The pipeline we propose consists of three parts: 1) recognizing the interaction type; 2) detecting the object that the interaction is targeting; and 3) learning incrementally the models from data recorded by the robot sensors. Our main contributions lie in the target object detection, guided by the recognized interaction, and in the incremental object learning. The novelty of our approach is the focus on natural, heterogeneous, and multimodal HRIs to incrementally learn new object models.

HRI pipeline

SoftPool: a fast and efficient method that sums exponentially weighted activations. Compared to a range of other pooling methods, SoftPool retains more information in the downsampled activation maps. More refined downsampling leads to better classification accuracy.

SoftPool

Refining activation downsampling with SoftPool

**NAS-FCOS** consists of two sub networks, an [FPN](https://paperswithcode.com/method/fpn) $f$ and a set of prediction heads $h$ which have shared structures. One notable difference with other FPN-based one-stage detectors is that our heads have partially shared weights. Only the last several layers of the predictions heads (marked as yellow) are tied by their weights. The number of layers to share is decided automatically by the search algorithm. Note that both FPN and head are in our actual search space; and have more layers than shown in this figure.

NAS-FCOS

NAS-FCOS: Fast Neural Architecture Search for Object Detection

**CenterPoint** is a two-stage 3D detector that finds centers of objects and their properties using a keypoint detector and regresses to other attributes, including 3D size, 3D orientation and velocity. In a second-stage, it refines these estimates using additional point features on the object. CenterPoint uses a standard Lidar-based backbone network, i.e., VoxelNet or PointPillars, to build a representation of the input point-cloud. CenterPoint predicts the relative offset (velocity) of objects between consecutive frames, which are then linked up greedily -- so in Centerpoint, 3D object tracking simplifies to greedy closest-point matching.

CenterPoint

Center-based 3D Object Detection and Tracking

**YOLOv1** is a single-stage object detection model. Object detection is framed as a regression problem to spatially separated bounding boxes and associated class probabilities. A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation. Since the whole detection pipeline is a single network, it can be optimized end-to-end directly on detection performance. 

The network uses features from the entire image to predict each bounding box. It also predicts all bounding boxes across all classes for an image simultaneously. This means the network reasons globally about the full image and all the objects in the image.

YOLOv1

You Only Look Once: Unified, Real-Time Object Detection

**GPT** is a [Transformer](https://paperswithcode.com/method/transformer)-based architecture and training procedure for natural language processing tasks. Training follows a two-stage procedure. First, a language modeling objective is used on
the unlabeled data to learn the initial parameters of a neural network model. Subsequently, these parameters are adapted to a target task using the corresponding supervised objective.

Improving Language Understanding by Generative Pre-Training

A **Spatial Transformer** is an image model block that explicitly allows the spatial manipulation of data within a [convolutional neural network](https://paperswithcode.com/methods/category/convolutional-neural-networks). It gives CNNs the ability to actively spatially transform feature maps, conditional on the feature map itself, without any extra training supervision or modification to the optimisation process. Unlike pooling layers, where the receptive fields are fixed and local, the spatial transformer module is a dynamic mechanism that can actively spatially transform an image (or a feature map) by producing an appropriate transformation for each input sample. The transformation is then performed on the entire feature map (non-locally) and can include scaling, cropping, rotations, as well as non-rigid deformations.

The architecture is shown in the Figure to the right. The input feature map $U$ is passed to a localisation network which regresses the transformation parameters $\theta$. The regular spatial grid $G$ over $V$ is transformed to the sampling grid $T\_{\theta}\left(G\right)$, which is applied to $U$, producing the warped output feature map $V$. The combination of the localisation network and sampling mechanism defines a spatial transformer.

