1,720,977 research outputs found
Ultimate Utils - the Ultimate Utils library for Machine Learning and Artificial Intelligence
No full textUlitmate-utils (or uutils) is a collection of common reusable code for machine learning - especially machine learning research.Open Restriction set for Item 120689 on 2021-11-30T17:48:00Z with date null by [email protected] by Brando Miranda ([email protected]) on 2021-11-30T17:51:59Z
No. of bitstreams: 1
ultimate-utils-master.zip: 377167 bytes, checksum: 39c134ec7174eff792ccde2719fc703c (MD5)Made available in DSpace on 2021-11-30T17:51:59Z (GMT). No. of bitstreams: 1
ultimate-utils-master.zip: 377167 bytes, checksum: 39c134ec7174eff792ccde2719fc703c (MD5)
Previous issue date: 2021-11-30Ope
Ultimate Anatome, the ultimate PyTorch library to analyze internal representation of neural networks
No full textUltimate Anatome is a PyTorch library to analyze internal representation of neural networks.
It is an adaptation and improvement on original Anatome library Ryuichiro Hataya.
In particular, we take careful attention to the selection of data points to make your neural network comparisons valid.
In addition, we implement the original way to process the features of convolution neural networks.
In addition we keep improving the library.
Code in https://github.com/brando90/ultimate-anatome and https://pypi.org/project/ultimate-anatome/.Open Restriction set for Item 120688 on 2021-11-30T17:37:01Z with date null by [email protected] by Brando Miranda ([email protected]) on 2021-11-30T17:45:44Z
No. of bitstreams: 1
ultimate-anatome-master.zip: 119664 bytes, checksum: 75796018d47016c0d5c5dbd8e715e833 (MD5)Made available in DSpace on 2021-11-30T17:45:44Z (GMT). No. of bitstreams: 1
ultimate-anatome-master.zip: 119664 bytes, checksum: 75796018d47016c0d5c5dbd8e715e833 (MD5)
Previous issue date: 2021-11-30Ope
Predicting Performance and Generalization in Deep Networks
No full textWe provide the data set and repository of code used for the summer project of 2019 for predicting performance and generalization of deep neural networks using neural deep networks.Open Restriction set for Item 120717 on 2021-12-09T21:41:11Z with date null by [email protected] Restriction set for Item 120717 on 2021-12-09T21:57:25Z with date null by [email protected] by Brando Miranda ([email protected]) on 2021-12-09T22:10:24Z
No. of bitstreams: 4
predicting generalization-dataset.zip: 113541842 bytes, checksum: 8d12b8814e7f79a8fe542652cf8f8737 (MD5)
predicting_performance.pptx: 652892 bytes, checksum: 8ff2dc04393e529088b7266cc6343894 (MD5)
predicting_performance.pdf: 972035 bytes, checksum: a5b2f641cd75403c4832d8b2274a215a (MD5)
predicting_performance_code.zip: 180708 bytes, checksum: f746596912f0461332b2dff3845d4623 (MD5)Made available in DSpace on 2021-12-09T22:10:24Z (GMT). No. of bitstreams: 4
predicting generalization-dataset.zip: 113541842 bytes, checksum: 8d12b8814e7f79a8fe542652cf8f8737 (MD5)
predicting_performance.pptx: 652892 bytes, checksum: 8ff2dc04393e529088b7266cc6343894 (MD5)
predicting_performance.pdf: 972035 bytes, checksum: a5b2f641cd75403c4832d8b2274a215a (MD5)
predicting_performance_code.zip: 180708 bytes, checksum: f746596912f0461332b2dff3845d4623 (MD5)
Previous issue date: 2019Ope
Uniform Stability and Generalization in Learning Theory - a tutorial with Tikhonov Regularization
No full textWe provide a tutorial of the foundations of Uniform Stability and
Generalization in Learning according to the MIT class 9.520 Statistical
Learning Theory and Applications [2] through the example of Tikhonov
Regularization. In particular, our main contributions are: 1) to provide
a new presentation of the proof 2) provide clear motivations and underlying intuition for pedagogy and 3) provide an example with Tikhonov Regularization for concreteness and pedagogy.Open Restriction set for Item 120677 on 2021-11-24T18:59:35Z with date null by [email protected] by Brando Miranda ([email protected]) on 2021-11-24T19:07:51Z
No. of bitstreams: 1
Uniform_Stability_and_Generalization_in_Statistical_Learning_theory.pdf: 283023 bytes, checksum: 9b8f69cc95b143b98f5832581124d23e (MD5)Made available in DSpace on 2021-11-24T19:07:51Z (GMT). No. of bitstreams: 1
Uniform_Stability_and_Generalization_in_Statistical_Learning_theory.pdf: 283023 bytes, checksum: 9b8f69cc95b143b98f5832581124d23e (MD5)
Previous issue date: 2015-01-01Ope
DiMO: Differentiable Model Optimization
Get PDFI am interested in automated methods for model selection - architecture selection, hyperparameter selection and optimizer selection.
