Loops

In this section are displayed the loops pertaining the articles that present applications of AI promising automation or enhancements in human labor through AI collaboration or AI support.

A total of 35 articles are analyzed in detail to map the continuous loops of interaction between humans and AI that are described by the researchers. These loops are then represented consistently to provide a comparable view through schemas.

M

Making Table Understanding Work in Practice

Data analysisDigital tracesHuman collaborationAutomationNot mentioned

  1. algorithm SIGMATYPER
    Matches each column name to the labels in the type ontology using syntactic and semantic matching.
  2. human User
    Reviews the matches and provides feedback on any incorrect matches.
  3. algorithm SIGMATYPER
    Uses an embedding of the entire table to calculate the cosine similarity between the column names and semantic types.
  4. human User
    Reviews the predicted semantic types and provides feedback on any incorrect predictions.
  5. algorithm SIGMATYPER
    Infers labeling functions from data used to generate new training data and prediction functions.
  6. human User
    Reviews the inferred labeling functions and provides feedback on any incorrect inferences.
  7. algorithm SIGMATYPER
    Generates new training data using data programming and the inferred labeling functions.
  8. human User
    Reviews the new training data and provides feedback on any incorrect data.
  9. algorithm SIGMATYPER
    Trains a local model on the new training data.
  10. human User
    Reviews the performance of the local model and provides feedback on any incorrect predictions.
  11. algorithm SIGMATYPER
    Uses the local model to make predictions on new data.
  12. human User
    Reviews the predictions and provides feedback on any incorrect predictions.
H

High-quality Conversational Systems

Quality controlLogsAutomationImplicit AnnotatorsEnd-UserAI Experts

  1. human Chatbot Trainer/Designer
    Humans create training data by labeling utterances with intent categories used to train the NLU system
  2. human User
    The User initiates a conversation with the Chatbot
  3. algorithm NLU System
    The NLU system classifies User queries into intent categories using natural language understanding techniques
  4. algorithm Dialogue System
    The dialogue system navigates the conversation to its successful completion using rule-based graph and learning elements
  5. human Chatbot Trainer/Designer
    Humans detect drift in the chatbot system by reviewing chat logs and identifying new topics or changes in User behavior, which is used to update the system
  6. human Chatbot Trainer/Designer
    Humans identify intent design issues and suggest improvements, ensuring the system's effectiveness
  7. human Chatbot Trainer/Designer
    Humans review and confirm the actionable insights provided by the chatbot system, ensuring the suggested changes are appropriate before implementation
  8. algorithm Chatbot System
    The chatbot system identifies challenges in providing a good response and may hand off the conversation to a human agent or raise a ticket for human intervention, aiming to provide consistent and suitable responses while delivering a high-quality customer experience
S

Society-in-the-Loop: Programming the Algorithmic Social Contract

OpaquenessNot mentionedAcceptanceHuman collaborationStakeholders

  1. human Stakeholders
    Humans and Stakeholders negotiate the values and goals that AI systems should strive towards, considering trade-offs between different societal interests and ethical considerations.
  2. human Institutions
    Institutions and tools are developed to program the algorithmic social contract between humans and governance algorithms, ensuring that AI systems align with societal values and norms.
  3. human Experts
    Mechanisms are established to debug and monitor the behavior of AI systems, allowing for transparency, fairness, and accountability in the governance of autonomous machines.
  4. human Policymakers/ethicists
    Policymakers and ethicists play a crucial role in overseeing the implementation of the algorithmic social contract, ensuring that AI systems operate in accordance with societal expectations.
  5. human Policymakers/public
    New metrics and methods are developed to evaluate AI behavior against quantifiable human values, enabling Policymakers and the public to articulate their expectations to machines.
  6. algorithm AI systems
    Algorithms must incorporate the values and goals negotiated by humans and Stakeholders, considering trade-offs between different societal interests and ethical considerations.
  7. algorithm AI systems
    Algorithms must operate transparently, allowing for debugging and monitoring of their behavior to ensure that they align with the algorithmic social contract.
  8. algorithm AI systems
    Algorithms must comply with the algorithmic social contract that is programmed and monitored by Institutions and tools, ensuring that they operate in accordance with societal expectations.
  9. algorithm AI systems
    Algorithms must adapt to human values, as new metrics and methods are developed to evaluate their behavior against quantifiable human values.
M