Spatial Transformer

Spatial Transformer Networks

We present TabPFN, a trained Transformer that can do supervised classification for small tabular datasets in less than a second, needs no hyperparameter tuning and is competitive with state-of-the-art classification methods. TabPFN is fully entailed in the weights of our network, which accepts training and test samples as a set-valued input and yields predictions for the entire test set in a single forward pass. TabPFN is a Prior-Data Fitted Network (PFN) and is trained offline once, to approximate Bayesian inference on synthetic datasets drawn from our prior. This prior incorporates ideas from causal reasoning: It entails a large space of structural causal models with a preference for simple structures. On the 18 datasets in the OpenML-CC18 suite that contain up to 1 000 training data points, up to 100 purely numerical features without missing values, and up to 10 classes, we show that our method clearly outperforms boosted trees and performs on par with complex state-of-the-art AutoML systems with up to 230× speedup. This increases to a 5 700× speedup when using a GPU. We also validate these results on an additional 67 small numerical datasets from OpenML. We provide all our code, the trained TabPFN, an interactive browser demo and a Colab notebook at https://github.com/automl/TabPFN.

TABPFN

TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second

**WaveGrad** is a conditional model for waveform generation through estimating gradients of the data density. This model is built on the prior work on score matching and diffusion probabilistic models. It starts from Gaussian white noise and iteratively refines the signal via a gradient-based sampler conditioned on the mel-spectrogram. WaveGrad is non-autoregressive, and requires only a constant number of generation steps during inference. It can use as few as 6 iterations to generate high fidelity audio samples.

WaveGrad

WaveGrad: Estimating Gradients for Waveform Generation

**BiGG** is an autoregressive model for generative modeling for sparse graphs. It utilizes sparsity to avoid generating the full adjacency matrix, and reduces the graph generation time complexity to $O(((n + m)\log n)$. Furthermore, during training this autoregressive model can be parallelized with $O(\log n)$ synchronization stages, which makes it much more efficient than other autoregressive models that require $\Omega(n)$. The approach is based on three key elements: (1) an $O(\log n)$ process for generating each edge using a binary tree data structure, inspired by R-MAT; (2) a tree-structured autoregressive model for generating the set of edges associated with each node; and (3) an autoregressive model defined over the sequence of nodes.

BiGG

Scalable Deep Generative Modeling for Sparse Graphs

**wav2vec-U** is an unsupervised method to train speech recognition models without any labeled data. It leverages self-supervised speech representations to segment unlabeled language and learn a mapping from these representations to phonemes via adversarial training. 

Specifically, we learn self-supervised representations with wav2vec 2.0 on unlabeled speech audio, then identify clusters in the representations with k-means to segment the audio data. Next, we build segment representations by mean pooling the wav2vec 2.0 representations, performing [PCA](https://paperswithcode.com/method/pca) and a second mean pooling step between adjacent segments. This is input to the generator which outputs a phoneme sequence that is fed to the discriminator, similar to phonemized unlabeled text to perform adversarial training.

wav2vec-U

Unsupervised Speech Recognition

**MacBERT** is a [Transformer](https://paperswithcode.com/methods/category/transformers)-based model for Chinese NLP that alters [RoBERTa](https://paperswithcode.com/method/roberta) in several ways, including a modified masking strategy. Instead of masking with [MASK] token, which never appears in the fine-tuning stage, MacBERT masks the word with its similar word. Specifically MacBERT shares the same pre-training tasks as [BERT](https://paperswithcode.com/method/bert) with several modifications. For the MLM task, the following modifications are performed:

- Whole word masking is used as well as Ngram masking strategies for selecting candidate tokens for masking, with a percentage of
40%, 30%, 20%, 10% for word-level unigram to 4-gram.
- Instead of masking with [MASK] token, which never appears in the fine-tuning stage, similar words are used for the masking purpose. A similar word is obtained by using Synonyms toolkit which is based on word2vec similarity calculations. If an N-gram is selected to mask, we will find similar words individually. In rare cases, when there is no similar word, we will degrade to use random word replacement.
- A percentage of 15% input words is used for masking, where 80% will replace with similar words, 10% replace with a random word, and keep with original words for the rest of 10%.