To this end, I am developing DiMO (Differentiable Model Optimization): a novel technique for
1) model optimization (i.e. model selection for a fixed dataset by improving the selected architecture, hyperparameters and custom optimizers) and
2) meta-learning (i.e. automatic selection of model architectures for new tasks not observed during training).
The main idea of DiMO is to develop a framework (flexible for both meta-learning and model optimization) that models the entire machine learning pipeline and minimize that objective directly with minimal hacks and tricks.
The hope is that by optimizing the objective directly - with minimal modifications - performance of constructed models can be optimal and the search efficient, due to it's differentiable formulation.Open Restriction set for Item 116758 on 2020-12-14T20:08:23Z with date by [email protected] by Brando Miranda ([email protected]) on 2020-12-14T20:12:43Z
No. of bitstreams: 1
DiMO_research_Statement.pdf: 116939 bytes, checksum: cf4fd7e9bbf25c375a5d2c5ac96f3d67 (MD5)Made available in DSpace on 2020-12-14T20:12:43Z (GMT). No. of bitstreams: 1
DiMO_research_Statement.pdf: 116939 bytes, checksum: cf4fd7e9bbf25c375a5d2c5ac96f3d67 (MD5)
Previous issue date: 2020-03-01Ope
Sketching: a Cognitively inspired Compositional Theorem Prover that Learns to Prove - a Proposal
No full textMathematics is a powerful tool that has benefited humanity for millenia. Proficiency in this subject has far reaching implications for society because physics, mechanical engineering, computer science, statistics, machine learning and the scientific method itself depend on it. For example, being able to automate mathematical proofs would allow us to create safe and interpretable A.I. because a system with provable guarantees is a system we understand and can trust. This is why in this proposal we suggest a novel way of implementing the concept of sketching for automated theorem provers using powerful ideas from cognitive science and artificial neural networks; these ideas are: compositionality, learning-to-learn, learning as model building and curriculum-learn.Open Restriction set for Item 116818 on 2020-12-23T17:59:11Z with date null by [email protected] by Brando Miranda ([email protected]) on 2020-12-23T18:02:24Z
No. of bitstreams: 2
IBM Theorm proving with cognitive science 1 page - Draft 2.pdf: 51397 bytes, checksum: 65858819b62061ecf1b341112cf35973 (MD5)
Proposed Plan of Graduate Study and Research and the applicant's long-range career goals Ford 2020 (abstractions) (Final Draft).pdf: 60604 bytes, checksum: b2793e1bb648cc38d365c6ad4fe0048f (MD5)Made available in DSpace on 2020-12-23T18:02:24Z (GMT). No. of bitstreams: 2
IBM Theorm proving with cognitive science 1 page - Draft 2.pdf: 51397 bytes, checksum: 65858819b62061ecf1b341112cf35973 (MD5)
Proposed Plan of Graduate Study and Research and the applicant's long-range career goals Ford 2020 (abstractions) (Final Draft).pdf: 60604 bytes, checksum: b2793e1bb648cc38d365c6ad4fe0048f (MD5)
Previous issue date: 2020-12-23Ope
Sketching: A Cognitively inspired Compositional Theorem Prover that Learns to Prove
Get PDFCurrent advances in Deep Learning have accelerated A.I. in an unprecedented way. Now we can solve tasks that seemed beyond our reach like language translation, beating top human Go players and generating images indistinguishable from real ones. The progress is impressive but it has been at the cost of transparency. As models become larger and more complex it has become increasingly difficult to understand what they do reliably at many levels of abstraction. One level that inspires me is the mathematical level. Do current Deep Learning models generalize because of implicit regularization? Is the role of over-parameterization the factor that leads to learnability via gradient descent in highly non-convex landscapes? Are classical statistical learning bounds useful in characterizing Deep Neural Networks? With the increased investment in A.I. and people fearlessly building, designing and hacking, the race to explain what all these models do becomes unmanageably complex with our current mathematical tools.Open Restriction set for Item 120622 on 2021-11-15T15:42:15Z with date null by [email protected] by Brando Miranda ([email protected]) on 2021-11-15T15:44:20Z
No. of bitstreams: 1
Draft 2, The Open Philanthropy Project AI Fellows Program .pdf: 89494 bytes, checksum: d0a477daa02332607e6dd0d667ff2975 (MD5)Made available in DSpace on 2021-11-15T15:44:20Z (GMT). No. of bitstreams: 1
