Since Scope Topics of interest include:
\nFine-grained categorization with humans in the loop
Embedding human experts’ knowledge into computational models
Machine teaching
Interpretable fine-grained models
Just insights about some of the topics above.
\nAuthors: Jörg Wagner, Jan Mathias Köhler, Tobias Gindele, Leon Hetzel, Jakob Thaddäus Wiedemer, Sven Behnke - arXiv:1908.02686v1
\n\"To verify and validate networks, it is essential to gain insight into their decisions, limitations as well as possible shortcomings of training data. In this work, the authors proposed a post-hoc, optimization based visual explanation method, which highlights the evidence in the input image for a specific prediction.\"
\n\"Their approach is based on a novel technique to defend against adversarial evidence (i.e. faulty evidence due to artefacts) by filtering gradients during optimization. The defense does not depend on human-tuned parameters. It enables explanations which are both fine-grained and preserve the characteristics of images, such as edges and colors.\"
\n\"The explanations are interpretable, suited for visualizing detailed evidence and can be tested as they are valid model inputs. The authors qualitatively and quantitatively evaluated their approach on a multitude of models and datasets.\"
\nhttps://arxiv.org/abs/1908.02686
\nAuthors: Zixuan Huang, Yin Li
\n\"The authors presented an interpretable deep model for fine-grained visual recognition. At the core of our method lies the integration of region-based part discovery and attribution within a deep neural network.\"
\n\"Their model is trained using image-level object labels, and provides an interpretation of its results via the segmentation of object parts and the identification of their contributions towards classification.\"
\n\"To facilitate the learning of object parts without direct supervision, they explored a simple prior of the occurrence of object parts. The authors demonstrated that this prior, when combined with their region-based part discovery and attribution, leads to an interpretable model that remains highly accurate. Their model is evaluated on major fine-grained recognition datasets, including CUB-200, CelebA and iNaturalist.
\n\"Their results compared favourably to state-of-the-art methods on classification tasks, and outperformed previous approaches on the localization of object parts.\"
\nPART DISCOVERY for FINE-GRAINED RECOGNITION.
\n\"Identifying discriminative object parts is important for fine-grained classification. For example, bounding box or landmark annotations can be used to learn object parts for fine-grained classification.\"
\n\"To avoid costly annotation of object parts, several recent works focused on unsupervised or weakly-supervised part learning using deep models. Spectral clustering on convolutional filters was performed to find representative filters for parts. It was proposed to learn a bank of
\nconvolutional filters that capture class-specific object parts.\"
\"Moreover, attention models have also been explored extensively for learning parts. It was made use of reinforcement learning to select region proposals for fine-grained classification.\"
\nTo read about the results: Results on CUB-200-2011 and Results on iNaturalist2017 and their Interpretability.
\nhttps://openaccess.thecvf.com/content_CVPR_2020/papers/Huang_Interpretable_and_Accurate_Fine-grained_Recognition_via_Region_Grouping_CVPR_2020_paper.pdf \u0002
Author: Xiaojin Zhu
\n\"What is machine teaching? Consider a “student” who is a machine learning algorithm, for example, a Support Vector Machine (SVM) or kmeans clustering. Now consider a “teacher” who wants the student to learn a target model θ. For example, θ can be a specific hyperplane in SVM, or the location of the k centroids in kmeans. The teacher knows θ and the student’s
\nlearning algorithm, and teaches by giving the student training examples. Machine teaching aims to design the optimal training set D.\"
\" What do you mean by optimal? One definition is the cardinality of D: the smaller |D| is,
\nthe better. But there are other definitions as we shall see.\"
\" If we already know the true model θ , why bother training a learner? The applications are such that the teacher and the learner are separate entities, and the teacher cannot directly “hard
