Add references and fix square brackets

contents/optimizations/optimizations.qmd

@@ -541,7 +541,7 @@ Upon deciding the type of clipping range, it is essential to tighten the range t
 ![Illustration of the main forms of quantization granularities. In layerwise quantization, the same clipping range is applied to all filters which belong to the same layer. Notice how this can result in lower quantization resolutions for channels with narrow distributions, e.g. Filter 1, Filter 2, and Filter C. A higher quantization resolution can be achieved using channelwise quantization which dedicates different clipping ranges to different channels (@surveyofquant).](images/png/efficientnumerics_granularity.png)
 
 1. Layerwise Quantization: This approach determines the clipping range by considering all of the weights in the convolutional filters of a layer. Then, the same clipping range is used for all convolutional filters. It's the simplest to implement, and, as such, it often results in sub-optimal accuracy due the wide variety of differing ranges between filters. For example, a convolutional kernel with a narrower range of parameters loses its quantization resolution due to another kernel in the same layer having a wider range.
-2. Groupwise Quantization: This approach groups different channels inside a layer to calculate the clipping range. This method can be helpful when the distribution of parameters across a single convolution/activation varies a lot. In practice, this method was useful in Q-BERT [Q-BERT: Hessian based ultra low precision quantization of bert] for quantizing Transformer [Attention Is All You Need] models that consist of fully-connected attention layers. The downside with this approach comes with the extra cost of accounting for different scaling factors.
+2. Groupwise Quantization: This approach groups different channels inside a layer to calculate the clipping range. This method can be helpful when the distribution of parameters across a single convolution/activation varies a lot. In practice, this method was useful in Q-BERT [@DBLP:journals/corr/abs-1909-05840] for quantizing Transformer [@vaswani2017attention] models that consist of fully-connected attention layers. The downside with this approach comes with the extra cost of accounting for different scaling factors.
 3. Channelwise Quantization: This popular method uses a fixed range for each convolutional filter that is independent of other channels. Because each channel is assigned a dedicated scaling factor, this method ensures a higher quantization resolution and often results in higher accuracy.
 4. Sub-channelwise Quantization: Taking channelwise quantization to the extreme, this method determines the clipping range with respect to any groups of parameters in a convolution or fully-connected layer. It may result in considerable overhead since different scaling factors need to be taken into account when processing a single convolution or fully-connected layer.