MCU-Net: A framework towards uncertainty representations for decision support system patient referrals in healthcare contexts

Lack of dataBodily tracesAutomationAccuracyNot mentioned

  1. human Curator
    Curate and label medical images for training and testing.
  2. algorithm MCU-Net model
    Generate stochastic predictions using Monte Carlo Dropout, capturing aleatoric and epistemic uncertainties in the segmentation outputs.
  3. human Expert
    Evaluate the uncertainty metrics (aleatoric uncertainty, epistemic uncertainty, entropy, and mutual information) to understand the model's confidence in its predictions.
  4. algorithm MCU-Net model
    Apply an uncertainty threshold (τ) to the model's predictions, determining the level of uncertainty for each case.
  5. human Expert
    Define and validate the threshold τ based on domain knowledge and the desired balance between automated decision-making and human intervention.
  6. algorithm MCU-Net model
    Flag cases with high uncertainty, surpassing the defined threshold τ, for referral to medical professionals.
  7. human Expert
    Review flagged cases, leveraging their Expertise to assess and make decisions based on the uncertain predictions provided by the model.
  8. human Expert
    Provide feedback and corrective actions based on the model's uncertain predictions, ensuring that critical cases are appropriately flagged for further assessment.
  9. algorithm MCU-Net model
    Incorporate human feedback to improve the model's performance and refine the uncertainty thresholding process.
U

Understanding Questions that Arise When Working with Business Documents

Lack of dataWritten textUser dataHuman collaborationAuthorsBusinessmen

  1. human Participant
    Participants submit questions and share links to documents through a Microsoft Word add-in.
  2. algorithm AI model
    The system uses an AI model to extract passages from documents that may contain answers to the questions.
  3. human Knowledge worker
    Workers determine whether an answer to the question can be provided at all, and place the question in different queues based on this criterion. Workers assign finer-grained tags to questions and iterate over tags already assigned to prior questions as new ones come in.
  4. algorithm ML Q&A model
    The system routes questions to the appropriate respondents, either the AI model or human Knowledge workers.
  5. human Knowledge worker
    For questions about document content, workers manually access the document using the share link and their personal credentials. Workers copy and paste the question and document content into a custom UI front-end for the ML Q&A model. If the AI-provided answer is unsatisfactory upon, workers provide a human response.
  6. human Participant
    The system returns the answer to the participant through the Microsoft Word add-in.
W

What can AI do for me: Evaluating Machine Learning Interpretations in Cooperative Play

RigorWritten textAccuracyDomain ExpertsNon-Expert User

  1. human Player
    The human Player forms a team with a computer teammate.
  2. human Player
    The human Player is in charge of the team.
  3. algorithm computer
    The computer periodically updates its guesses and interpretations (every 4 words in the experiments described in the PDF).
  4. human Player
    At any point before the question is fully read, the human can decide to buzz, interrupt the readout, and provide an answer.
  5. algorithm computer
    The interpretations provided by the computer should help the human better decide whether to trust the computer's prediction or not.
C

Cogment: Open Source Framework For Distributed Multi-actor Training, Deployment & Operations