Draft 2, The Open Philanthropy Project AI Fellows Program .pdf: 89494 bytes, checksum: d0a477daa02332607e6dd0d667ff2975 (MD5)
Previous issue date: 2018-12-11Ope
DiMO: Differentiable Model Optimization and metaDiMO - Code
Get PDFWe propose an AutoML paradigm that is end-to-end differentiable, versatile, scalable and by design optimizes to transfer to different data sets. The proposed method is scalable because it's fully differentiable and therefore trainable with Gradient Descent (GD). Consequently it does not rely on any Reinforcement Learning (RL) or Evolutionary methods. The proposed method is versatile because every part of the system can be easily learned or constrained with appropriate priors (e.g. biologically plausible priors). The pipeline is highly modular; the architecture search and its hyper parameters, optimizer and how credit is assigned all the way to performance predictor are all separate modules. This gives rise to rich combinations to experiment, deploy, train and constrain the system. Last but not least, the paradigm is formulated to consider multiple data sets at once and thus every module that has been learned is optimized to transfer to different tasks. Additionally, with a set of unknown data sets as part of the validation process, the system could potentially be trained to generalize to data sets it has not seen before.Open Restriction set for Item 120718 on 2021-12-09T21:54:55Z with date null by [email protected] by Brando Miranda ([email protected]) on 2021-12-09T22:00:58Z
No. of bitstreams: 1
dimo-differentiable-model-optimization.zip: 151980 bytes, checksum: 5753e2ff14a65524e52d8619d7b78dc3 (MD5)Made available in DSpace on 2021-12-09T22:00:58Z (GMT). No. of bitstreams: 1
dimo-differentiable-model-optimization.zip: 151980 bytes, checksum: 5753e2ff14a65524e52d8619d7b78dc3 (MD5)
Previous issue date: 2019Ope
Evaluating sublinear estimators for big data
Get PDFIncreasingly, databases are storing more and more data, making it costly to go through all the data one may have in a database. However, users are still interested in being able to query a database holding their data to get some understanding of the data that they have. In this paper we propose three different sampling-based methods to estimate the total mean value of one particular attribute in a particular group of records in a data set. First we approximate the number of elements pertaining to one group and then, estimate their mean value. With these two approximated quantities, we can easily estimate the total amount one group contributes by multiplying both averages. We will also argue the correctness of the algorithms that we propose. We evaluate each algorithm in practice by comparing them on real data.Open Restriction set for Item 116819 on 2020-12-23T18:08:44Z with date null by [email protected] by Brando Miranda ([email protected]) on 2020-12-23T18:09:57Z
No. of bitstreams: 1
Evaluating_sublinear_estimators_for_big_data__UAP_.pdf: 215942 bytes, checksum: c8c39f1dbb9600fe64a9f7d891c7f1e1 (MD5)Made available in DSpace on 2020-12-23T18:09:57Z (GMT). No. of bitstreams: 1
Evaluating_sublinear_estimators_for_big_data__UAP_.pdf: 215942 bytes, checksum: c8c39f1dbb9600fe64a9f7d891c7f1e1 (MD5)
Previous issue date: 2014-05-16Ope
Reducing Training Time with More Data - a Review
Get PDFIn most machine learning problems, we tend to think that training algorithms require more computation time as the number of training samples increases. In this paper we discuss two contexts in which this is not true. In the case of SVM optimization, assuming some desired generalization error, the PEGASOS algorithm statistically needs less runtime with more data. In the case of learning halfspaces over sparse vectors, more training examples reduce the training runtime from exponential to polynomial time.Open Restriction set for Item 120678 on 2021-11-24T19:35:02Z with date null by [email protected] by Brando Miranda ([email protected]) on 2021-11-24T19:37:36Z
No. of bitstreams: 1
Reducing_Training_Timewith_More_Data___a_Review___18409_final_project.pdf: 226858 bytes, checksum: 8b44ce09fc5b54661d51c570c112b345 (MD5)Made available in DSpace on 2021-11-24T19:37:36Z (GMT). No. of bitstreams: 1
Reducing_Training_Timewith_More_Data___a_Review___18409_final_project.pdf: 226858 bytes, checksum: 8b44ce09fc5b54661d51c570c112b345 (MD5)
Previous issue date: 2015-05-14Ope
- …