\nwire” the learner. One application is education where the learner is a human student. A more sinister “application” is security where the learner is an adaptive spam filter, and the “teacher” is a hacker who wishes to change the filtering behavior by sending the spam filter specially designed messages. Regardless of the intention, machine teaching aims to maximally influence the learner via optimal training data.\"
Applying Machine Teaching:
\n\"Optimization: Despite our early successes in certain teaching settings, solving for the optimal training data D is still difficult in general. The author expected that many tools developed in the optimization community can be brought to bear on difficult problems.\"
\n\"Theory: Machine teaching originated from the theoretical study of teaching dimension. It is important to understand the theoretical properties of the optimal training set under more general teaching settings. The author speculated that information theory may be a suitable tool here: the teacher is the encoder, the learner is the decoder, and the message is the target model. But there is a twist: the decoder is not ideal. It is specified by whatever machine learning algorithm it runs.\"
\n\"Education: Arguably more complex, education first needs to identify computable cognitive models of the student. Existing intelligent tutoring systems are a good place to start: with a little effort, one may hypothesize the inner works of the student black-box.\"
\n\"Novel application: Consider computer security. As mentioned earlier, machine teaching also describes the optimal attack strategy if a hacker wants to influence a learning agent, see and the references therein. The question is, knowing the optimal attack strategy predicted by machine teaching, can we effectively defend the learning agent? There may be other serendipitous applications of machine teaching besides education and computer security. \"
\nhttps://pages.cs.wisc.edu/~jerryzhu/machineteaching/pub/MachineTeachingAAAI15.pdf
\nBy Jennifer Langston - April 23, 2019
\n\"Machine teaching seeks to gain knowledge from people rather than extracting knowledge from data alone. A person who understands the task at hand — whether how to decide which department in a company should receive an incoming email or how to automatically position wind turbines to generate more energy — would first decompose that problem into smaller parts. Then they would provide a limited number of examples, or the equivalent of lesson plans, to help the machine learning algorithms solve it.\"
\n\"In supervised learning scenarios, machine teaching is particularly useful when little or no labeled training data exists for the machine learning algorithms because an industry or company’s needs are so specific.\"
\n“Even the smartest AI will struggle by itself to learn how to do some of the deeply complex tasks that are common in the real world. So you need people guiding AI systems to learn the things that we already know,” . “Taking this turnkey AI and having non-experts use it to do much more complex tasks is really the sweet spot for machine teaching.”
\n\"Deep reinforcement learning, a branch of AI in which algorithms learn by trial and error based on a system of rewards, has successfully outperformed people in video games. But those models have struggled to master more complicated real-world industrial tasks.\"
\n\"Adding a machine teaching layer — or infusing an organization’s unique subject matter expertise directly into a deep reinforcement learning model — can dramatically reduce the time it takes to find solutions to these deeply complex real-world problems.\"
\n\"Telling to AI brain what’s important to focus on at the outset can short circuit a lot of fruitless and time-consuming exploration as it tries to learn in simulation what does and doesn’t work\"
\n“The reason machine teaching proves critical is because if you just use reinforcement learning naively and don’t give it any information on how to solve the problem, it’s going to explore randomly and will maybe hopefully — but frequently not ever — hit on a solution that works.”
\n“It makes problems truly solvable whereas without machine teaching they aren’t.”
\nhttps://blogs.microsoft.com/ai/machine-teaching/
", "messages": [ { "id": 1695992, "postDate": "2022-02-18T14:26:57.973Z", "content": "Since Scope Topics of interest include:
\nFine-grained categorization with humans in the loop
Embedding human experts’ knowledge into computational models
Machine teaching
Interpretable fine-grained models
Just insights about some of the topics above.
