+### input IDs
+
+The input ids are often the only required parameters to be passed to the model as input. They are token indices,
+numerical representations of tokens building the sequences that will be used as input by the model.
 
 <Youtube id="VFp38yj8h3A"/>
 
@@ -171,7 +202,7 @@ Tokenizers](https://github.com/huggingface/tokenizers) for peak performance.
 ```
 
 The tokenizer returns a dictionary with all the arguments necessary for its corresponding model to work properly. The
-token indices are under the key "input_ids":
+token indices are under the key `input_ids`:
 
 ```python
 >>> encoded_sequence = inputs["input_ids"]
@@ -199,7 +230,7 @@ because this is the way a [`BertModel`] is going to expect its inputs.
 
 ## L
 
-### Labels
+### labels
 
 The labels are an optional argument which can be passed in order for the model to compute the loss itself. These labels
 should be the expected prediction of the model: it will use the standard loss in order to compute the loss between its
@@ -207,28 +238,34 @@ predictions and the expected value (the label).
 
 These labels are different according to the model head, for example:
 
-- For sequence classification models (e.g., [`BertForSequenceClassification`]), the model expects a tensor of dimension
+- For sequence classification models ([`BertForSequenceClassification`]), the model expects a tensor of dimension
   `(batch_size)` with each value of the batch corresponding to the expected label of the entire sequence.
-- For token classification models (e.g., [`BertForTokenClassification`]), the model expects a tensor of dimension
+- For token classification models ([`BertForTokenClassification`]), the model expects a tensor of dimension
   `(batch_size, seq_length)` with each value corresponding to the expected label of each individual token.
-- For masked language modeling (e.g., [`BertForMaskedLM`]), the model expects a tensor of dimension `(batch_size,
+- For masked language modeling ([`BertForMaskedLM`]), the model expects a tensor of dimension `(batch_size,
   seq_length)` with each value corresponding to the expected label of each individual token: the labels being the token
   ID for the masked token, and values to be ignored for the rest (usually -100).
-- For sequence to sequence tasks,(e.g., [`BartForConditionalGeneration`], [`MBartForConditionalGeneration`]), the model
+- For sequence to sequence tasks,([`BartForConditionalGeneration`], [`MBartForConditionalGeneration`]), the model
   expects a tensor of dimension `(batch_size, tgt_seq_length)` with each value corresponding to the target sequences
-  associated with each input sequence. During training, both *BART* and *T5* will make the appropriate
-  *decoder_input_ids* and decoder attention masks internally. They usually do not need to be supplied. This does not
-  apply to models leveraging the Encoder-Decoder framework. See the documentation of each model for more information on
-  each specific model's labels.
+  associated with each input sequence. During training, both BART and T5 will make the appropriate
+  `decoder_input_ids` and decoder attention masks internally. They usually do not need to be supplied. This does not
+  apply to models leveraging the Encoder-Decoder framework. 
+  
+<Tip>
+
+Each model's labels may be different, so be sure to always check the documentation of each model for more information
+about their specific labels!
+
+</Tip>
 
-The base models (e.g., [`BertModel`]) do not accept labels, as these are the base transformer models, simply outputting
+The base models ([`BertModel`]) do not accept labels, as these are the base transformer models, simply outputting
 features.
 
 ## M
 
-### MLM
+### masked language modeling
 
-Masked language modeling, a pretraining task where the model sees a corrupted version of the texts, usually done by
+A pretraining task where the model sees a corrupted version of the texts, usually done by
 masking some tokens randomly, and has to predict the original text.
 
 ### multimodal
@@ -237,22 +274,30 @@ A task that combines texts with another kind of inputs (for instance images).
 
 ## N
 
-### NLG
+### Natural language generation
 
-Natural language generation, all tasks related to generating text (for instance talk with transformers, translation).
+All tasks related to generating text (for instance talk with transformers, translation).
 
-### NLP
+### Natural language processing
 
-Natural language processing, a generic way to say "deal with texts".
+A generic way to say "deal with texts".
 
-### NLU
+### Natural language understanding
 
-Natural language understanding, all tasks related to understanding what is in a text (for instance classifying the
+All tasks related to understanding what is in a text (for instance classifying the
 whole text, individual words).
 
 ## P
 
-### Position IDs
+### pixel values
+
+A tensor of the numerical representations of an image that is passed to a model. The pixel values have a shape of [`batch_size`, `num_channels`, `height`, `width`], and are generated from a feature extractor.
+
+### pooling
+
+An operation that reduces a matrix into a smaller matrix, either by taking the maximum or average of the pooled dimension(s). Pooling layers are commonly found between convolutional layers to downsample the feature representation.
+
+### position IDs
 
 Contrary to RNNs that have the position of each token embedded within them, transformers are unaware of the position of
 each token. Therefore, the position IDs (`position_ids`) are used by the model to identify each token's position in the
@@ -268,26 +313,37 @@ other types of positional embeddings, such as sinusoidal position embeddings or
 ### pretrained model
 
 A model that has been pretrained on some data (for instance all of Wikipedia). Pretraining methods involve a
-self-supervised objective, which can be reading the text and trying to predict the next word (see CLM) or masking some
-words and trying to predict them (see MLM).
+self-supervised objective, which can be reading the text and trying to predict the next word (see [causal language
+modeling](#causal-language-modeling)) or masking some words and trying to predict them (see [masked language
+modeling](#masked-language-modeling)). 
+
+  Speech and vision models have their own pretraining objectives. For example, Wav2Vec2 is a speech model pretrained on a contrastive task which requires the model to identify the "true" speech representation from a set of "false" speech representations. On the other hand, BEiT is a vision model pretrained on a masked image modeling task which masks some of the image patches and requires the model to predict the masked patches (similar to the masked language modeling objective).
 
 ## R
 
-### RNN
+### recurrent neural network
 
-Recurrent neural network, a type of model that uses a loop over a layer to process texts.
+A type of model that uses a loop over a layer to process texts.
 
 ## S
 
+### sampling rate
+
+A measurement in hertz of the number of samples (the audio signal) taken per second. The sampling rate is a result of discretizing a continuous signal such as speech.
+
 ### self-attention
 
 Each element of the input finds out which other elements of the input they should attend to.
 
-### seq2seq or sequence-to-sequence
+### sequence-to-sequence (seq2seq)
 
 Models that generate a new sequence from an input, like translation models, or summarization models (such as
 [Bart](model_doc/bart) or [T5](model_doc/t5)).
 
+### stride
+
+In [convolution](#convolution) or [pooling](#pooling), the stride refers to the distance the kernel is moved over a matrix. A stride of 1 means the kernel is moved one pixel over at a time, and a stride of 2 means the kernel is moved two pixels over at a time.
+
 ## T
 
 ### token
@@ -295,7 +351,7 @@ Models that generate a new sequence from an input, like translation models, or s
 A part of a sentence, usually a word, but can also be a subword (non-common words are often split in subwords) or a
 punctuation symbol.
 
-### Token Type IDs
+### token Type IDs
 
 Some models' purpose is to do classification on pairs of sentences or question answering.