ComplexityNot mentionedOpen sourceHuman supportAI ExpertsGeneric humans

  1. human General
    Humans act in the environment on par with agents to ensure safe exploratory actions in sensitive contexts like autonomous driving.
  2. human General
    Humans provide rewards for several learning algorithms, for example in the context of evaluating machine-generated dialogues, summaries, semantic parsers, natural language, machine translation, and many others.
  3. algorithm learning
    Agents learn from human demonstrations under the imitation learning (IL) paradigm when it is challenging to design a reward function or when the reward function could be sparse, thus making it hard for an RL agent to learn.
  4. human General
    Humans generate tasks for agents to achieve.
  5. algorithm learning
    Agents learn from the tasks generated by humans.
  6. algorithm learning
    If no human input is possible, a different AI agent can be used to generate the curriculum.
E

Esports Agents with a Theory of Mind: Towards Better Engagement, Education, and Engineering

Data analysisBehavioral dataOptimizationHuman supportSport Players

  1. algorithm Computational model
    The algorithm processes observed behaviors and generates data representations of Player actions within the game environment.
  2. human Player
    Players interact with the system by providing labels for behaviors, indicating the time and spatial contexts of their actions within the game environment.
  3. algorithm Computational model
    The algorithm integrates the labeled behaviors provided by Players into the computational models of esports Players, incorporating qualitative and quantitative inputs.
  4. human Player
    Players engage with the system to interactively correct the computational model's probabilities as well as nodes that make up the graphical model, contributing to a more accurate representation of Player cognition.
  5. algorithm Computational model
    The algorithm utilizes the corrected computational model to refine the understanding of Player intents, strategies, and tactics within the gaming environment.
  6. human Player
    Players provide insights and feedback that contribute to the development of AI agents acting as intelligent tutoring systems for esports, enabling personalized coaching and gameplay experiences.
I

IGLU 2022: Interactive Grounded Language Understanding in a Collaborative Environment at NeurIPS 2022

Human interpretationNot mentionedHuman supportEnd-User

  1. algorithm data collection algorithm
    Using a single-turn data collection strategy to increase the speed of data collection.
  2. human Researcher
    Providing natural language instructions to the algorithm for task completion.
  3. algorithm training algorithm
    Using a new gridworld environment for fast and scalable experiments.
  4. human Researcher
    Collaborating with the algorithm to provide feedback and guidance during the training process.
  5. algorithm NLP algorithm
    Tackling the generation of clarifying questions for truly interactive agents.
  6. human evaluator
    Interacting with the algorithm to provide feedback on the effectiveness of the generated clarifying questions.
I

Integrative Imaging Informatics for Cancer Research: Workflow Automation for Neuro-oncology (I3CR-WANO)

HealthManual laborBodily tracesAutomationMedical Experts

  1. human Clinician
    Obtain DICOM format MRI data from patients with high- and low-grade gliomas.
  2. algorithm AI system
    Classify MRI sequences using an ensemble of NLP and CNN models. Preprocess the MRI data, including image registration, skull stripping, and bias field correction.
  3. human Radiologist
    Review the classification results and ensure the accuracy of sequence identification.
  4. algorithm AI system
    Segment tumor tissue subtypes using CNNs to generate quantitative tumor measurements. Optionally allow Expert-in-the-loop manual refinement of segmentation results.
  5. human Oncologist
    Utilize the quantitative tumor measurements for personalized treatment planning and response assessment. Leverage the system to streamline data curation, model prototyping, and standardized dataset creation for research collaborations.
  6. algorithm AI system
    Provide segmentation masks for longitudinal tumor tracking and quantitative growth assessment. Generate standardized reports and visualizations based on the processed MRI data.
L

Lizard: A Large-Scale Dataset for Colonic Nuclear Instance Segmentation and Classification

Manual laborBodily tracesAbundance of dataAutomationMedical Experts

  1. human Domain Experts
    Provide input for accurate delineation of object boundaries
  2. algorithm Multi-stage annotation pipeline
    Utilize input from Domain Experts to collect accurate annotations at scale
  3. human Pathologists
    Assess representative samples provided by the dataset
  4. algorithm Quantitative concordance statistics computation
    Compute quantitative concordance statistics between pathologists and the dataset to ensure the accuracy of the annotations
  5. human Researchers
    Test their developed models on the dataset
  6. algorithm Performance benchmarking
    Provide performance benchmarks to encourage the development of accurate and interpretable downstream models for the computational analysis of H&E stained colon tissue
A