\nAuthors: Jörg Wagner, Jan Mathias Köhler, Tobias Gindele, Leon Hetzel, Jakob Thaddäus Wiedemer, Sven Behnke - arXiv:1908.02686v1
\n\"To verify and validate networks, it is essential to gain insight into their decisions, limitations as well as possible shortcomings of training data. In this work, the authors proposed a post-hoc, optimization based visual explanation method, which highlights the evidence in the input image for a specific prediction.\"
\n\"Their approach is based on a novel technique to defend against adversarial evidence (i.e. faulty evidence due to artefacts) by filtering gradients during optimization. The defense does not depend on human-tuned parameters. It enables explanations which are both fine-grained and preserve the characteristics of images, such as edges and colors.\"
\n\"The explanations are interpretable, suited for visualizing detailed evidence and can be tested as they are valid model inputs. The authors qualitatively and quantitatively evaluated their approach on a multitude of models and datasets.\"
\nhttps://arxiv.org/abs/1908.02686
\nAuthors: Zixuan Huang, Yin Li
\n\"The authors presented an interpretable deep model for fine-grained visual recognition. At the core of our method lies the integration of region-based part discovery and attribution within a deep neural network.\"
\n\"Their model is trained using image-level object labels, and provides an interpretation of its results via the segmentation of object parts and the identification of their contributions towards classification.\"
\n\"To facilitate the learning of object parts without direct supervision, they explored a simple prior of the occurrence of object parts. The authors demonstrated that this prior, when combined with their region-based part discovery and attribution, leads to an interpretable model that remains highly accurate. Their model is evaluated on major fine-grained recognition datasets, including CUB-200, CelebA and iNaturalist.
\n\"Their results compared favourably to state-of-the-art methods on classification tasks, and outperformed previous approaches on the localization of object parts.\"
\nPART DISCOVERY for FINE-GRAINED RECOGNITION.
\n\"Identifying discriminative object parts is important for fine-grained classification. For example, bounding box or landmark annotations can be used to learn object parts for fine-grained classification.\"
\n\"To avoid costly annotation of object parts, several recent works focused on unsupervised or weakly-supervised part learning using deep models. Spectral clustering on convolutional filters was performed to find representative filters for parts. It was proposed to learn a bank of
\nconvolutional filters that capture class-specific object parts.\"
\"Moreover, attention models have also been explored extensively for learning parts. It was made use of reinforcement learning to select region proposals for fine-grained classification.\"
\nTo read about the results: Results on CUB-200-2011 and Results on iNaturalist2017 and their Interpretability.
\nhttps://openaccess.thecvf.com/content_CVPR_2020/papers/Huang_Interpretable_and_Accurate_Fine-grained_Recognition_via_Region_Grouping_CVPR_2020_paper.pdf \u0002
Author: Xiaojin Zhu
\n\"What is machine teaching? Consider a “student” who is a machine learning algorithm, for example, a Support Vector Machine (SVM) or kmeans clustering. Now consider a “teacher” who wants the student to learn a target model θ. For example, θ can be a specific hyperplane in SVM, or the location of the k centroids in kmeans. The teacher knows θ and the student’s
\nlearning algorithm, and teaches by giving the student training examples. Machine teaching aims to design the optimal training set D.\"
\" What do you mean by optimal? One definition is the cardinality of D: the smaller |D| is,
\nthe better. But there are other definitions as we shall see.\"
\" If we already know the true model θ , why bother training a learner? The applications are such that the teacher and the learner are separate entities, and the teacher cannot directly “hard
\nwire” the learner. One application is education where the learner is a human student. A more sinister “application” is security where the learner is an adaptive spam filter, and the “teacher” is a hacker who wishes to change the filtering behavior by sending the spam filter specially designed messages. Regardless of the intention, machine teaching aims to maximally influence the learner via optimal training data.\"
Applying Machine Teaching:
\n\"Optimization: Despite our early successes in certain teaching settings, solving for the optimal training data D is still difficult in general. The author expected that many tools developed in the optimization community can be brought to bear on difficult problems.\"
