Knowledge Distillation

Contrastive Language-Image Pre-training (CLIP), consisting of a simplified version of ConVIRT trained from scratch, is an efficient method of image representation learning from natural language supervision. , CLIP jointly trains an image encoder and a text encoder to predict the correct pairings of a batch of (image, text) training examples. At test time the learned text encoder synthesizes a zero-shot linear classifier by embedding the names or descriptions of the target dataset’s classes. For pre-training, CLIP is trained to predict which of the $N X N$ possible (image, text) pairings across a batch actually occurred. CLIP learns a multi-modal embedding space by jointly training an image encoder and text encoder to maximize the cosine similarity of the image and text embeddings of the $N$ real pairs in the batch while minimizing the cosine similarity of the embeddings of the $N^2 - N$ incorrect pairings. A symmetric cross entropy loss is optimized over these similarity scores. Image credit: Learning Transferable Visual Models From Natural Language Supervision

A very simple way to improve the performance of almost any machine learning algorithm is to train many different models on the same data and then to average their predictions. Unfortunately, making predictions using a whole ensemble of models is cumbersome and may be too computationally expensive to allow deployment to a large number of users, especially if the individual models are large neural nets. Caruana and his collaborators have shown that it is possible to compress the knowledge in an ensemble into a single model which is much easier to deploy and we develop this approach further using a different compression technique. We achieve some surprising results on MNIST and we show that we can significantly improve the acoustic model of a heavily used commercial system by distilling the knowledge in an ensemble of models into a single model. We also introduce a new type of ensemble composed of one or more full models and many specialist models which learn to distinguish fine-grained classes that the full models confuse. Unlike a mixture of experts, these specialist models can be trained rapidly and in parallel. Source: Distilling the Knowledge in a Neural Network

A Focal Loss function addresses class imbalance during training in tasks like object detection. Focal loss applies a modulating term to the cross entropy loss in order to focus learning on hard misclassified examples. It is a dynamically scaled cross entropy loss, where the scaling factor decays to zero as confidence in the correct class increases. Intuitively, this scaling factor can automatically down-weight the contribution of easy examples during training and rapidly focus the model on hard examples. Formally, the Focal Loss adds a factor $(1 - p\_{t})^\gamma$ to the standard cross entropy criterion. Setting $\gamma>0$ reduces the relative loss for well-classified examples ($p\_{t}>.5$), putting more focus on hard, misclassified examples. Here there is tunable focusing parameter $\gamma \ge 0$. $$ {\text{FL}(p\_{t}) = - (1 - p\_{t})^\gamma \log\left(p\_{t}\right)} $$

Shrink and Fine-Tune, or SFT, is a type of distillation that avoids explicit distillation by copying parameters to a student student model and then fine-tuning. Specifically it extracts a student model from the maximally spaced layers of a fine-tuned teacher. Each layer $l \in L'$ is copied fully from $L$. For example, when creating a BART student with 3 decoder layers from the 12 encoder layer 12 decoder layer teacher, we copy the teacher’s full $Enc^{L}$ and decoder layers 0, 6, and 11 to the student. When deciding which layers to copy, we break ties arbitrarily; copying layers 0, 5, and 11 might work just as well. When copy only 1 decoder layer, we copy layer 0. This was found this to work better than copying layer 11. The impact of initialization on performance is measured experimentally in Section 6.1. After initialization, the student model continues to fine-tune on the summarization dataset, with the objective of minimizing $\mathcal{L}\_{Data}$.

Laplacian eigenvectors represent a natural generalization of the Transformer positional encodings (PE) for graphs as the eigenvectors of a discrete line (NLP graph) are the cosine and sinusoidal functions. They help encode distance-aware information (i.e., nearby nodes have similar positional features and farther nodes have dissimilar positional features). Hence, Laplacian Positional Encoding (PE) is a general method to encode node positions in a graph. For each node, its Laplacian PE is the k smallest non-trivial eigenvectors.