A Deep Learning and Gamification Approach to Energy Conservation at Nanyang Technological University

Social behaviorLogsOptimizationGeneric User

  1. algorithm IoT sensors and cyber-physical systems sensing/actuation platforms
    The algorithms collect data from IoT sensors and cyber-physical systems sensing/actuation platforms.
  2. algorithm deep learning models
    The algorithms analyze the data to predict the energy consumption of each resource based on the behavior of Occupants.
  3. algorithm deep learning models
    The algorithms optimize the energy consumption of the building as a whole.
  4. algorithm deep learning models
    The algorithms provide feedback to Occupants about their energy usage.
  5. algorithm deep learning models
    The algorithms suggest ways to improve the energy-saving behavior of Occupants.
  6. human Occupants
    The Occupants modify their behavior based on the feedback and suggestions provided by the algorithms.
  7. human Occupants
    The Occupants earn points based on their energy-saving behavior.
  8. human Occupants
    The points earned by Occupants can be used to compete with other Occupants in the building.
T

Towards Model-informed Precision Dosing with Expert-in-the-loop Machine Learning

High costsLack of dataOpacityAnnotationsAutomationLow costsDomain Experts

  1. human Expert
    Experts can directly edit the learned rules generated by the model to improve its interpretability and accuracy.
  2. human Expert
    Experts can annotate observations/predicted values to help identify errors or areas where the model needs improvement.
  3. algorithm machine learning model
    The model is updated based on the feedback provided by the Experts, and the learning process continues iteratively.
I

Iterative Human and Automated Identification of Wildlife Images

Data analysisEnvironmental dataAnnotationsHuman supportAnnotators

  1. algorithm deep learning model
    Predict categories of images using deep learning models trained from previous periods
  2. algorithm deep learning model
    Calculate confidence levels for predicted image categories
  3. human Annotator
    Verify low-confidence predictions through human annotation
  4. human Annotator
    Provide annotations for low-confidence predictions
  5. algorithm deep learning model
    Include pseudo-labels in final data set for further model updates or ecological analyses
A

AstronomicAL: An interactive dashboard for visualisation, integration and classification of data using Active Learning

InaccuraciesIntegrationsImagesAstronomical dataHuman supportAccuracyOptimizationEnd-User

  1. human Researcher
    Import dataset into AstronomicAL and visualize and integrate data from different sources using customizable domain-specific plots.
  2. algorithm AstronomicAL
    Utilize active learning techniques to prioritize data that offer high information gain and explore each data point chosen, injecting Domain Expertise directly into the training process to ensure accurate and reliable labels.
  3. human Researcher
    Adapt AstronomicAL for research to allow for domain-specific plots, novel query strategies, and improved models. Customize models and query strategies to improve performance.
  4. algorithm AstronomicAL
    Curate a labeled test set to demonstrate the validity and Generalizability of the model. Mark any example as unsure, ensuring that all training data are of high quality.
  5. algorithm AstronomicAL
    Run AstronomicAL entirely locally on the User’s system, providing a private space to experiment. Export a simple configuration file to share entire layouts, models, and assigned labels with the community, allowing for complete transparency and effortless reproduction of results.
P

Proactive Decision Support using Automated Planning

Decision makingLogsHuman collaborationDecision makers

  1. human Decision maker
    Human identifies the tasks or goals to be accomplished in a complex planning environment.
  2. algorithm automated planner
    Automated planning system (RADAR) receives the model of the world, initial state, and goals/tasks from the human Decision maker. RADAR's automated planner analyzes the provided information and generates a plan or course-of-action (COA) based on the given model and goals.
  3. algorithm automated planner
    The automated planner continuously monitors the planning process of the human Decision maker and the current state of the environment.
  4. algorithm automated planner
    Based on the current state and resource availability, RADAR's automated planner provides alerts and suggestions to the human Decision maker regarding potential drawbacks in the plan, resource constraints, or problems that may arise in the future.
  5. human Decision maker
    Human Decision maker evaluates the alerts and suggestions provided by RADAR and incorporates them into their decision-making process as they see fit.
S