\n\"Theory: Machine teaching originated from the theoretical study of teaching dimension. It is important to understand the theoretical properties of the optimal training set under more general teaching settings. The author speculated that information theory may be a suitable tool here: the teacher is the encoder, the learner is the decoder, and the message is the target model. But there is a twist: the decoder is not ideal. It is specified by whatever machine learning algorithm it runs.\"
\n\"Education: Arguably more complex, education first needs to identify computable cognitive models of the student. Existing intelligent tutoring systems are a good place to start: with a little effort, one may hypothesize the inner works of the student black-box.\"
\n\"Novel application: Consider computer security. As mentioned earlier, machine teaching also describes the optimal attack strategy if a hacker wants to influence a learning agent, see and the references therein. The question is, knowing the optimal attack strategy predicted by machine teaching, can we effectively defend the learning agent? There may be other serendipitous applications of machine teaching besides education and computer security. \"
\nhttps://pages.cs.wisc.edu/~jerryzhu/machineteaching/pub/MachineTeachingAAAI15.pdf
\nBy Jennifer Langston - April 23, 2019
\n\"Machine teaching seeks to gain knowledge from people rather than extracting knowledge from data alone. A person who understands the task at hand — whether how to decide which department in a company should receive an incoming email or how to automatically position wind turbines to generate more energy — would first decompose that problem into smaller parts. Then they would provide a limited number of examples, or the equivalent of lesson plans, to help the machine learning algorithms solve it.\"
\n\"In supervised learning scenarios, machine teaching is particularly useful when little or no labeled training data exists for the machine learning algorithms because an industry or company’s needs are so specific.\"
\n“Even the smartest AI will struggle by itself to learn how to do some of the deeply complex tasks that are common in the real world. So you need people guiding AI systems to learn the things that we already know,” . “Taking this turnkey AI and having non-experts use it to do much more complex tasks is really the sweet spot for machine teaching.”
\n\"Deep reinforcement learning, a branch of AI in which algorithms learn by trial and error based on a system of rewards, has successfully outperformed people in video games. But those models have struggled to master more complicated real-world industrial tasks.\"
\n\"Adding a machine teaching layer — or infusing an organization’s unique subject matter expertise directly into a deep reinforcement learning model — can dramatically reduce the time it takes to find solutions to these deeply complex real-world problems.\"
\n\"Telling to AI brain what’s important to focus on at the outset can short circuit a lot of fruitless and time-consuming exploration as it tries to learn in simulation what does and doesn’t work\"
\n“The reason machine teaching proves critical is because if you just use reinforcement learning naively and don’t give it any information on how to solve the problem, it’s going to explore randomly and will maybe hopefully — but frequently not ever — hit on a solution that works.”
\n“It makes problems truly solvable whereas without machine teaching they aren’t.”
\nhttps://blogs.microsoft.com/ai/machine-teaching/
", "rawMarkdown": "Since Scope Topics of interest include:\n\n- Fine-grained categorization with humans in the loop\n\n- Embedding human experts’ knowledge into computational models\n\n- Machine teaching\n\n- Interpretable fine-grained models\n\nJust insights about some of the topics above.\n\n\n#Interpretable and Fine-Grained Visual Explanations for Convolutional Neural Networks\n\nAuthors: Jörg Wagner, Jan Mathias Köhler, Tobias Gindele, Leon Hetzel, Jakob Thaddäus Wiedemer, Sven Behnke - arXiv:1908.02686v1\n\n\"To verify and validate networks, it is essential to gain insight into their decisions, limitations as well as possible shortcomings of training data. In this work, the authors proposed a post-hoc, optimization based visual explanation method, which highlights the evidence in the input image for a specific prediction.\"\n\n\"Their approach is based on a novel technique to defend against adversarial evidence (i.e. faulty evidence due to artefacts) by filtering gradients during optimization. The defense does not depend on human-tuned parameters. It enables explanations which are both fine-grained and preserve the characteristics of images, such as edges and colors.