MacBERT

Revisiting Pre-Trained Models for Chinese Natural Language Processing

**Natural Gradient Descent** is an approximate second-order optimisation method. It has an interpretation as optimizing over a Riemannian manifold using an intrinsic distance metric, which implies the updates are invariant to transformations such as whitening. By using the positive semi-definite (PSD) Gauss-Newton matrix to approximate the (possibly negative definite) Hessian, NGD can often work better than exact second-order methods.

Given the gradient of $z$, $g = \frac{\delta{f}\left(z\right)}{\delta{z}}$, NGD computes the update as:

$$\Delta{z} = \alpha{F}^{−1}g$$

where the Fisher information matrix $F$ is defined as:

$$ F = \mathbb{E}\_{p\left(t\mid{z}\right)}\left[\nabla\ln{p}\left(t\mid{z}\right)\nabla\ln{p}\left(t\mid{z}\right)^{T}\right] $$

The log-likelihood function $\ln{p}\left(t\mid{z}\right)$ typically corresponds to commonly used error functions such as the cross entropy loss.

Source: [LOGAN](https://paperswithcode.com/method/logan)

Image: [Fast Convergence of Natural Gradient Descent for Overparameterized Neural Networks
](https://arxiv.org/abs/1905.10961)

Natural Gradient Descent

**NPID++** (Non-Parametric Instance Discrimination) is a self-supervision approach that takes a non-parametric classification approach. It approves upon [NPID](https://paperswithcode.com/method/npid) by using more negative samples and training for more epochs.

NPID++

Self-Supervised Learning of Pretext-Invariant Representations

A **FuseFormer block** is used in the [FuseFormer](https://paperswithcode.com/method/fuseformer) model for video inpainting. It is the same to standard [Transformer](https://paperswithcode.com/method/transformer) block except that feed forward network is replaced with a Fusion Feed Forward Network (F3N). F3N brings no extra parameter into the standard feed forward net and the difference is that F3N inserts a soft-split and a soft composite operation between the two layer of MLPs.

FuseFormer Block

FuseFormer: Fusing Fine-Grained Information in Transformers for Video Inpainting

A **Variational Autoencoder** is a type of likelihood-based generative model. It consists of an encoder, that takes in data $x$ as input and transforms this into a latent representation $z$,  and a decoder, that takes a latent representation $z$ and returns a reconstruction $\hat{x}$. Inference is performed via variational inference to approximate the posterior of the model.

Auto-Encoding Variational Bayes

**Sarsa_INLINE_MATH_1** extends eligibility-traces to action-value methods. It has the same update rule as for **TD_INLINE_MATH_1** but we use the action-value form of the TD erorr:

$$ \delta\_{t} = R\_{t+1} + \gamma\hat{q}\left(S\_{t+1}, A\_{t+1}, \mathbb{w}\_{t}\right) - \hat{q}\left(S\_{t}, A\_{t}, \mathbb{w}\_{t}\right) $$

and the action-value form of the [eligibility trace](https://paperswithcode.com/method/eligibility-trace):

$$ \mathbb{z}\_{-1} = \mathbb{0} $$

$$ \mathbb{z}\_{t} = \gamma\lambda\mathbb{z}\_{t-1} + \nabla\hat{q}\left(S\_{t}, A\_{t}, \mathbb{w}\_{t} \right), 0 \leq t \leq T$$

Source: Sutton and Barto, Reinforcement Learning, 2nd Edition

Sarsa Lambda

RelDiff generates entity-relation-entity embeddings in a single embedding space. RelDiff adopts two fundamental vector algebraic operators to transform entity and relation embeddings from knowledge graphs into entity-relation-entity embeddings. In particular, RelDiff can encode finer-grained information about the relations than is captured when separate embeddings are learned for the entities and the relations.

AI Glossary

Human Robot Interaction Pipeline

Soft Pooling