Scones: Towards Conversational Authoring of Sketches

InefficienciesWritten textAutomationEnd-User

  1. human User
    Provides natural language text instructions describing the desired scene or modifications to an existing sketch.
  2. algorithm deep learning model
    Utilizes deep learning models to interpret the natural language input and understand the User's instructions.
  3. algorithm sketch generation model
    Generates sketched scenes based on the interpreted text instructions, using state-of-the-art deep neural network architectures.
  4. human User
    Reviews the generated sketches and provides feedback or additional instructions for refinement.
  5. algorithm sketch generation model
    Incorporates User feedback and iteratively refines the sketches based on the provided instructions.
T

THEaiTRE: Artificial Intelligence to Write a Theatre Play

Not mentionedNot mentionedAutomationContent creationEnd-Users

  1. human Play Expert
    Generate a synopsis of the play using various options, such as play background/setting from play databases, more detailed synopses from fan websites, or scenic remarks extracted from texts of plays themselves.
  2. human Play Expert
    Generate a list of characters based on the synopsis.
  3. algorithm Language Model
    Seed a language model (LM) with a prompt that is the beginning of a dramatic situation.
  4. algorithm Language Model
    Fine-tune the LM to theatre plays to see how far this approach can go.
  5. algorithm Language model
    Restrict the generation by enforcing that only certain predetermined characters speak, possibly in a pregenerated order. This can be achieved by stopping the generation.
  6. algorithm Language model
    Generate the final text using a similar approach as the base LM generation.
I

Iterative annotation to ease neural network training: Specialized machine learning in medical image analysis

AdoptionImagesAutomationMedical Experts

  1. algorithm semantic segmentation network
    The algorithm processes whole slide images using a semantic segmentation network.
  2. human Expert
    Human Experts interact with the network's predictions and correct any inaccuracies in the annotations.
  3. algorithm semantic segmentation network
    The corrected annotations provided by human Experts are used to retrain the semantic segmentation network.
  4. algorithm semantic segmentation network
    The network's predictions are converted back to a format for display in WSI viewing software, such as Aperio ImageScope.
  5. human Expert
    Human Experts review the updated network predictions and provide further feedback, initiating additional training iterations if necessary.
  6. algorithm semantic segmentation network
    The network is fine-tuned based on human feedback, leading to improved accuracy and reduced burden of manual WSI annotation.
P

Plan Explanations as Model Reconciliation: Moving Beyond Explanation as Soliloquy

OpaquenessExplanationsHuman collaborationAcceptanceEnd-User

  1. algorithm AI system
    The AI system identifies differences between its own model and the human's model.
  2. algorithm AI system
    The AI system utilizes algorithms to automatically compute explanations that suggest changes to the human's model.
  3. algorithm AI system
    The AI system presents these explanations to the human, highlighting the suggested changes to the human's model.
  4. human User
    The human reviews the explanations and considers the suggested changes to their model.
  5. human User
    Based on the explanations and suggested changes, the human updates their model to align it with the AI system's model.
  6. algorithm AI system
    With the reconciled model, the AI system optimizes its plans and behavior to be in line with the updated human model.
N