RetinaNet is a one-stage object detection model that utilizes a focal loss function to address class imbalance during training. Focal loss applies a modulating term to the cross entropy loss in order to focus learning on hard negative examples. RetinaNet is a single, unified network composed of a backbone network and two task-specific subnetworks. The backbone is responsible for computing a convolutional feature map over an entire input image and is an off-the-shelf convolutional network. The first subnet performs convolutional object classification on the backbone's output; the second subnet performs convolutional bounding box regression. The two subnetworks feature a simple design that the authors propose specifically for one-stage, dense detection. We can see the motivation for focal loss by comparing with two-stage object detectors. Here class imbalance is addressed by a two-stage cascade and sampling heuristics. The proposal stage (e.g., Selective Search, EdgeBoxes, DeepMask, RPN) rapidly narrows down the number of candidate object locations to a small number (e.g., 1-2k), filtering out most background samples. In the second classification stage, sampling heuristics, such as a fixed foreground-to-background ratio, or online hard example mining (OHEM), are performed to maintain a manageable balance between foreground and background. In contrast, a one-stage detector must process a much larger set of candidate object locations regularly sampled across an image. To tackle this, RetinaNet uses a focal loss function, a dynamically scaled cross entropy loss, where the scaling factor decays to zero as confidence in the correct class increases. Intuitively, this scaling factor can automatically down-weight the contribution of easy examples during training and rapidly focus the model on hard examples. Formally, the Focal Loss adds a factor $(1 - p\_{t})^\gamma$ to the standard cross entropy criterion. Setting $\gamma>0$ reduces the relative loss for well-classified examples ($p\_{t}>.5$), putting more focus on hard, misclassified examples. Here there is tunable focusing parameter $\gamma \ge 0$. $$ {\text{FL}(p\_{t}) = - (1 - p\_{t})^\gamma \log\left(p\_{t}\right)} $$

FastSpeech2 is a text-to-speech model that aims to improve upon FastSpeech by better solving the one-to-many mapping problem in TTS, i.e., multiple speech variations corresponding to the same text. It attempts to solve this problem by 1) directly training the model with ground-truth target instead of the simplified output from teacher, and 2) introducing more variation information of speech (e.g., pitch, energy and more accurate duration) as conditional inputs. Specifically, in FastSpeech 2, we extract duration, pitch and energy from speech waveform and directly take them as conditional inputs in training and use predicted values in inference. The encoder converts the phoneme embedding sequence into the phoneme hidden sequence, and then the variance adaptor adds different variance information such as duration, pitch and energy into the hidden sequence, finally the mel-spectrogram decoder converts the adapted hidden sequence into mel-spectrogram sequence in parallel. FastSpeech 2 uses a feed-forward Transformer block, which is a stack of self-attention and 1D-convolution as in FastSpeech, as the basic structure for the encoder and mel-spectrogram decoder.

Local Prior Matching is a semi-supervised objective for speech recognition that distills knowledge from a strong prior (e.g. a language model) to provide learning signal to a discriminative model trained on unlabeled speech. The LPM objective minimizes the cross entropy between the local prior and the model distribution, and is minimized when $q\_{y\mid{x}} = \bar{p}\_{y\mid{x}}$. Intuitively, LPM encourages the ASR model to assign posterior probabilities proportional to the linguistic probabilities of the proposed hypotheses.

Collaborative Distillation is a new knowledge distillation method (named Collaborative Distillation) for encoder-decoder based neural style transfer to reduce the number of convolutional filters. The main idea is underpinned by a finding that the encoder-decoder pairs construct an exclusive collaborative relationship, which is regarded as a new kind of knowledge for style transfer models.

Teacher-Tutor-Student Knowledge Distillation is a method for image virtual try-on models. It treats fake images produced by the parser-based method as "tutor knowledge", where the artifacts can be corrected by real "teacher knowledge", which is extracted from the real person images in a self-supervised way. Other than using real images as supervisions, knowledge distillation is formulated in the try-on problem as distilling the appearance flows between the person image and the garment image, enabling the finding of dense correspondences between them to produce high-quality results.

The NA method can be divided into two steps: (i) Training a neural network approximation of f , and (ii) inference of xˆ. Step (i) is conventional and involves training a generic neural network on a dataset ˆ of input/output pairs from the simulator, denoted D, resulting in f, an approximation of the forward ˆ model. This is illustrated in the left inset of Fig 1. In step (ii), our goal is to use ∂f/∂x to help us gradually adjust x so that we achieve a desired output of the forward model, y. This is similar to many classical inverse modeling approaches, such as the popular Adjoint method [8, 9]. For many practical ˆ expression for the simulator, from which it is trivial to compute ∂f/∂x, and furthermore, we can use modern deep learning software packages to efficiently estimate gradients, given a loss function L. More formally, let y be our target output, and let xˆi be our current estimate of the solution, where i indexes each solution we obtain in an iterative gradient-based estimation procedure. Then we compute xˆi+1 with inverse problems, however, obtaining ∂f/∂x requires significant expertise and/or effort, making these approaches challenging. Crucially, fˆ from step (i) provides us with a closed-form differentiable