\"\n\n\"The explanations are interpretable, suited for visualizing detailed evidence and can be tested as they are valid model inputs. The authors qualitatively and quantitatively evaluated their approach on a multitude of models and datasets.\"\n\nhttps://arxiv.org/abs/1908.02686\n\n#Interpretable and Accurate Fine-grained Recognition via Region Grouping\n\nAuthors: Zixuan Huang, Yin Li\n\n\"The authors presented an interpretable deep model for fine-grained visual recognition. At the core of our method lies the integration of region-based part discovery and attribution within a deep neural network.\"\n\n\"Their model is trained using image-level object labels, and provides an interpretation of its results via the segmentation of object parts and the identification of their contributions towards classification.\"\n\n\"To facilitate the learning of object parts without direct supervision, they explored a simple prior of the occurrence of object parts. The authors demonstrated that this prior, when combined with their region-based part discovery and attribution, leads to an interpretable model that remains highly accurate. Their model is evaluated on major fine-grained recognition datasets, including CUB-200, CelebA and iNaturalist. \n\n\"Their results compared favourably to state-of-the-art methods on classification tasks, and outperformed previous approaches on the localization of object parts.\"\n\nPART DISCOVERY for FINE-GRAINED RECOGNITION.\n\n\"Identifying discriminative object parts is important for fine-grained classification. For example, bounding box or landmark annotations can be used to learn object parts for fine-grained classification.\"\n\n\"To avoid costly annotation of object parts, several recent works focused on unsupervised or weakly-supervised part learning using deep models. Spectral clustering on convolutional filters was performed to find representative filters for parts. It was proposed to learn a bank of\nconvolutional filters that capture class-specific object parts.\"\n\n\"Moreover, attention models have also been explored extensively for learning parts. It was made use of reinforcement learning to select region proposals for fine-grained classification.\"\n\nTo read about the results: Results on CUB-200-2011 and Results on iNaturalist2017 and their Interpretability.\nhttps://openaccess.thecvf.com/content_CVPR_2020/papers/Huang_Interpretable_and_Accurate_Fine-grained_Recognition_via_Region_Grouping_CVPR_2020_paper.pdf \u0002\n\n#Machine Teaching: An Inverse Problem to Machine Learning and an Approach Toward Optimal Education\n\nAuthor: Xiaojin Zhu\n\n \"What is machine teaching? Consider a “student” who is a machine learning algorithm, for example, a Support Vector Machine (SVM) or kmeans clustering. Now consider a “teacher” who wants the student to learn a target model θ. For example, θ can be a specific hyperplane in SVM, or the location of the k centroids in kmeans. The teacher knows θ and the student’s\nlearning algorithm, and teaches by giving the student training examples. Machine teaching aims to design the optimal training set D.\"\n\n\" What do you mean by optimal? One definition is the cardinality of D: the smaller |D| is,\nthe better. But there are other definitions as we shall see.\"\n\n\" If we already know the true model θ , why bother training a learner? The applications are such that the teacher and the learner are separate entities, and the teacher cannot directly “hard\nwire” the learner. One application is education where the learner is a human student. A more sinister “application” is security where the learner is an adaptive spam filter, and the “teacher” is a hacker who wishes to change the filtering behavior by sending the spam filter specially designed messages. Regardless of the intention, machine teaching aims to maximally influence the learner via optimal training data.\"\n\nApplying Machine Teaching:\n\n\"Optimization: Despite our early successes in certain teaching settings, solving for the optimal training data D is still difficult in general. The author expected that many tools developed in the optimization community can be brought to bear on difficult problems.\"\n\n\"Theory: Machine teaching originated from the theoretical study of teaching dimension. It is important to understand the theoretical properties of the optimal training set under more general teaching settings. The author speculated that information theory may be a suitable tool here: the teacher is the encoder, the learner is the decoder, and the message is the target model. But there is a twist: the decoder is not ideal. It is specified by whatever machine learning algorithm it runs.