Natural Language Sentence Generation from API Specifications

Manual laborWritten textAutomationEnd-UserImplicit Annotators

  1. algorithm NLG Pipeline
    Algorithm extracts phrases from OpenAPI specification and possibly augments them with human-provided phrases.
  2. algorithm NLG Pipeline
    Algorithm generates equivalent sentences using a variety of language models.
  3. human API Developer
    API developer interacts with the system through a web UI to provide feedback on the generated sentences.
  4. algorithm NLG Pipeline
    Algorithm filters the generated sentences to eliminate noisy ones and selects a diverse set of sentences based on human feedback.
  5. human API Developer
    API developer selects appropriate sentences for training the intent classifier.
  6. algorithm Intent Recognition Model
    Algorithm trains the intent recognition model for the chatbot.
  7. algorithm Intent Recognition Model
    Trained intent classifier is deployed and serves to classify Users’ intents based on their input utterances in the dialog system while attempting to invoke an API endpoint conversationally.
I

Interactive Evolutionary Multi-Objective Optimization via Learning-to-Rank

ComplexityDecision makingFeedbackSpeedHuman supportDecision makers

  1. algorithm EMO algorithm
    The EMO algorithm generates a set of candidate solutions.
  2. algorithm consultation module
    The consultation module presents a set of selected candidate solutions to the Decision Maker.
  3. human Decision Maker
    The Decision Maker provides feedback on the candidate solutions based on her preference.
  4. algorithm LTR neural network
    The LTR neural network learns the Decision Maker's preference based on the feedback.
  5. algorithm LTR neural network
    The LTR model is applied to guide the EMO algorithm towards the SOI.
  6. algorithm EMO algorithm
    The EMO algorithm generates a new set of candidate solutions based on the guidance from the LTR model.
  7. algorithm consultation module
    The consultation module presents the new set of candidate solutions to the Decision Maker.
  8. human Decision Maker
    The Decision Maker provides feedback on the new candidate solutions based on her preference.
Z

Zero Shot Learning for Code Education: Rubric Sampling with Deep Learning Inference

CodeHuman supportSpeedHuman collaborationTeachers

  1. human Experts
    In this step, we ask Experts to describe a student’s thought process, enumerating strategies to get to a right or wrong answer. Given a detailed enough description, we can use it to label indefinitely. These labels will be noisy but the quantity should make up for any uncertainty.
  2. algorithm Deep learning
    Once we have elicited samples from the Expert prior, we can use deep learning techniques to infer the correct label for a given input x. This involves training a deep model on the labeled data and using it to predict labels for new inputs.
  3. algorithm Rubric sampling
    Once we have inferred the correct label for a given input x, we can provide feedback to the student based on their performance. This feedback can be associated with specific parts of a student's solution and can articulate their misconceptions in the language of the instructor.
A

Acquire, Augment, Segment & Enjoy: Weakly Supervised Instance Segmentation of Supermarket Products

Manual laborInefficienciesImagesAutomationImplicit Annotators

  1. algorithm Proposed system
    Automatically segment the corresponding object from the background using basic image processing techniques.
  2. human Annotator
    Manually input the class of the object on the turntable.
  3. algorithm Proposed system
    Create annotations for instance-aware semantic segmentation of reasonable quality with minimal effort.
A

AgentBuddy: A Contextual Bandit based Decision Support System for Customer Support Agents

ProductivityFeedback dataAutomationCustomer support

  1. algorithm Bandit algorithm
    Choose one of the models among 'Search', 'Hand curated answers' etc.
  2. algorithm Chosen model
    Provide the answer which is observed by the CSA (care Agent).
  3. human CSA
    Receive feedback on a scale of 1-5 from the CSA.
  4. algorithm Bandit algorithm
    Update its weights based on this feedback.
A

An Image Processing Pipeline for Camera Trap Time-Lapse Recordings

Data analysisEnvironmental dataAnnotationsHuman supportNot mentioned

  1. algorithm machine learning
    The machine learning algorithm generates a draft segmentation of the video.
  2. algorithm machine learning
    The algorithm identifies at least one image containing tortoises per ground truth segment.
  3. algorithm machine learning
    The algorithm draws the human's attention to that region of the recording.
  4. human Annotator
    The human quickly creates a highly accurate segmentation with the assistance of the algorithm.
A