Learning From Multiple Experts is a self-paced knowledge distillation framework that aggregates the knowledge from multiple 'Experts' to learn a unified student model. Specifically, the proposed framework involves two levels of adaptive learning schedules: Self-paced Expert Selection and Curriculum Instance Selection, so that the knowledge is adaptively transferred to the 'Student'. The self-paced expert selection automatically controls the impact of knowledge distillation from each expert, so that the learned student model will gradually acquire the knowledge from the experts, and finally exceed the expert. The curriculum instance selection, on the other hand, designs a curriculum for the unified model where the training samples are organized from easy to hard, so that the unified student model will receive a less challenging learning schedule, and gradually learns from easy to hard samples.

Margin Rectified Linear Unit, or Margin ReLU, is a type of activation function based on a ReLU, but it has a negative threshold for negative values instead of a zero threshhold.

PocketNet is a face recognition model family discovered through neural architecture search. The training is based on multi-step knowledge distillation.

TernaryBERT is a Transformer-based model which ternarizes the weights of a pretrained BERT model to $\{-1,0,+1\}$, with different granularities for word embedding and weights in the Transformer layer. Instead of directly using knowledge distillation to compress a model, it is used to improve the performance of ternarized student model with the same size as the teacher model. In this way, we transfer the knowledge from the highly-accurate teacher model to the ternarized student model with smaller capacity.

FastSpeech 2s is a text-to-speech model that abandons mel-spectrograms as intermediate output completely and directly generates speech waveform from text during inference. In other words there is no cascaded mel-spectrogram generation (acoustic model) and waveform generation (vocoder). FastSpeech 2s generates waveform conditioning on intermediate hidden, which makes it more compact in inference by discarding the mel-spectrogram decoder. Two main design changes are made to the waveform decoder. First, considering that the phase information is difficult to predict using a variance predictor, adversarial training is used in the waveform decoder to force it to implicitly recover the phase information by itself. Secondly, the mel-spectrogram decoder of FastSpeech 2 is leveraged, which is trained on the full text sequence to help on the text feature extraction. As shown in the Figure, the waveform decoder is based on the structure of WaveNet including non-causal convolutions and gated activation. The waveform decoder takes a sliced hidden sequence corresponding to a short audio clip as input and upsamples it with transposed 1D-convolution to match the length of audio clip. The discriminator in the adversarial training adopts the same structure in Parallel WaveGAN, which consists of ten layers of non-causal dilated 1-D convolutions with leaky ReLU activation function. The waveform decoder is optimized by the multi-resolution STFT loss and the LSGAN discriminator loss following Parallel WaveGAN. In inference, the mel-spectrogram decoder is discarded and only the waveform decoder is used to synthesize speech audio.

MT-PET is a multi-task version of Pattern Exploiting Training (PET) for exaggeration detection, which leverages knowledge from complementary cloze-style QA tasks to improve few-shot learning. It defines pairs of complementary pattern-verbalizer pairs for a main task and auxiliary task. These PVPs are then used to train PET on data from both tasks. PET uses the masked language modeling objective of pretrained language models to transform a task into one or more cloze-style question answering tasks. In the original PET implementation, PVPs are defined for a single target task. MT-PET extends this by allowing for auxiliary PVPs from related tasks, adding complementary cloze-style QA tasks during training. The motivation for the multi-task approach is two-fold: 1) complementary cloze-style tasks can potentially help the model to learn different aspects of the main task, i.e. the similar tasks of exaggeration detection and claim strength prediction; 2) data on related tasks can be utilized during training, which is important in situations where data for the main task is limited.

Semi-Supervised Knowledge Distillation is a type of knowledge distillation for person re-identification that exploits weakly annotated data by assigning soft pseudo labels to YouTube-Human to improve models' generalization ability. SSKD first trains a student model (e.g. ResNet-50) and a teacher model (e.g. ResNet-101) using labeled data from multi-source domain datasets. Then, SSKD develops an auxiliary classifier to imitate the soft predictions of unlabeled data generated by the teacher model. Meanwhile, the student model is also supervised by hard labels and predicted soft labels by the teacher model for labeled data.