\"\n\n\"Education: Arguably more complex, education first needs to identify computable cognitive models of the student. Existing intelligent tutoring systems are a good place to start: with a little effort, one may hypothesize the inner works of the student black-box.\"\n\n\"Novel application: Consider computer security. As mentioned earlier, machine teaching also describes the optimal attack strategy if a hacker wants to influence a learning agent, see and the references therein. The question is, knowing the optimal attack strategy predicted by machine teaching, can we effectively defend the learning agent? There may be other serendipitous applications of machine teaching besides education and computer security. \"\n\nhttps://pages.cs.wisc.edu/~jerryzhu/machineteaching/pub/MachineTeachingAAAI15.pdf\n\n\n#Machine teaching: How people’s expertise makes AI even more powerful\nBy Jennifer Langston - April 23, 2019\n\n\"Machine teaching seeks to gain knowledge from people rather than extracting knowledge from data alone. A person who understands the task at hand — whether how to decide which department in a company should receive an incoming email or how to automatically position wind turbines to generate more energy — would first decompose that problem into smaller parts. Then they would provide a limited number of examples, or the equivalent of lesson plans, to help the machine learning algorithms solve it.\"\n\n\"In supervised learning scenarios, machine teaching is particularly useful when little or no labeled training data exists for the machine learning algorithms because an industry or company’s needs are so specific.\"\n\n“Even the smartest AI will struggle by itself to learn how to do some of the deeply complex tasks that are common in the real world. So you need people guiding AI systems to learn the things that we already know,” . “Taking this turnkey AI and having non-experts use it to do much more complex tasks is really the sweet spot for machine teaching.”\n\n\"Deep reinforcement learning, a branch of AI in which algorithms learn by trial and error based on a system of rewards, has successfully outperformed people in video games. But those models have struggled to master more complicated real-world industrial tasks.\"\n\n\"Adding a machine teaching layer — or infusing an organization’s unique subject matter expertise directly into a deep reinforcement learning model — can dramatically reduce the time it takes to find solutions to these deeply complex real-world problems.\"\n\n\"Telling to AI brain what’s important to focus on at the outset can short circuit a lot of fruitless and time-consuming exploration as it tries to learn in simulation what does and doesn’t work\" \n\n“The reason machine teaching proves critical is because if you just use reinforcement learning naively and don’t give it any information on how to solve the problem, it’s going to explore randomly and will maybe hopefully — but frequently not ever — hit on a solution that works.” \n\n “It makes problems truly solvable whereas without machine teaching they aren’t.”\n\nhttps://blogs.microsoft.com/ai/machine-teaching/\n\n", "votes": 16 }, { "id": 1696413, "postDate": "2022-02-18T20:09:10.103Z", "content": "Thank you for sharing, @mpwolke! You are always helpful!
", "rawMarkdown": "Thank you for sharing, @mpwolke! You are always helpful! ", "votes": 1, "replies": [ { "id": 1696447, "postDate": "2022-02-18T20:52:51.603Z", "content": "Thank you so much Vad since it's very difficult to a non-coder, an outsider Kaggler to try to bring any significant Content about Machine Learning/Machine Teaching.
\nThough I've just copied and pasted the work of those relevant professionals, at least our community can read something mentioned by the Competition hosts on their Discussion topics.
\nIndeed, I hope it helps since I've learned a little bit more, after reading those precious papers.
", "rawMarkdown": "Thank you so much Vad since it's very difficult to a non-coder, an outsider Kaggler to try to bring any significant Content about Machine Learning/Machine Teaching.\n\nThough I've just copied and pasted the work of those relevant professionals, at least our community can read something mentioned by the Competition hosts on their Discussion topics.\n\nIndeed, I hope it helps since I've learned a little bit more, after reading those precious papers." } ] } ], "comments": [ { "id": 1696413, "author_name": "Vadim Irtlach", "author_url": "", "post_date": "2022-02-18T20:09:10.103000", "content": "Thank you for sharing, @mpwolke! You are always helpful!
", "votes": 1, "replies": [ { "id": 1696447, "author_name": "Marília Prata", "author_url": "", "post_date": "2022-02-18T20:52:51.603000", "content": "Thank you so much Vad since it's very difficult to a non-coder, an outsider Kaggler to try to bring any significant Content about Machine Learning/Machine Teaching.