An Interactive Explanatory AI System for Industrial Quality Control

InaccuraciesQuality controlLogsHuman supportDomain Experts

  1. algorithm ILP system
    Based on difficult-to-formalize experiential prior knowledge from human Domain Experts
  2. algorithm CNN
    The system provides transparent explanations for the results to the human Domain Experts.
  3. human Domain Expert
    The human Domain Experts use the explanations to make decisions.
  4. algorithm support system
    From the human Domain Experts to improve its performance
L

Learning to Segment Anatomical Structures Accurately from One Exemplar

Lack of dataQuality of dataImagesAnnotationsAutomationMedical Experts

  1. algorithm CTN
    CTN takes the exemplar contour as an initialization and gradually evolves it to minimize the weighted loss for each unlabeled image.
  2. human User
    If any part in the predicted contour is inaccurate, Users can correct them by drawing line segments.
  3. algorithm CTN
    CTN formats these corrections as partial contours and incorporates them back into the training via an additional Chamfer loss.
  4. algorithm CTN
    CTN consistently improves with more human corrections, potentially achieving better performance than fully supervised methods with considerably less annotation efforts.
I

IPLAN: Interactive and Procedural Layout Planning

Manual laborImagesCreativityFreedomAutomationDesigners

  1. algorithm
    iPLAN reverse-engineers the final design to obtain the stage-to-stage process, based on principles that are widely adopted by professional designers . This process is performed by the algorithm.
  2. algorithm
    iPLAN uses a Markov chain model to capture the full design procedure. The algorithm designs and uses the Markov chain model to capture the full design procedure.
  3. human Designer
    iPLAN accepts User guidance at every stage of the design process. This means that the User can provide input at different stages across a wide range of levels of detail.
  4. algorithm
    iPLAN automatically suggests possible designs based on the User's input. The algorithm is designed to offer the capability of fully automated generation .
  5. algorithm
    Learning from human inputs on designs augmented by reverse-engineered processes.
R

Race Driver Evaluation at a Driving Simulator using a physical Model and a Machine Learning Approach

Human interpretationInaccuraciesUser dataAutomationDrivers

  1. human Driver
    Provide data to the algorithm through a driving simulation.
  2. algorithm optCog optimizer
    Optimizing the vehicle's acceleration in the center of gravity (Cog) using the optCog optimizer.
  3. algorithm optTire optimizer
    Evaluating the maximum possible force of each tire independently using the optTire optimizer.
  4. algorithm N/A
    Defining scores for overall, vehicle-trajectory, and handling performance based on the outcome of the optimizers.
  5. algorithm machine learning model architecture
    Proposing a machine learning model architecture to directly determine the scores from the initial data.
A

Augmenting Scientific Creativity with an Analogical Search Engine

Information retrievalWritten textScientific dataHuman supportAnnotators

  1. algorithm machine learning model
    The fully automated system employs advanced machine learning algorithms to process a large corpus of scientific papers.
  2. algorithm machine learning model
    Algorithms are utilized to emulate human judgment of purpose match, ensuring that the system finds partial purpose matches in the top results.
  3. algorithm machine learning model
    The system presents the top results to the Users through an intuitive and User-friendly search interface.
  4. human User
    Users interact with the system to explore and adapt distant inspirations found in the scientific papers, leveraging the analogical similarities identified by the algorithms.
  5. algorithm machine learning model
    As Users interact with the system and explore the retrieved scientific papers, their interactions and feedback can be used to further refine the algorithms and improve the relevance and accuracy of the results over time.
L

Learning from Thresholds: Fully Automated Classification of Tumor Infiltrating Lymphocytes for Multiple Cancer Types