\nThough I've just copied and pasted the work of those relevant professionals, at least our community can read something mentioned by the Competition hosts on their Discussion topics.
\nIndeed, I hope it helps since I've learned a little bit more, after reading those precious papers.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1695992": "Since Scope Topics of interest include:\n\n- Fine-grained categorization with humans in the loop\n\n- Embedding human experts’ knowledge into computational models\n\n- Machine teaching\n\n- Interpretable fine-grained models\n\nJust insights about some of the topics above.\n\n\n#Interpretable and Fine-Grained Visual Explanations for Convolutional Neural Networks\n\nAuthors: Jörg Wagner, Jan Mathias Köhler, Tobias Gindele, Leon Hetzel, Jakob Thaddäus Wiedemer, Sven Behnke - arXiv:1908.02686v1\n\n\"To verify and validate networks, it is essential to gain insight into their decisions, limitations as well as possible shortcomings of training data. In this work, the authors proposed a post-hoc, optimization based visual explanation method, which highlights the evidence in the input image for a specific prediction.\"\n\n\"Their approach is based on a novel technique to defend against adversarial evidence (i.e. faulty evidence due to artefacts) by filtering gradients during optimization. The defense does not depend on human-tuned parameters. It enables explanations which are both fine-grained and preserve the characteristics of images, such as edges and colors.\"\n\n\"The explanations are interpretable, suited for visualizing detailed evidence and can be tested as they are valid model inputs. The authors qualitatively and quantitatively evaluated their approach on a multitude of models and datasets.\"\n\nhttps://arxiv.org/abs/1908.02686\n\n#Interpretable and Accurate Fine-grained Recognition via Region Grouping\n\nAuthors: Zixuan Huang, Yin Li\n\n\"The authors presented an interpretable deep model for fine-grained visual recognition. At the core of our method lies the integration of region-based part discovery and attribution within a deep neural network.\"\n\n\"Their model is trained using image-level object labels, and provides an interpretation of its results via the segmentation of object parts and the identification of their contributions towards classification.\"\n\n\"To facilitate the learning of object parts without direct supervision, they explored a simple prior of the occurrence of object parts. The authors demonstrated that this prior, when combined with their region-based part discovery and attribution, leads to an interpretable model that remains highly accurate. Their model is evaluated on major fine-grained recognition datasets, including CUB-200, CelebA and iNaturalist. \n\n\"Their results compared favourably to state-of-the-art methods on classification tasks, and outperformed previous approaches on the localization of object parts.\"\n\nPART DISCOVERY for FINE-GRAINED RECOGNITION.\n\n\"Identifying discriminative object parts is important for fine-grained classification. For example, bounding box or landmark annotations can be used to learn object parts for fine-grained classification.\"\n\n\"To avoid costly annotation of object parts, several recent works focused on unsupervised or weakly-supervised part learning using deep models. Spectral clustering on convolutional filters was performed to find representative filters for parts. It was proposed to learn a bank of\nconvolutional filters that capture class-specific object parts.\"\n\n\"Moreover, attention models have also been explored extensively for learning parts. It was made use of reinforcement learning to select region proposals for fine-grained classification.\"\n\nTo read about the results: Results on CUB-200-2011 and Results on iNaturalist2017 and their Interpretability.\nhttps://openaccess.thecvf.com/content_CVPR_2020/papers/Huang_Interpretable_and_Accurate_Fine-grained_Recognition_via_Region_Grouping_CVPR_2020_paper.pdf \u0002\n\n#Machine Teaching: An Inverse Problem to Machine Learning and an Approach Toward Optimal Education\n\nAuthor: Xiaojin Zhu\n\n \"What is machine teaching? Consider a “student” who is a machine learning algorithm, for example, a Support Vector Machine (SVM) or kmeans clustering. Now consider a “teacher” who wants the student to learn a target model θ. For example, θ can be a specific hyperplane in SVM, or the location of the k centroids in kmeans. The teacher knows θ and the student’s\nlearning algorithm, and teaches by giving the student training examples. Machine teaching aims to design the optimal training set D.\"\n\n\" What do you mean by optimal? One definition is the cardinality of D: the smaller |D| is,\nthe better. But there are other definitions as we shall see.