High costsTime consumingBodily tracesAutomationNot mentioned

  1. human Annotator
    Manually annotate image patches with TIL positive or TIL negative labels for a subset of cancer types.
  2. algorithm TIL classification network
    Apply a classification algorithm to Whole Slide Images (WSIs) and adjust the predicted TIL probability maps by applying thresholds to generate semi-automatic annotations.
  3. algorithm AI model
    Combine manually annotated patches and semi-automatically annotated patches to form the training set for the AI model.
  4. algorithm AI model
    Train deep neural network models, such as VGG 16-layer network and Inception-V4, using the combined dataset to Generalize across multiple cancer types.
  5. algorithm AI model
    Automatically apply the trained AI model to each cancer type without human adjustment, providing TIL prediction results.
L

Learning to Learn in Simulation

MimickryLogsSpeedAutomationNot mentioned

  1. algorithm Curiosity agent
    The curiosity agent selects actions to navigate the robot within the exploration space.
  2. algorithm Curiosity agent
    The curiosity agent may request ground truth annotations from the human operator when additional information is needed.
  3. human Operator
    The human operator provides feedback by providing ground truth annotations based on the robot's requests.
  4. algorithm Robot
    The robot uses the feedback to improve its performance and adjust its exploration strategy.
  5. algorithm Curiosity agent
    The curiosity agent continues to select actions and seek out new information, iteratively improving the robot's learning process.
  6. algorithm Robot
    The collaboration aims to reduce the burden on the human operator and enhance the robot's autonomous learning capabilities.
S

Shared Autonomy via Deep Reinforcement Learning

PrivacyFeedbackTasksSpeedHuman supportHuman interpretationEnd-User

  1. algorithm deep reinforcement learning algorithm
    The algorithm uses deep reinforcement learning with neural network function approximation to learn an end-to-end mapping from observation and input to agent action values, with task reward as the only form of supervision.
  2. algorithm deep reinforcement learning algorithm
    The algorithm provides real-time action feedback to the User based on the learned mapping.
  3. human User
    The User provides feedback to the algorithm by providing a terminal reward upon succeeding or failing at the task.
  4. algorithm deep reinforcement learning algorithm
    The algorithm decomposes the agent's reward function into known terms computed for every state and a terminal reward provided by the User, enabling the system to learn efficiently from a dense reward signal that captures Generally useful behaviors and adapt to individual Users through feedback.
  5. algorithm deep reinforcement learning algorithm
    The algorithm is capable of incorporating inferred goals into the agent's observations when the goal space and User model are known, further improving sample efficiency.
  6. algorithm deep reinforcement learning algorithm
    The algorithm balances the need to follow User commands closely while also deviating from the User's actions when they are suboptimal, discarding actions whose values fall below some threshold and selecting the remaining action closest to the User's input.
  7. algorithm deep reinforcement learning algorithm
    The algorithm learns to assist the User without access to private information, implicitly inferring it from the User's input.
U

Understanding Aesthetic Evaluation using Deep Learning

Manual laborLack of dataUser dataAutomationHuman supportEnd-User

  1. algorithm CNN classifier
    Learn visual features important for aesthetic evaluation
  2. algorithm CNN classifier
    Increase accuracy in automating personal aesthetic judgement
  3. algorithm generative system
    Generate a set of phenotype images
  4. algorithm CNN classifier
    Use CNN classifier to find visual similarity between generated phenotype images and database of examples
  5. algorithm CNN classifier
    Assign a score to each generated phenotype image based on its similarity to the examples in the database
  6. algorithm CNN classifier
    Rank the generated phenotype images in order of aesthetic quality based on the assigned scores

Author, PhD Candidate - Andrea Benedetti

Scientific supervisor - Donato Ricci

Researched and developed in 2023.

The AI Compendium is developed within the scope of Shaping AI, s a multi-national and multi-disciplinary social research project that examines the global trajectories of public discourse on artificial intelligence (AI) in four countries (Germany, UK, Canada and France) over a ten-year period 2012-2021. The AI Compendium is part of the PhD Thesis of Andrea Benedetti.