\"\n\n\" If we already know the true model θ , why bother training a learner? The applications are such that the teacher and the learner are separate entities, and the teacher cannot directly “hard\nwire” the learner. One application is education where the learner is a human student. A more sinister “application” is security where the learner is an adaptive spam filter, and the “teacher” is a hacker who wishes to change the filtering behavior by sending the spam filter specially designed messages. Regardless of the intention, machine teaching aims to maximally influence the learner via optimal training data.\"\n\nApplying Machine Teaching:\n\n\"Optimization: Despite our early successes in certain teaching settings, solving for the optimal training data D is still difficult in general. The author expected that many tools developed in the optimization community can be brought to bear on difficult problems.\"\n\n\"Theory: Machine teaching originated from the theoretical study of teaching dimension. It is important to understand the theoretical properties of the optimal training set under more general teaching settings. The author speculated that information theory may be a suitable tool here: the teacher is the encoder, the learner is the decoder, and the message is the target model. But there is a twist: the decoder is not ideal. It is specified by whatever machine learning algorithm it runs.\"\n\n\"Education: Arguably more complex, education first needs to identify computable cognitive models of the student. Existing intelligent tutoring systems are a good place to start: with a little effort, one may hypothesize the inner works of the student black-box.\"\n\n\"Novel application: Consider computer security. As mentioned earlier, machine teaching also describes the optimal attack strategy if a hacker wants to influence a learning agent, see and the references therein. The question is, knowing the optimal attack strategy predicted by machine teaching, can we effectively defend the learning agent? There may be other serendipitous applications of machine teaching besides education and computer security. \"\n\nhttps://pages.cs.wisc.edu/~jerryzhu/machineteaching/pub/MachineTeachingAAAI15.pdf\n\n\n#Machine teaching: How people’s expertise makes AI even more powerful\nBy Jennifer Langston - April 23, 2019\n\n\"Machine teaching seeks to gain knowledge from people rather than extracting knowledge from data alone. A person who understands the task at hand — whether how to decide which department in a company should receive an incoming email or how to automatically position wind turbines to generate more energy — would first decompose that problem into smaller parts. Then they would provide a limited number of examples, or the equivalent of lesson plans, to help the machine learning algorithms solve it.\"\n\n\"In supervised learning scenarios, machine teaching is particularly useful when little or no labeled training data exists for the machine learning algorithms because an industry or company’s needs are so specific.\"\n\n“Even the smartest AI will struggle by itself to learn how to do some of the deeply complex tasks that are common in the real world. So you need people guiding AI systems to learn the things that we already know,” . “Taking this turnkey AI and having non-experts use it to do much more complex tasks is really the sweet spot for machine teaching.”\n\n\"Deep reinforcement learning, a branch of AI in which algorithms learn by trial and error based on a system of rewards, has successfully outperformed people in video games. But those models have struggled to master more complicated real-world industrial tasks.\"\n\n\"Adding a machine teaching layer — or infusing an organization’s unique subject matter expertise directly into a deep reinforcement learning model — can dramatically reduce the time it takes to find solutions to these deeply complex real-world problems.\"\n\n\"Telling to AI brain what’s important to focus on at the outset can short circuit a lot of fruitless and time-consuming exploration as it tries to learn in simulation what does and doesn’t work\" \n\n“The reason machine teaching proves critical is because if you just use reinforcement learning naively and don’t give it any information on how to solve the problem, it’s going to explore randomly and will maybe hopefully — but frequently not ever — hit on a solution that works.” \n\n “It makes problems truly solvable whereas without machine teaching they aren’t.”\n\nhttps://blogs.microsoft.com/ai/machine-teaching/\n\n", "1696413": "Thank you for sharing, @mpwolke! You are always helpful! " } }